System-on-a-wafer with programmable interconnects and methods for making same
The SoW architecture with programmable logic circuits addresses memory bandwidth limitations and latency issues by dynamically reconfiguring interconnection pathways, enhancing flexibility and performance in handling diverse computing workloads.
Patent Information
- Application Number
- PCT/US2026/027256
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-06-03
- Filing Date
- 2026-05-08
- Publication Date
- 2026-11-12
AI Technical Summary
Conventional computing systems face challenges in accommodating rapidly growing workloads, particularly in high-performance computing and AI, due to insufficient memory bandwidth and latency issues, which are exacerbated by the rapid growth of large Al models, leading to bottlenecks in performance.
The integration of heterogeneous computing components on a semiconductor wafer (SoW) with programmable logic circuits that allow dynamic reconfiguration of interconnection pathways, enabling flexible and adaptable computing systems that can support various workloads without extensive redesign.
The SoW architecture provides enhanced flexibility, performance, and adaptability, facilitating efficient handling of diverse computing requirements by dynamically reallocating computational resources and improving memory architectures to meet the demands of dynamic workloads.
Smart Images

Figure US2026027256_12112026_PF_FP_ABST
Abstract
Description
Attorney Docket No. MTSH-152WO01SYSTEM-ON-A-WAFER WITH PROGRAMMABLE INTERCONNECTS AND METHODS FOR MAKING SAME CROSS-REFERENCE TO RELATED APPLICATION(S) (0001] This application claims the priority benefit, under 35 U.S.C. 119(e), of U.S. Application No. 63 / 802,988, filed May 9, 2025 and entitled, “SYSTEM-ON-A-WAFER WITH PROGRAMMABLE LOGIC CIRCUITS AND METHODS FOR MAKING SAME,” U.S. Application No. 63 / 816,848, filed June 3, 2025 and entitled, “SYSTEM-ON-A-WAFER MODULES AND COMPUTING SYSTEMS INCLUDING SAME,” U.S. Application No.64 / 018,967, filed March 27, 2026 and entitled, “SYSTEM-ON-A-WAFER WITH PROGRAMMABLE INTERCONNECTS AND METHODS FOR MAKING SAME,” U.S. Application No. 64 / 018,983, filed March 27, 2026 and entitled, “UNIVERSAL BRIDGE FOR CHIPLET INTERFACE AND COMMUNICATION IN SYSTEM ON WAFER,” and U.S. Application No. 64 / 021,981, filed March 30, 2026 and entitled, “MEMORY ARCHITECTURE IN SYSTEM ON WAFER Al ACCELERATOR.” Each of the aforementioned applications is incorporated herein by reference in its entirety.BACKGROUND(0002] Recent advances in artificial intelligence (Al) and other computational techniques have spurred a need for improvements to computational density, computing power, and heat transfer as a means to improve system performance. For a given integrated circuit (IC), increasing a computational frequency produces increased computational throughput, but also increases the amount of heat generated as a byproduct of the increased computational throughput.
[0003] ICs (also referred to as chips, chiplets, or semiconductor dies) may be manufactured on silicon wafers, with each silicon wafer containing dozens or hundreds of ICs. These ICs may then be separated, e.g., using a die saw, and further integrated into a server or other computing system. However, advances in manufacturing technology have led to the creation of systems-on-a-wafer (SoWs), in which a fully integrated computational system is manufactured using a wafer as a base. These SoWs are not segmented after manufacturing; rather, each SoW may contain a plurality of ICs, interconnections between ICs such as traces, leads, pads, etc., and other components for networking, power distribution, and other functionalities.
[0004] A data center may include multiple networked computing systems to process, store, and distribute large amounts of data. Conventional data centers typically include aggregated computational resources, such as rack servers with central processing units (CPUs) andAttorney Docket No. MTSH-152WO01memory mounted to a single printed circuit board (PCB). For these data centers, the CPUs are configured to perform a wide array of workloads, such as managing databases and providing applications and services to a large user base. However, some workloads, particularly workloads involving high performance computing, Al, and data analytics, have grown in size to such an extent that CPUs alone are no longer suitable to facilitate execution of these workloads without incurring significant time and costs for computation. For example, large language models have recently doubled in size every few months.(0005] For example, the rapid growth and deployment of Al models, such as large language models and image diffusion models, have significantly increased the amount of data processed during Al training and Al inference resulting in appreciably larger workloads for computing systems. The larger workloads have driven the development of computing systems with higher computational densities. However, not all components of the computing system have kept pace in development. In particular, the computation density of computing systems has increased at a far greater pace than memory in these computing systems. As a result, conventional memory architectures are typically unable to provide sufficient memory bandwidth to process the large amounts of data in Al models, thus creating a bottleneck in performance.(0006] Accordingly, data centers configured to execute these workloads have undergone significant changes in recent years to accommodate the rapid growth of these workloads. The integration of semiconductor chips on a common semiconductor wafer in the SoW form factor reduces latency compared to conventional servers. SoWs are a recent technological advancement and lack supporting structures and apparatuses.SUMMARY(0007] Some implementations of the present disclosure relate to computing systems that may include heterogeneous computing components integrated onto a back end of line (BEOL). The computing components may include, but are not limited to, GPUs, central processing units (CPUs), switches, network interface cards (NICs), memory, and the like. Further, the BEOL may be shaped like a semiconductor wafer. These computing systems are referred to herein as a system-on-a-wafer (SoW).(0008] Some implementations of the present disclosure relate to various inventive implementations of a SoW that incorporates at least one computing component with a programmable logic circuit. The programmable logic circuit may provide a way to dynamically reconfigure the interconnection pathways used by the computing component to receive and / orAttorney Docket No. MTSH-152WO01transmit data signals (e.g., signals containing information), to receive and / or transmit power, and / or to electrically ground the computing component. To simplify discussion of the interconnection pathways herein, the interconnection pathways are described as being able to carry a “signal.” The term “signal” is intended to cover any electrical signal transmitted by the interconnection pathways including, but not limited to, a data signal, a power signal, a ground signal, and the like. In this manner, the programmable logic circuit may provide greater flexibility, performance, and / or adaptability to facilitate the use of different combinations of computing components and / or support dynamic workloads.
[0009] Some implementations of the present disclosure relate to SoWs that may include multiple computing components that each have one or more programmable logic circuits. In some implementations, the SoW may include one or more computing components that do not include a programmable logic circuit, i.e., the connections to the network of interconnections are fixed, together with one or more computing components that include a programmable logic circuit. In some implementations, the SoW may include multiple computing components that each have a programmable logic circuit connected to a different number of interconnection pathways in the BEOL. In some implementations, a computing component may be connected to a subset of the interconnection pathways available from the network of interconnections. In some implementations, all the interconnections in the subset may be connected to the programmable logic circuit. In some implementations, only a fraction of the interconnections in the subset may be connected to the programmable logic circuit with the remaining interconnections connected, for example, directly to a functional circuit block.10010] Some implementations of the present disclosure relate to manufacturing SoWs via a chip-first manufacturing process or a chip-last manufacturing process (also referred to as a “chip-on-wafer process”).
[0011] Some implementations of the present disclosure relate to computing systems with heterogeneous computing components (e.g., GPUs, central processing units (CPUs), switches, network interface cards (NICs), memory, and the like) integrated onto a back end of line (BEOL). In some implementations, the BEOL may be shaped like a semiconductor wafer. These computing systems are referred to herein as an SoW. The SoW architecture offers a convenient way to design and build computing systems in a highly modular manner. For example, an SoW may include different combinations of computing components to meet the requirements of different applications (e.g., high memory bandwidth, low latency, high compute density, low power consumption, and the like). Moreover, the SoW architectureAttorney Docket No. MTSH-152WO01facilitates the integration of computing components fabricated using different process nodes, thus reducing manufacturing cost.(0012] Some implementations of the present disclosure relate to various inventive implementations of an SoW that incorporates one or more programmable interconnects. The programmable interconnects provide a way to dynamically reconfigure the interconnection pathways within the SoW. For example, the network of interconnections in the SoW may be adjusted to accommodate different computing component configurations without requiring a physical change to the arrangement and / or layout of the interconnection pathways. In another example, the network of interconnections in the SoW may be adjusted to allocate computational resources to different computing components (e.g., a particular CPU or GPU) in response to dynamically changing workloads managed by the SoW. In this manner, the programmable interconnects may provide greater flexibility, performance, and / or adaptability to facilitate the use of different combinations of computing components and / or support dynamic workloads without requiring extensive redesign of the interconnection network. (00131 Some implementations of the present disclosure relate to computing systems including heterogeneous computing components integrated onto an individual semiconductor wafer. The computing components may include, but are not limited to, a GPU, a CPU, a switch, an I / O computing component, a NIC, a memory module, and the like. A single wafer with a plurality of computing components integrated thereon is referred to herein as an SoW.(0014| Some implementations of the present disclosure relate to various inventive implementations of SoW modules that separate and disaggregate various computing components conventionally integrated into a single conventional SoW. The present disclosure in some implementations is further directed to computing systems incorporating one or more SoW modules.(0015| Some implementations of the present disclosure relate to communication interfaces that may support various communication protocols including, but not limited to, a die-to-die protocol (e.g., Universal Chiplet Interconnect Express (UCIe)), peripheral component interconnect express (PCI Express), Ultra Accelerator Link (UALink), Ethernet, InfiniBand, Ultra Ethernet, Optical I / O, and the like. For example, PCI Express may be used to facilitate communication between a GPU-only SoW module and a CPU-only SoW module. In another example, a UALink may be used to facilitate communication between a pair of GPU-only SoW modules for GPU-to-GPU scaling.Attorney Docket No. MTSH-152WO01
[0016] Some implementations of the present disclosure relate to modular and configurable memory architectures and hierarchies designed for SoW technology. Some implementations of the present disclosure relate to the integration of different memory modules types — such as on-chip SRAM, stacked memory, and side memory — into a cohesive and flexible memory hierarchy. By carefully designing the memory configuration and SoW layout, some implementations of the present technology allows for the modulation of memory types and their arrangement to meet various computing requirements, such as Al model training and inference, high-performance tasks demanding rapid data access, and large-scale models needing substantial storage capacity.
[0017] Some implementations of the present disclosure relate to flexible, programmable interconnect layers within an SoW structure. Where traditional designs utilize passive interconnections, the present technology includes active interconnection components (such as FPGAs) that can be programmed to adapt routing for inter-component communication as needed. Some implementations of the present disclosure may provide the benefits, including but not limited to, programmable active interconnections, enhanced routing flexibility, strategic placement of active interconnects, and / or scalability.
[0018] In some aspects, the techniques described herein relate to a system-on-a-wafer (SoW) including a back end of the line (BEOL) including a plurality of interconnects, the plurality of interconnects collectively providing a plurality of interconnection pathways and a computing component disposed on the BEOL, the computing component including a substrate, a functional circuit block disposed on the substrate, and a programmable logic circuit disposed on the substrate and electrically coupled to the plurality of interconnects, wherein the programmable logic circuit is configured to redirect transmission of a signal from a first interconnection pathway of the plurality of interconnection pathways to a second interconnection pathway of the plurality of interconnection pathways and the first interconnection pathway and the second interconnection pathway are communicatively coupled to the functional circuit block via the programmable logic circuit.
[0019] In some aspects, the techniques described herein relate to a SoW, wherein the computing component is a first computing component, and the SoW further includes a second computing component communicatively coupled to the first interconnection pathway, and a third computing component communicatively coupled to the second interconnection pathway.Attorney Docket No. MTSH-152WO01
[0020] In some aspects, the techniques described herein relate to a SoW, wherein the programmable logic circuit is configured to, when redirecting transmission of the signal from the first interconnection pathway to the second interconnection pathway activate the second interconnection pathway and deactivate the first interconnection pathway.[00211 In some aspects, the techniques described herein relate to a SoW, wherein the programmable logic circuit includes a transistor.
[0022] In some aspects, the techniques described herein relate to a SoW, wherein the programmable logic circuit includes a gate array.
[0023] In some aspects, the techniques described herein relate to a SoW, wherein the programmable logic circuit includes a field programmable gate array (FPGA).
[0024] In some aspects, the techniques described herein relate to a SoW, wherein the programmable logic circuit includes a micro-electromechanical systems (MEMS) switch.
[0025] In some aspects, the techniques described herein relate to a SoW, wherein the computing component is a first computing component, the substrate is a first substrate, the functional circuit block is a first functional circuit block, the programmable logic circuit is a first programmable logic circuit, the SoW further includes a second computing component, including a second substrate, a second functional circuit block disposed on the second substrate, and a second programmable logic circuit disposed on the second substrate, the second programmable logic circuit is electrically coupled to the plurality of interconnects, the second programmable logic circuit is configured to redirect transmission of a signal from a third interconnection pathway of the plurality of interconnection pathways to a fourth interconnection pathway of the plurality of interconnection pathways, and the third interconnection pathway and the fourth interconnection pathway are communicatively coupled to the second computing component.
[0026] In some aspects, the techniques described herein relate to an SoW, wherein the first computing component is different from the second computing component.
[0027] In some aspects, the techniques described herein relate to a SoW, wherein the first programmable logic circuit is communicatively coupled to a first subset of interconnection pathways of the plurality of interconnection pathways, the second programmable logic circuit is communicatively coupled to a second subset of interconnection pathways of the plurality of interconnection pathways, and the first subset of interconnection pathways and the second subset of interconnection pathways include a different number of interconnection pathways.Attorney Docket No. MTSH-152WO01
[0028] In some aspects, the techniques described herein relate to a SoW, wherein the computing component is a first computing component, the substrate is a first substrate, the functional circuit block is a first functional circuit block, the SoW further includes a second computing component, including a second substrate, and a second functional circuit block disposed on the second substrate, and the second computing component does not include a programmable logic circuit to redirect a signal between different interconnection pathways of the plurality of interconnection pathways.(0029] In some aspects, the techniques described herein relate to a SoW, wherein the computing component is communicatively coupled to a subset of interconnection pathways of the plurality of interconnection pathways, and the programmable logic circuit is directly connected to all interconnection pathways of the subset of interconnection pathways.(0030] In some aspects, the techniques described herein relate to a SoW, wherein the computing component is communicatively coupled to a subset of interconnection pathways of the plurality of interconnection pathways, the programmable logic circuit is directly connected to one or more interconnection pathways of the subset of interconnection pathways, and the programmable logic circuit is not directly connected to all interconnection pathways of the subset of interconnection pathways.(0031] In some aspects, the techniques described herein relate to a SoW, wherein the computing component is communicatively coupled to a subset of interconnection pathways of the plurality of interconnection pathways, one or more first interconnection pathways of the subset of interconnection pathways is directly connected to the programmable logic circuit, and one or more second interconnection pathways of the subset of interconnection pathways is directly connected to the functional circuit block.
[0032] In some aspects, the techniques described herein relate to an SoW, wherein the BEOL is shaped as a semiconductor wafer.
[0033] In some aspects, the techniques described herein relate to an SoW, wherein the computing component includes at least one of a graphics processing unit (GPU), a central processing unit (CPU), a memory module, a switch, an input / output (I / O) controller, or a network interface controller (NIC).
[0034] In some aspects, the techniques described herein relate to a SoW, wherein the SoW is configured to operate as an artificial intelligence (Al) accelerator.Attorney Docket No. MTSH-152WO01
[0035] In some aspects, the techniques described herein relate to a SoW, wherein the plurality of interconnects includes at least one interconnect providing only one interconnection pathway of the plurality of interconnection pathways.
[0036] In some aspects, the techniques described herein relate to a SoW, wherein the plurality of interconnects includes at least one interconnect providing two or more interconnection pathways of the plurality of interconnection pathways.
[0037] In some aspects, the techniques described herein relate to a system-on-a-wafer (SoW), including a back end of the line (BEOL), including a plurality of interconnects, a plurality of computing components, disposed on the BEOL, each computing component of the plurality of computing components including a substrate, and a circuit disposed on the substrate, and at least one programmable interconnect communicatively coupled to the plurality of interconnects, wherein the plurality of interconnects and the at least one programmable interconnect provide a plurality of interconnection pathways communicatively coupled to respective computing components of the plurality of computing components, and the at least one programmable interconnect is configured to change at least one interconnection pathway of the plurality of interconnection pathways.
[0038] In some aspects, the techniques described herein relate to an SoW, wherein the BEOL is shaped as a semiconductor wafer.
[0039] In some aspects, the techniques described herein relate to a SoW, wherein the plurality of computing components includes at least one first computing component, and at least one second computing component different from the at least one first computing component.
[0040] In some aspects, the techniques described herein relate to a SoW, wherein the plurality of computing components includes at least one of a graphics processing unit (GPU), a central processing unit (CPU), a memory module, a switch, an input / output (I / O) controller, or a network interface controller (NIC).
[0041] In some aspects, the techniques described herein relate to a SoW, wherein the at least one programmable interconnect includes a transistor.
[0042] In some aspects, the techniques described herein relate to a SoW, wherein the at least one programmable interconnect includes a gate array.
[0043] In some aspects, the techniques described herein relate to a SoW, wherein the at least one programmable interconnect includes a field programmable gate array (FPGA).Attorney Docket No. MTSH-152WO01
[0044] In some aspects, the techniques described herein relate to a SoW, wherein the at least one programmable interconnect includes a micro-electromechanical systems (MEMS) switch. (0045] In some aspects, the techniques described herein relate to a SoW, wherein the plurality of computing components includes a first computing component containing the at least one programmable interconnect.
[0046] In some aspects, the techniques described herein relate to a SoW, wherein the plurality of interconnects includes a first plurality of interconnects disposed in a first redistribution layer of the BEOL and communicatively coupled to the plurality of computing components, a second plurality of interconnects disposed in a bridge layer of the BEOL and communicatively coupled to the first plurality of interconnects, and a third plurality of interconnects disposed in a second redistribution layer of the BEOL and communicatively coupled to the second plurality of interconnects.(0047] In some aspects, the techniques described herein relate to a SoW, wherein the at least one programmable interconnect is disposed in the bridge layer of the BEOL.(0048] In some aspects, the techniques described herein relate to a SoW, further including at least one local silicon interconnect (LSI) disposed in the bridge layer of the BEOL.
[0049] In some aspects, the techniques described herein relate to a SoW, further including a plurality of solder balls disposed in a bump layer of the BEOL, wherein the plurality of solder balls is communicatively coupled to the third plurality of interconnects.
[0050] In some aspects, the techniques described herein relate to a SoW, wherein the at least one programmable interconnect is disposed in the bump layer of the BEOL.
[0051] In some aspects, the techniques described herein relate to a SoW, wherein the plurality of computing components is configured to operate as an artificial intelligence (Al) accelerator.
[0052] In some aspects, the techniques described herein relate to a system-on-a-wafer (SoW), including a computing component layer, the computing component layer including a plurality of computing components, each computing component of the plurality of computing components including a substrate, and a circuit integrally formed on the substrate, a back end of the line (BEOL), including a first redistribution layer disposed on the computing component layer, the first redistribution layer including a first plurality of interconnects communicatively coupled to the plurality of computing components, a bridge layer disposed on the first redistribution layer, the bridge layer including a second plurality of interconnectsAttorney Docket No. MTSH-152WO01communicatively coupled to the first plurality of interconnects, a second redistribution layer disposed on the bridge layer, the second redistribution layer including a third plurality of interconnects communicatively coupled to the second plurality of interconnects, and a bump layer disposed on the second redistribution layer, the bump layer including a plurality of solder balls communicatively coupled to the third plurality of interconnects, and at least one programmable interconnect disposed in at least one of the computing component layer, the bridge layer, or the bump layer, wherein the BEOL is shaped as a semiconductor wafer.(0053] In some aspects, the techniques described herein relate to a SoW, wherein the bridge layer further includes at least one local silicon interconnect (LSI).(0054| In some aspects, the techniques described herein relate to a computing system, including one or more system-on-a-wafer (SoW) modules, each SoW module of the one or more SoW modules including a back end of line (BEOL), and a plurality of computing components electrically coupled to the BEOL, the plurality of computing components consisting of at least one of an input / output (I / O) computing component or a network interface controller (NIC), and a plurality of functional computing components, each functional computing component of the plurality of functional computing components being functionally identical to every other functional computing component of the plurality of functional computing components.
[0055] In some aspects, the techniques described herein relate to a computing system, wherein the one or more SoW modules includes a SoW module, the SoW module including a plurality of central processing units (CPUs) as the plurality of functional computing components of the SoW module.
[0056] In some aspects, the techniques described herein relate to a computing system, wherein the one or more SoW modules includes a SoW module, the SoW module including a plurality of graphics processing units (GPUs) as the plurality of functional computing components of the SoW module.
[0057] In some aspects, the techniques described herein relate to a computing system, wherein the one or more SoW modules includes a SoW module, the SoW module including a plurality of switches as the plurality of functional computing components of the SoW module.
[0058] In some aspects, the techniques described herein relate to a computing system, wherein the one or more SoW modules includes a SoW module, the SoW module including a plurality of memory modules as the plurality of functional computing components of the SoW module.Attorney Docket No. MTSH-152WO01
[0059] In some aspects, the techniques described herein relate to a computing system, wherein the one or more SoW modules includes a first SoW module including a plurality of central processing units (CPUs) as the plurality of functional computing components of the first SoW module, a second SoW module including a plurality of graphics processing units (GPUs) as the plurality of functional computing components of the second SoW module, a third SoW module including a plurality of switches as the plurality of functional computing components of the third SoW module, and a fourth SoW module including a plurality of memory modules as the plurality of functional computing components of the fourth SoW module.
[0060] In some aspects, the techniques described herein relate to a computing system, wherein each SoW module of the one or more SoW modules is communicatively coupled to every other SoW module of the one or more SoW modules.
[0061] In some aspects, the techniques described herein relate to a computing system, wherein each SoW module of the one or more SoW modules is communicatively coupled to up to four other SoW modules of the one or more SoW modules.
[0062] In some aspects, the techniques described herein relate to a computing system, further including a plurality of cables, wherein each cable of the plurality of cables is communicatively coupled to a pair of SoW modules of the one or more SoW modules and is configured to transmit one or more signals between the pair of SoW modules.
[0063] In some aspects, the techniques described herein relate to a computing system, wherein each SoW module of the one or more SoW modules further includes a connector having a first end and a second end, wherein the first end is communicatively coupled to the at least one of the I / O computing component or the NIC, and the second end is communicatively coupled to one cable of the plurality of cables.
[0064] In some aspects, the techniques described herein relate to a computing system, further including a printed circuit board (PCB) to communicatively couple each SoW module of the one or more SoW modules to every other SoW module of the one or more SoW modules.
[0065] In some aspects, the techniques described herein relate to a computing system, wherein the one or more SoW modules includes a first SoW module, a second SoW module communicatively coupled to the first SoW module via a first connection standard, and a third SoW module communicatively coupled to the first SoW module via a second connection standard different from the first connection standard.Attorney Docket No. MTSH-152WO01
[0066] In some aspects, the techniques described herein relate to a computing system, wherein the at least one of the I / O computing component or the NIC supports at least one of a Universal Chiplet Interconnect Express (UCIe) connection, a Peripheral Component Express (PCIe) connection, an Ultra Accelerator Link (UALink) connection, an Ethernet connection, an Ultra Ethernet connection, an InfiniBand connection, or an optical input / output (I / O) connection.
[0067] In some aspects, the techniques described herein relate to a computing system, including one or more system-on-a-wafer (SoW) modules, each SoW module of the one or more SoW modules including a back end of line (BEOL), and a plurality of computing components electrically coupled to the BEOL, the plurality of computing components including at least one of an input / output (I / O) computing component or a network interface controller (NIC), and a plurality of functional computing components predominantly including a first type of functional computing component.(0068] In some aspects, the techniques described herein relate to a computing system, wherein a percentage of the plurality of functional computing components corresponding to the first type of functional computing component is greater than or equal to 90%.
[0069] In some aspects, the techniques described herein relate to a computing system, wherein a quantity of the plurality of functional computing components corresponding to one or more other type of functional computing components different from the first type of functional computing components is less than or equal to 150.(0070] In some aspects, the techniques described herein relate to a computing system, the computing system including a substrate including a redistribution layer (RDL), a first unit block disposed on a first surface of the substrate, wherein the first unit block includes a first system-on-a-chip (SOC) and a second SOC, a first memory module disposed on the first SOC, a second memory module disposed on the second SOC, a third memory module disposed on the substrate, the third memory module being adjacent to the first SOC, and a fourth memory module disposed on the substrate, the fourth memory module being adjacent to the second SOC, wherein the first SOC is communicatively coupled to the fourth memory module through at least one functional block, and wherein the first SOC and the second SOC each includes at least one logic integrated circuit (IC).(00711 In some aspects, the techniques described herein relate to a computing system, further including a second functional block disposed in the second SOC, wherein the second memoryAttorney Docket No. MTSH-152WO01module is communicatively coupled to the third memory module through the second functional block.(0072] In some aspects, the techniques described herein relate to a computing system, wherein the third memory module and the fourth memory module are communicatively coupled through the RDL.(0073] In some aspects, the techniques described herein relate to a computing system, further including a third functional block disposed in the first memory module and a fourth functional block disposed in the second memory module, wherein the first memory module is communicatively coupled to the second memory module through the third functional block and the fourth functional block.(0074] In some aspects, the techniques described herein relate to a computing system, further including a fifth functional block disposed in the first SOC and a sixth functional block disposed in the second SOC, wherein the first memory module is further communicatively coupled to the second memory module through the fifth functional block and the sixth functional block.(0075] In some aspects, the techniques described herein relate to a computing system, further including a voltage regulation module (VRM) configured to provide regulated voltage to the first unit block.(0076] In some aspects, the techniques described herein relate to a computing system, wherein the VRM is disposed on a second surface of the substrate opposite the first surface of the substrate.(0077] In some aspects, the techniques described herein relate to a computing system, wherein the VRM is electrically coupled to the first unit block through the RDL.(0078] In some aspects, the techniques described herein relate to a computing system, further including a second unit block communicatively coupled to the first unit block through the RDL.(0079] In some aspects, the techniques described herein relate to a computing system, further including at least 50 logic ICs disposed on the substrate.(0080] In some aspects, the techniques described herein relate to a computing system, further including at least 100 logic ICs disposed on the substrate.(0081] In some aspects, the techniques described herein relate to a computing system, further including at least 200 logic ICs disposed on the substrate.Attorney Docket No. MTSH-152WO01
[0082] In some aspects, the techniques described herein relate to a computing system, further including at least one input / output (I / O) chip communicatively coupled to the first unit block through the RDL and at least one wafer connector communicatively coupled to the first unit block through the RDL.
[0083] In some aspects, the techniques described herein relate to a computing system, the computing system including a substrate including a redistribution layer (RDL), a plurality of three-dimensional integrated circuit (3DIC) stacks disposed on a first surface of the substrate, wherein each 3DIC stack of the plurality of 3DIC stacks includes a logic IC, at least one memory module disposed on the logic IC, and a plurality of functional blocks communicatively coupling each 3DIC stack with a component, and a plurality of wafer connectors communicatively coupled to the RDL, wherein the plurality of wafer connectors is configured to transmit data between each 3DIC stacks and an external logic IC disposed externally to the computing system, wherein the plurality of 3DIC stacks are disposed on the first surface in a first pattern.
[0084] In some aspects, the techniques described herein relate to a computing system, wherein the plurality of 3DIC stacks includes at least 503DIC stacks.
[0085] In some aspects, the techniques described herein relate to a computing system, wherein the plurality of 3DIC stacks includes at least 1003DIC stacks.
[0086] In some aspects, the techniques described herein relate to a computing system, wherein the plurality of 3DIC stacks includes at least 2003DIC stacks.
[0087] In some aspects, the techniques described herein relate to a computing system, wherein the plurality of 3DIC stacks is arranged in a corresponding plurality of unit blocks, each unit block communicatively coupled through the RDL to respective other unit blocks of the plurality of unit blocks.
[0088] In some aspects, the techniques described herein relate to a computing system, wherein the plurality of unit blocks includes a first unit block and a second unit block, the first unit block is communicatively coupled to the second unit block through a first functional block disposed in a first logic IC, and a second functional block disposed in a second logic IC, and the first unit block and the second unit block are communicatively coupled to an external logic IC through at least one wafer connector of the plurality of wafer connectors.
[0089] In some aspects, the techniques described herein relate to a computing system, wherein each unit block of the plurality of unit blocks further includes a plurality of memory modulesAttorney Docket No. MTSH-152WO01disposed on the first surface of the substrate and each memory module of the plurality of memory modules is communicatively coupled to a respective logic IC of the respective unit block through a second plurality of functional blocks.
[0090] In some aspects, the techniques described herein relate to a computing system, the computing system including a substrate, a first component disposed on the substrate, wherein the first component includes first circuitry configured to perform a first function, and a first functional block having a first plurality of communications paths, a second component disposed on the substrate, wherein the second component includes second circuitry configured to perform a second function, and a second functional block having a second plurality of communications paths, and a universal bridge communicatively coupling the first component to the second component, wherein the first functional block and the second functional block each includes respective surfaces having respective exposed leads arranged in a primary pattern, and the universal bridge has a plurality of surfaces having exposed universal bridge leads arranged to interface with the primary pattern.
[0091] In some aspects, the techniques described herein relate to a computing system, wherein the first component and the second component include at least two different components selected from a graphics processing unit (GPU), a central processing unit (CPU), a memory module, an input / output (I / O) chip, a link chiplet, a network interface card (NIC), a tensor processing unit (TPU), a data processing unit (DPU), a switch, or a system-on-a-chip (SOC).
[0092] In some aspects, the techniques described herein relate to a computing system, wherein the exposed universal bridge leads are symmetric about a plane parallel to the plurality of surfaces.
[0093] In some aspects, the techniques described herein relate to a computing system, wherein the exposed universal bridge leads are glide symmetric about a plane parallel to the plurality of surfaces.
[0094] In some aspects, the techniques described herein relate to a computing system, wherein the computing system is a component of an immersion cooling system, and the component is immersed in an immersion cooling liquid.
[0095] In some aspects, the techniques described herein relate to a computing system, wherein the exposed universal bridge leads are arranged in the primary pattern.Attorney Docket No. MTSH-152WO01
[0096] In some aspects, the techniques described herein relate to a computing system, wherein the exposed universal bridge leads are arranged in a second pattern that mirrors the primary pattern.
[0097] In some aspects, the techniques described herein relate to a computing system, wherein the universal bridge includes an universal bridge substrate and a plurality of traces.
[0098] In some aspects, the techniques described herein relate to a computing system, wherein the universal bridge substrate includes at least one of silicon, silicon nitride, silicon dioxide, or printed circuit board (PCB).
[0099] In some aspects, the techniques described herein relate to a computing system, wherein the plurality of traces include gold, copper, or aluminum.
[0100] In some aspects, the techniques described herein relate to a computing system, wherein the universal bridge includes a first integrated circuit (IC) configured to control a routing of communications through the plurality of traces.
[0101] In some aspects, the techniques described herein relate to a computing system, wherein the first IC includes a field-programmable gate array (FPGA).
[0102] In some aspects, the techniques described herein relate to a computing system, wherein the first component is not configured to perform the second function, and the second component is not configured to perform the first function.
[0103] In some aspects, the techniques described herein relate to a computing system, wherein the first component is configured to perform the second function, and the second component is configured to perform the first function.
[0104] In some aspects, the techniques described herein relate to a computing system, wherein the universal bridge is disposed above the substrate.
[0105] In some aspects, the techniques described herein relate to a computing system, wherein the universal bridge is disposed in the substrate.
[0106] In some aspects, the techniques described herein relate to a computing system, wherein the universal bridge is disposed in at least one of a redistribution layer (RDL) or a local silicon interconnect (LSI) layer.
[0107] It should be appreciated that all combinations of the foregoing concepts and additional concepts discussed in greater detail below (provided such concepts are not mutuallyAttorney Docket No. MTSH-152WO01inconsistent) are contemplated as being part of the inventive subject matter disclosed herein. In particular, all combinations of claimed subject matter appearing at the end of this disclosure are contemplated as being part of the inventive subject matter disclosed herein. It should also be appreciated that terminology explicitly employed herein that also may appear in any disclosure incorporated by reference should be accorded a meaning most consistent with the particular concepts disclosed herein.BRIEF DESCRIPTION OF THE DRAWINGS|0108] The skilled artisan will understand that the drawings primarily are for illustrative purposes and are not intended to limit the scope of the inventive subject matter described herein. The drawings are not necessarily to scale; in some instances, various aspects of the inventive subject matter disclosed herein may be shown exaggerated or enlarged in the drawings to facilitate an understanding of different features. In the drawings, like reference characters generally refer to like features (e.g., functionally similar and / or structurally similar elements).
[0109] FIG. 1 A shows a top view of a system-on-a-wafer (SoW).
[0110] FIG. IB shows a partial cross-section view of the SoW of FIG. 1A along the plane A-A of FIG. 1A.
[0111] FIG. 2 shows a partial cross-section view of another SoW with a different combination of computing components along the same plane A- A of FIG. 1 A.
[0112] FIG. 3 A shows a top view of an example SoW with multiple computing components that each have at least one programmable logic circuit.
[0113] FIG. 3B shows a partial cross-section view of the SoW of FIG. 3 A along the plane A-A of FIG. 3 A.[01.14] FIG. 4 shows a partial cross-section view of another example SoW with a different combination of computing components that each have at least one programmable logic circuit.
[0115] FIG. 5 shows a top view of another example SoW with multiple computing components that each have at least one programmable logic circuit.
[0116] FIG. 6 shows a method of making a SoW with a computing component that includes at least one programmable logic circuit.
[0117] FIG. 7A shows a top view of an example SoW.Attorney Docket No. MTSH-152WO01
[0118] FIG. 7B shows a cross-section view of the SoW of FIG. 7A corresponding to the plane A- A of FIG. 7 A.
[0119] FIG. 8 shows a cross-section view of another example SoW with a different combination of computing components.[0120J FIG. 9 shows a cross-section view of an example SoW that includes a computing component with at least one programmable interconnect (PI).
[0121] FIG. 10 shows a cross-section view of another example SoW that includes at least one programmable interconnect in a bridge layer of the SoW.(0122] FIG. 11 shows a cross-section view of another example SoW that includes at least one programmable interconnect and at least one local silicon interconnect (LSI) in a bridge layer of the SoW.[01231 FIG. 12 shows a cross-section view of another example SoW that includes at least one programmable interconnect in a bump layer of the SoW.(0124] FIG. 13 shows a cross-section view of another example SoW that includes a computing component with at least one programmable interconnect and programmable interconnects in a bridge layer and a bump layer of the SoW.(0125] FIG. 14 shows a method of making a SoW with at least one programmable interconnect.(0126] FIG. 15A shows a top view of an example SoW module that includes graphics processing units (GPUs).(0127] FIG. 15B shows a top view of an example SoW module that includes memory modules.(0128] FIG. 15C shows a top view of an example SoW module that includes switches.(0129] FIG. 15D shows a top view of an example SoW module that includes central processing units (CPUs).
[0130] FIG. 16 shows an example computing system that includes multiple SoW modules communicatively coupled via cables.
[0131] FIG. 17 shows an example computing system that includes multiple SoW modules communicatively coupled via a printed circuit board (PCB).
[0132] FIG. 18 shows an example computing system for a server that includes multiple SoW modules.Attorney Docket No. MTSH-152WO01
[0133] FIG. 19 shows a method of making a computing system with multiple SoW modules.
[0134] FIG. 20A illustrates a top view of an SoW in accordance with the present technology.
[0135] FIG. 20B illustrates a top view of an exemplary unit block.
[0136] FIG. 20C illustrates a side view of an exemplary unit block.
[0137] FIG. 21 A illustrates a top view of an SoW in accordance with the present technology.
[0138] FIG. 2 IB illustrates a top view of an exemplary unit block.
[0139] FIG. 21C illustrates a side view of an exemplary unit block.
[0140] FIG. 22A illustrates a top view of an SoW in accordance with the present technology.[0141 { FIG. 22B illustrates a top view of an exemplary unit block.
[0142] FIG. 22C illustrates a side view of an exemplary unit block.
[0143] FIG. 23 A illustrates a top view of an SoW in accordance with the present technology.
[0144] FIG. 23B illustrates a top view of an exemplary unit block.
[0145] FIG. 23 C illustrates a side view of an exemplary unit block.
[0146] FIG. 24 shows a method of making a memory architecture in an SoW.
[0147] FIG. 25 is a top view of an SoW according to the present technology.
[0148] FIG. 26 illustrates a cross-sectional view of an SoW in accordance with the present technology.
[0149] FIG. 27A illustrates a top view of an SoW in accordance with the present technology.
[0150] FIG. 27B illustrates an exemplary unit block.
[0151] FIG. 28A illustrates an SoW in accordance with the present technology.
[0152] FIG. 28B illustrates an exemplary unit block.
[0153] FIG. 29A illustrates an SoW in accordance with the present technology.
[0154] FIG. 29B illustrates an exemplary unit block.
[0155] FIG. 30 shows a method of making a SoW with a universal bridge for communication in an SoW.Attorney Docket No. MTSH-152WO01
[0156] FIG. 31 depicts aspects of an immersion cooling system for dissipating heat from one or more heat-generating components such as semiconductor die packages via immersion cooling.DETAILED DESCRIPTION
[0157] Following below are more detailed descriptions of various concepts related to, and implementations of, systems and apparatuses for systems-on-a-wafer (SoWs) and computing systems including the same. Also following below are more detailed descriptions of methods for making the SoWs disclosed herein. It should be appreciated that various concepts introduced above and discussed in greater detail below may be implemented in multiple ways. Examples of specific implementations and applications are provided primarily for illustrative purposes so as to enable those skilled in the art to practice the implementations and alternatives apparent to those skilled in the art. It should also be appreciated that one or more features discussed in connection with a given example of a SoW may be employed in other examples of SoWs, respectively, according to the present disclosure, such that the various features disclosed herein may be readily combined in a given SoW according to the present disclosure (provided that respective features are not mutually inconsistent).
[0158] The figures and example implementations described below are not meant to limit the scope of the present implementations to a single embodiment. Other implementations are possible by way of interchange of some or all of the described or illustrated elements. Moreover, where certain elements of the disclosed example implementations may be partially or fully implemented using known components, in some instances only those portions of such known components that are necessary for an understanding of the present implementations are described, and detailed descriptions of other portions of such known components are omitted so as not to obscure the present implementations.|0159| Certain dimensions and features of the SoWs are described herein using the terms “approximately,” “about,” “substantially,” and / or “similar.” As used herein, the terms “approximately,” “about,” “substantially,” and / or “similar” indicates that each of the described dimensions or features is not a strict boundary or parameter and does not exclude functionally similar variations therefrom. Unless context or the description indicates otherwise, the use of the terms “approximately,” “about,” “substantially,” and / or “similar” in connection with a numerical parameter indicates that the numerical parameter includes variations that, using mathematical and industrial principles accepted in the art (e.g., rounding, measurement or other systematic errors, manufacturing tolerances, etc.), would not vary the least significant digit.Attorney Docket No. MTSH-152WO011. An Example SOW with Programmable Logic Circuits
[0160] Traditional computing architectures are unable to scale with the large amounts of data processed in modem Al models due to limitations on power efficiency and memory bandwidth. For example, computing systems based on the von Neumann architecture typically include a processor (e.g., central processing unit (CPU), a graphics processing unit (GPU)) and memory that communicate via relatively long electrical interconnects integrated in a printed circuit board (PCB). The transmission of large amounts of data in these systems is limited by the power required to transmit data and the latency of transmission. Furthermore, the PCB-integrated format often results in computing systems with a relatively large physical footprint, which limits the computational density attainable for the overall system. For example, the PCB of a server may have an area that is two times larger than the area of the constituent computing components mounted to the PCB.
[0161] To address some or all of the foregoing limitations, some implementations of the present disclosure relate to computing systems that may include heterogeneous computing components integrated onto a back end of line (BEOL). The computing components may include, but are not limited to, GPUs, central processing units (CPUs), switches, network interface cards (NICs), memory, and the like. Further, the BEOL may be shaped like a semiconductor wafer. These computing systems are referred to herein as a system-on-a-wafer (SoW).
[0162] The integration of multiple computing components directly onto a wafer-shaped BEOL provides a way to significantly reduce the physical distance between computing components, especially processing and memory, resulting in higher density, improved power efficiency, and greater memory bandwidth. The SoW architecture also offers a convenient way to design and build computing systems in a highly modular manner. For example, a SoW may include different combinations of computing components to meet the requirements of different applications (e.g., high memory bandwidth, low latency, high compute density, low power consumption, and the like). Moreover, the SoW architecture facilitates the integration of computing components fabricated using different process nodes, thus reducing manufacturing cost.
[0163] However, the design of the computing components and the manner in which they electrically interface with the electrical wiring of the SoW are typically different for computing components that perform different functions and often change over time as improvements areAttorney Docket No. MTSH-152WO01made to the design of the computing components. Accordingly, one challenge with SoWs is providing an electrical interface on the BEOL that can readily accommodate different computing components while maintaining efficient routing of interconnection pathways and interconnectivity between the computing components. Conventional interconnections, such as the traces in RDLs and LSI, are passive, i.e., the interconnections are static and fixed over time. As a result, if a computing component in the SoW is replaced with another computing component having a different design, significant modifications to the network of interconnections within the SoW are often required. The lack of flexibility in the interconnections can result in increased design time, greater complexity, and potential performance bottlenecks, especially with dynamically changing workloads commonly encountered in Al training and Al inference.(0164] In view of the foregoing limitations of passive interconnections in SoWs, some implementations of the present disclosure relate to various inventive implementations of a SoW that incorporates at least one computing component with a programmable logic circuit. The programmable logic circuit provides a way to dynamically reconfigure the interconnection pathways used by the computing component to receive and / or transmit data signals (e.g., signals containing information), to receive and / or transmit power, and / or to electrically ground the computing component. To simplify discussion of the interconnection pathways herein, the interconnection pathways are described as being able to carry a “signal.” The term “signal” is intended to cover any electrical signal transmitted by the interconnection pathways including, but not limited to, a data signal, a power signal, a ground signal, and the like. In this manner, the programmable logic circuit may provide greater flexibility, performance, and / or adaptability to facilitate the use of different combinations of computing components and / or support dynamic workloads.
[0015] In some implementations, the programmable logic circuit may facilitate the integration of the computing component onto a BEOL without requiring a physical change to the arrangement and / or layout of the interconnection pathways in the BEOL. The programmable logic circuit may thus allow the computing components in the SoW to be updated and / or replaced while reducing or, in some instances, avoiding any redesign of the network of interconnection pathways in the BEOL. In this manner, the inclusion of a programmable logic circuit in the computing component may reduce the overall cost of designing a SoW and provide a scalable approach to change the arrangement and / or combination of computing components in the SoW.Attorney Docket No. MTSH-152WO01
[0166] In some implementations, the programmable logic circuit may provide a way to dynamically allocate computational resources to different computing components in the SoW, particularly in response to dynamically changing workloads managed by the SoW. For example, the programmable logic circuit may dynamically change the number of interconnection pathways available to a given computing component and, in the process, control the number and / or type of computing components communicatively coupled to that given computing component. For instance, the SoW may include multiple CPUs that each include a programmable logic circuit. Each respective programmable logic circuit of the CPUs may operate in coordination with each other to dynamically change the number of memory modules and / or GPUs connected to each CPU. The respective programmable logic circuits may thus allow for more efficient allocation of computational resources to each CPU based on their respective workloads.
[0167] In some implementations, the programmable logic circuit may be communicatively coupled directly to the network of interconnections in the BEOL. In some implementations, the computing component may include a functional circuit block and the programmable logic circuit may manage communication between the functional circuit block and other computing components of the SoW. This communication may be accomplished, for example, by the programmable logic circuit activating and deactivating different interconnection pathways connected to the computing component. In one non-limiting example, the programmable logic circuit may include a transistor. In another non-limiting example, the programmable logic circuit may include a gate array. In yet another non-limiting example, the programmable logic circuit may include a field programmable gate array (FPGA). In yet another non-limiting example, the programmable logic circuit may include a micro-electromechanical systems (MEMS) switch.
[0168] Some implementations of the present disclosure relate to SoWs that may include multiple computing components that each have one or more programmable logic circuits. In some implementations, the SoW may include one or more computing components that do not include a programmable logic circuit, i.e., the connections to the network of interconnections are fixed, together with one or more computing components that include a programmable logic circuit. In some implementations, the SoW may include multiple computing components that each have a programmable logic circuit connected to a different number of interconnection pathways in the BEOL. In some implementations, a computing component may be connected to a subset of the interconnection pathways available from the network of interconnections. InAttorney Docket No. MTSH-152WO01some implementations, all the interconnections in the subset may be connected to the programmable logic circuit. In some implementations, only a fraction of the interconnections in the subset may be connected to the programmable logic circuit with the remaining interconnections connected, for example, directly to a functional circuit block.
[0169] In yet another aspect, the SoWs disclosed herein may be manufactured via a chip-first manufacturing process or a chip-last manufacturing process (also referred to as a “chip-on-wafer process”).|0170] In one example implementation, a system-on-a-wafer (SoW) includes a back end of the line (BEOL) including a plurality of interconnects where the plurality of interconnects collectively provide a plurality of interconnection pathways; and a computing component disposed on the BEOL where the computing component includes a substrate, a functional circuit block disposed on the substrate, and a programmable logic circuit disposed on the substrate and electrically coupled to the plurality of interconnects, wherein the programmable logic circuit is configured to redirect transmission of a signal from a first interconnection pathway of the plurality of interconnection pathways to a second interconnection pathway of the plurality of interconnection pathways, and the first interconnection pathway and the second interconnection pathway are communicatively coupled to the functional circuit block via the programmable logic circuit.
[0171] The computing component may be a first computing component, and the SoW may further include a second computing component communicatively coupled to the first interconnection pathway, and a third computing component communicatively coupled to the second interconnection pathway. The programmable logic circuit may be configured to, when redirecting transmission of the signal from the first interconnection pathway to the second interconnection pathway, activate the second interconnection pathway and deactivate the first interconnection pathway.
[0172] The programmable logic circuit may include a transistor. The programmable logic circuit may include a gate array. The programmable logic circuit may include a field programmable gate array (FPGA). The programmable logic circuit may include a microelectromechanical systems (MEMS) switch.
[0173] The computing component may be a first computing component, the substrate may be a first substrate, the functional circuit block may be a first functional circuit block, the programmable logic circuit may be a first programmable logic circuit, the SoW may furtherAttorney Docket No. MTSH-152WO01include a second computing component that includes a second substrate, a second functional circuit block disposed on the second substrate, and a second programmable logic circuit disposed on the second substrate, the second programmable logic circuit may be electrically coupled to the plurality of interconnects, the second programmable logic circuit may be configured to redirect transmission of a signal from a third interconnection pathway of the plurality of interconnection pathways to a fourth interconnection pathway of the plurality of interconnection pathways, and the third interconnection pathway and the fourth interconnection pathway may be communicatively coupled to the second computing component. The first computing component may be different from the second computing component. The first programmable logic circuit may be communicatively coupled to a first subset of interconnection pathways of the plurality of interconnection pathways, the second programmable logic circuit may be communicatively coupled to a second subset of interconnection pathways of the plurality of interconnection pathways, and the first subset of interconnection pathways and the second subset of interconnection pathways may include a different number of interconnection pathways.
[0174] The computing component may be a first computing component, the substrate may be a first substrate, the functional circuit block may be a first functional circuit block, the SoW may further include a second computing component that includes a second substrate, and a second functional circuit block disposed on the second substrate, and the second computing component may not include a programmable logic circuit to redirect a signal between different interconnection pathways of the plurality of interconnection pathways.
[0175] The computing component may be communicatively coupled to a subset of interconnection pathways of the plurality of interconnection pathways, and the programmable logic circuit may be directly connected to all interconnection pathways of the subset of interconnection pathways. The computing component may be communicatively coupled to a subset of interconnection pathways of the plurality of interconnection pathways, the programmable logic circuit may be directly connected to one or more interconnection pathways of the subset of interconnection pathways, and the programmable logic circuit may not be directly connected to all interconnection pathways of the subset of interconnection pathways. The computing component may be communicatively coupled to a subset of interconnection pathways of the plurality of interconnection pathways, one or more first interconnection pathways of the subset of interconnection pathways may be directly connected to the programmable logic circuit, and one or more second interconnection pathways of the subset ofAttorney Docket No. MTSH-152WO01interconnection pathways may be directly connected to the functional circuit block.
[0176] The BEOL may be shaped as a semiconductor wafer. The computing component may include at least one of a graphics processing unit (GPU), a central processing unit (CPU), a memory module, a switch, an input / output (VO) controller, or a network interface controller (NIC). The SoW may be configured to operate as an artificial intelligence (Al) accelerator. The plurality of interconnects may include at least one interconnect providing only one interconnection pathway of the plurality of interconnection pathways. The plurality of interconnects may include at least one interconnect providing two or more interconnection pathways of the plurality of interconnection pathways.
[0177] Following below are more detailed descriptions of various concepts related to, and implementations of, a SoW with a programmable logic circuit and methods for making the SoWs disclosed herein. It should be appreciated that various concepts introduced above and discussed in greater detail below may be implemented in multiple ways. Examples of specific implementations and applications are provided primarily for illustrative purposes so as to enable those skilled in the art to practice the implementations and alternatives apparent to those skilled in the art.
[0178] The figures and example implementations described below are not meant to limit the scope of the present implementations to a single embodiment. Other implementations are possible by way of interchange of some or all of the described or illustrated elements. Moreover, where certain elements of the disclosed example implementations may be partially or fully implemented using known components, in some instances only those portions of such known components that are necessary for an understanding of the present implementations are described, and detailed descriptions of other portions of such known components are omitted so as not to obscure the present implementations.[0179| In the discussion below, various examples of a SoW with a programmable logic circuit are provided, wherein a given example or set of examples showcases a computing component in a computing component layer with the programmable logic circuit, a plurality of interconnects in one or more redistribution layers and / or a bridge layer, a local silicon interconnect, a plurality of solder balls in a bump layer, a thermal module, a connector, and a power module. It should be appreciated that one or more features discussed in connection with a given example of a SoW may be employed in other examples of SoWs, respectively, according to the present disclosure, such that the various features disclosed herein may beAttorney Docket No. MTSH-152WO01readily combined in a given SoW according to the present disclosure (provided that respective features are not mutually inconsistent).
[0180] Certain dimensions and features of the SoWs are described herein using the terms “approximately,” “about,” “substantially,” and / or “similar.” As used herein, the terms “approximately,” “about,” “substantially,” and / or “similar” indicates that each of the described dimensions or features is not a strict boundary or parameter and does not exclude functionally similar variations therefrom. Unless context or the description indicates otherwise, the use of the terms “approximately,” “about,” “substantially,” and / or “similar” in connection with a numerical parameter indicates that the numerical parameter includes variations that, using mathematical and industrial principles accepted in the art (e.g., rounding, measurement or other systematic errors, manufacturing tolerances, etc.), would not vary the least significant digit.1.1. An Example of a Conventional System-on-a-Wafer (SoW)
[0181] A system-on-a-wafer (SoW) architecture may facilitate the heterogenous integration of computing components into a single computing system. This architecture provides a way to increase the compute density of a computing system, which can reduce the amount of energy consumed for operation, reduce latency (e.g., the distance between computing components is shorter), and provide greater scalability. Herein, a computing component may be defined as an electronic circuit formed on a substrate (e.g., a semiconductor substrate). A computing component may also be referred to as a “chip” or a “chiplet.”
[0182] As an illustrative example, FIGS. 1 A and IB show a SoW 10a, which includes a back end of the line (BEOL) 11 and computing components 12 disposed on the BEOL 11. The computing components 12 may include an input / output (VO) computing component 12a, a network interface controller (NIC) 12b, a switch 12c, a central processing unit (CPU) 12d, a graphics processor unit (GPU) 12e, and a memory module 12f. It should be noted that a semiconductor wafer may be used to facilitate fabrication of the BEOL 11 and / or assembly of the computing components 12 onto the BEOL 11. Once the BEOL 11 and computing components 12 are assembled, the wafer may be removed. As shown, the computing components 12 may be communicatively coupled via a combination of traces (also referred to as “electrical wiring”) in the BEOL 11 (see, for example, the traces 13a, 14a, and 15a). The traces may be disposed in multiple redistribution layers (RDLs) and a bridge layer. Local silicon interconnects (LSI) may further be disposed in the bridge layer. Together, the traces 13a, 14a, and 15a and the LSI may form a network of interconnections in the BEOL 11 that provides a plurality of interconnection pathways to transmit signals to and from the computingAttorney Docket No. MTSH-152WO01components 12. The term “signal” as used herein is intended to cover any electrical signal transmitted by the interconnection pathways including, but not limited to, a data signal, a power signal, a ground signal, and the like.
[0183] The SoW 10a may include a chip connector 18 and a power module 19. The chip connector 18 may provide an electrical connection point to an external electrical component and / or to another SoW. The power module 19 may include a voltage converter that modulates an external supply voltage and provides to the computing components 12 a chip supply voltage at an appropriate level to power some or all of the computing components 12. As shown in FIG. IB, the chip connector 18 and the power module 19 may be disposed onto the BEOL 11, e.g., the bump layer, in connection with the network of interconnections.
[0184] The SoW 10a may include one or more thermal modules 17 in thermal communication with the computing components 12. In some implementations, the thermal module 17 may be disposed in direct physical contact with the computing components 12 as shown in FIG. IB. Each thermal module 17 is configured to dissipate heat generated by the SoW 10a and, in particular, the computing components 12 during operation. The thermal module 17 may include, but is not limited to, a fluid cooler (e.g., an air and / or liquid cooler), one or more cooling fins, a heat exchanger, and / or other cooling systems. A liquid cooler may include, for example, an immersion cooling system, such as a two-phase immersion cooling system. Examples of thermal modules, which may be used in the present technology, are described in Patent Cooperation Treaty (PCT) Application No. PCT / US2023 / 067058 filed on May 16, 2023 and entitled, “Electronic Package Construction for Immersion Cooling of Integrated Circuits,” and PCT Application No. PCT / US2024 / 055034 filed on November 8, 2024 and entitled, “Systems for Thermal and Power Management in Three-Dimensional Integrated Circuits.” Each of the foregoing applications are incorporated herein by reference in their entirety for all purposes.
[0185] Additional examples and features of SoWs, which may be used in the present technology, are described in PCT Application No. PCT / US2024 / 059364 filed on December 10, 2024, and entitled “WAFER-LEVEL HETEROGENEOUS INTEGRATION FOR ARTIFICIAL INTELLIGENCE ACCELERATORS AND SYSTEMS,” which is incorporated herein by reference in its entirety.
[0186] In conventional SoWs, the BEOL is typically designed for a specific combination and arrangement of computing components. Accordingly, if a computing component is changed,Attorney Docket No. MTSH-152WO01e.g., replaced with a computing component providing a different function, it’s often necessary to redesign the BEOL, which can result in increased design time and greater complexity. Moreover, the lack of flexibility in the interconnections can lead to potential performance bottlenecks, especially with dynamically changing workloads commonly encountered in Al training and Al inference.
[0187] As an illustrative example, FIG. IB shows a cross-section of the SoW 10a, which includes multiple GPUs 12e communicatively coupled to the BEOL 11 via a combination of traces and local silicon interconnects (LSI). FIG. 2 shows a cross-section of a SoW 10b along the same plane as the SoW 10a in FIG. IB. As shown, the SoW 10b includes a different combination of computing components 12 compared to the SoW 10a, such as an input / output (I / O) computing component 12a, a GPU 12e, a high bandwidth memory (HBM) computing component or memory module 12f, and a CPU 12d. As shown, the traces in the first RDL, the bridge layer, and a portion of the second RDL (see, for example, the traces 13b, 14b, and 15b) and the arrangement of the LSI in the bridge layer are changed to accommodate the different combination of computing components in the SoW 10b.1.2. A SoW with a Programmable Logic Circuit
[0188] To address some or all of the limitations of conventional SoWs, various inventive implementations of a SoW are disclosed herein incorporating one or more programmable logic circuits. The programmable logic circuits may be integrated into respective computing components of the SoW and communicatively coupled to a network of interconnections. The network of interconnections may provide interconnection pathways to transmit signals between different computing components in the SoW and / or between a computing component of the SoW and an external computing component, e.g., communicatively coupled to the SoW via a ball grid array. In some implementations, the network of interconnections may be formed in the BEOL of the SoW from one or more interconnects, such as a trace or a local silicon interconnect (LSI).
[0189] The programmable logic circuit provides a way to dynamically reroute signals transmitted to and from a computing component while requiring little or, in some instances, no physical modifications to the network of interconnections. Thus, in some implementations, the design of the BEOL and, by extension, the design of the network of interconnections, may be used across multiple SoWs to support different combinations and / or arrangements of computing components. In this manner, the programmable logical circuits may appreciablyAttorney Docket No. MTSH-152WO01reduce the time to design the BEOL when different combinations and / or arrangements of computing components are used in the SoW. Additionally, by incorporating the programmable logic circuits into the computing components, the flexibility afforded by the programmable logic circuits may readily scale to SoWs of varying size and / or complexity (e.g., more computing components).
[0190] The programmable logic circuits may also improve the operation of the SoW. For instance, the programmable logic circuits of the computing components may provide a way to dynamically allocate computational resources within the SoW to different computing components (e.g., CPUs, GPUs) to facilitate more efficient utilization of computational resources available in the SoW as workloads managed by the SoW change over time. For example, the programmable logic circuits for a computing component may dynamically change the number of interconnection pathways available to that computing component. This change in the number of interconnection pathways, in turn, may directly affect the number and / or type of computing components communicatively coupled to that computing component. For instance, the SoW may include multiple CPUs that each include a programmable logic circuit. Each respective programmable logic circuit of the CPUs may operate in coordination with each other to dynamically change the number of memory modules and / or GPUs connected to each CPU.[01911 In some implementations, the programmable logic circuit of a computing component may be electrically coupled to a plurality of interconnects defining a plurality of interconnection pathways. To dynamically reroute signals to and from the computing component, the programmable logic circuit may selectively activate and / or deactivate different interconnection pathways connected to the computing component.|0192| For example, the SoW may include a CPU with a programmable logic circuit where the programmable logic circuit is connected to a plurality of memory modules via corresponding interconnection pathways. During operation, it may be desirable to provide the CPU access to a relatively large number of memory modules available in the SoW and / or to more interconnection pathways to the memory modules to increase the memory bandwidth available to that CPU, e.g., to facilitate execution of a relatively large workload. Accordingly, the programmable logic circuit may activate interconnection pathways connected to the plurality of memory modules, thus increasing the number of interconnection pathways available to the CPU. Alternatively, it may be desirable to allocate computational resources associated with the memory modules to another computing component of the SoW. Accordingly, theAttorney Docket No. MTSH-152WO01programmable logic circuit may deactivate interconnection pathways connected to the plurality of memory modules, thus decreasing the number of interconnection pathways connected to the plurality of memory modules.
[0193] In another example, the programmable logic circuit may provide a way to integrate a computing component onto a BEOL while requiring little or, in some instances, no changes to the BEOL itself. As an illustrative example, a first SoW may include a first computing component mounted to a portion of the BEOL of the first SoW such that the programmable logic circuit of the first computing component is connected to a plurality of interconnects in the BEOL. The first computing component may only use a first subset of the plurality of interconnects for operation. Accordingly, the programmable logic circuit of the first computing component may activate the interconnection pathways in the first subset and deactivate all remaining interconnection pathways that are unused by the first computing component. A second SoW sharing the same BEOL design as the first SoW may include a second computing component mounted to the same portion of the BEOL as the first computing component in the first SoW. The programmable logic circuit of the second computing component may be further connected to the same plurality of interconnects as the first computing component in the first SoW. The second computing component may only use a second subset of the plurality of interconnects for operation. The second subset may be different from the first subset. Accordingly, the programmable logic circuit of the second computing component may activate the interconnection pathways in the second subset and deactivate all remaining interconnection pathways that are unused by the second computing component.
[0014] The programmable logic circuit may be electrically coupled to a functional circuit block in the computing component, e.g., to transmit signals between the functional circuit block and the plurality of interconnection pathways. Herein, a functional circuit block refers to electronic circuity designed to perform one or more specific functions associated with the computing component. For example, if the computing component is a CPU, the functional circuit block may include electronic circuitry that facilitates execution of one or more functions typically associated with CPUs. The electronic circuitry in this example may include, but is not limited to, an arithmetic logic unit (ALU), a control unit (CU), a register, and the like. In another example, if the computing component is a GPU, the functional circuit block may include electronic circuity that facilitates execution of one or more functions typically associated with GPUs. The electronic circuitry in this example may include, but is not limited to, a multiprocessor, a register, shared memory, and the like. In yet another example, if theAttorney Docket No. MTSH-152WO01computing component is a memory module, the functional circuit block may include electronic circuity that facilitates execution of one or more functions typically associated with memory modules. The electronic circuitry in this example may include, but is not limited to, a memory cell, and the like.
[0195] It should be appreciated that, in some implementations, the BEOL of the SoWs disclosed herein may include interconnection pathways arranged to accommodate different combinations of computing components. As a consequence, not all of the interconnection pathways available in the BEOL may be utilized for specific combinations of computing components. Said another way, the network of interconnections may include redundant interconnection pathways that may be unused depending on the combination and arrangement of computing components used for a given SoW.
[0196] It should also be appreciated that the SoWs disclosed herein may generally incorporate any component and / or feature of the SoWs 10a and 10b shown in FIGS. 1A, IB, and 2 and described in Section 1. For brevity, a discussion of these features and / or components may not be provided below.[0197 FIGS. 3 A and 3B show an example SoW 100a that includes multiple computing components 120 with respective programmable logic circuits 170. As shown, the SoW 100a may include a layered structure with computing components 120 and a network of interconnections 130 (also referred to herein as a “network 130”) communicatively coupled to the computing components 120. The layered structure may include the computing component layer 102a, a RDL 102b disposed on the computing component layer 102a, a bridge layer 102c disposed on the RDL 102b, a RDL 102d disposed on the bridge layer 102c, and a bump layer 102e disposed on the RDL 102d. The RDL 102b, the bridge layer 102c, the RDL 102d, and the bump layer 102e together form a BEOL 110. It should be appreciated that one or more wafers may be used to facilitate fabrication of the SoW 100a (see, for example, the method 200 in FIG. 6). However, the wafer may be removed from the final SoW 100a.
[0198] The computing component layer 102a may contain the computing components 120 of the SoW 100a. The network of interconnections 130 may be formed from traces distributed in the RDLs 102b and 102d and the bridge layer 102c. Specifically, the network of interconnections 130 may include traces 132 disposed in the RDL 102b, traces 142 disposed in the bridge layer connected to the traces 132, traces 152 disposed in the RDL 102d connected to the traces 142, contact pads 162 (also referred to as “under bump metallization 162” orAttorney Docket No. MTSH-152WO01“UBM 162”) disposed in the bump layer 102e connected to the traces 152 and solder balls 166. The solder balls 166 may provide an electrical connection to an external electronic device, such as a PCB with other computing components in a computing system (e.g., a server) or an auxiliary component. In some implementations, the network of interconnections 130 may further include one or more local silicon interconnects (LSIs) 146 disposed in the bridge layer 102c. In some implementations, the traces 132, 142, and 152 may vary in size. For example, the traces 132 may be smaller (e.g., in width) compared to the traces 142 and the traces 142 may be smaller (e.g., in width) compared to the traces 152.
[0199] It should be appreciated that the layered structure of the SoW 100a is a non-limiting example. More generally, the SoWs disclosed herein may include a computing component layer, one or more redistribution layers (RDLs), a bridge layer, and / or a bump layer.
[0200] As shown in FIGS. 3A and 3B, the SoW 100a may include computing components 120 (also referred to as “chips 120” or “chiplets 120”) configured to perform a variety of functions during operation of the SoW 100a. For example, the computing components 120 may include, but are not limited to, a logic chip, a switch, a memory module (e.g., high-bandwidth memory (HBM)), a network interface controller (NIC), an input / output (I / O) controller, and the like. Examples of a logic chip include, but are not limited to, a graphics processing unit (GPU), a central processing unit (CPU), a data processing unit (DPU), a tensor processing unit (TPU), a systems on a chip (SoC), and the like. A memory module may include logic circuitry configured to store data. Examples of memory module includes, but are not limited to, a dynamic random access memory (DRAM) module, a static random access memory (SRAM) module, a flash memory module, a solid-state drive (SSD), a non-volatile random access memory (NVRAM) module, a read-only memory (ROM) module (e.g., a floating-gate ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), one-time programmable ROM (OTPROM), and the like), and any other suitable form of memory.
[0201] The computing components 120 of the SoW 100a may be selected in any combination and / or arrangement to meet any desired specifications for the SoW 100a. In the example shown in FIGS. 3 A and 3B, the SoW 100a may include computing components 120 selected such that the SoW 100a can function as a standalone computing system, e.g., as a server or at least a portion of a server. Examples of servers incorporating the SoW 100a include, but are not limited to, an artificial intelligence (Al) accelerator (e.g., for an Al training server, or an Al inference server), a web server, a database server, an email server, a web proxy server, a domainAttorney Docket No. MTSH-152WO01name server (DNS), an application server, and the like.
[0202] As shown, the SoW 100a may include a plurality of input / output (I / O) computing components 120a, a plurality of NICs 120b, a plurality of switches 120c, a plurality of CPUs 120d, a plurality of GPUs 120e, and a plurality of high bandwidth memory (HBM) computing components 120f. In some implementations, the computing components 120 may be arranged such that computing components 120 that are more likely to communicate with one another during operation are located closer to each other. For example, the GPUs 120e may more frequently communicate with the HBM computing components 120f. Accordingly, the GPUs 120e may be arranged to be at least partially surrounded by the HBM computing components 120f. In another example, the switches 120c and the CPUs 120d may more frequently communicate with external devices (e.g., an external computing system). Accordingly, the switches 120c and the CPUs 120d may be located near or, in some instances, surrounded by the I / O computing components 120a, which are responsible for facilitating communication between the SoW 100a and the external devices.
[0203] It should be appreciated that the combination and arrangement of computing components 120 shown in FIGS. 3A and 3B are a non-limiting example. In particular, the frequency of communication between computing components 120 is only one factor that may influence the overall design of the SoW 100a. Other factors may include, for example, the SoW 100a providing a sufficient quantity of different types of computing components 120 (e.g., a sufficient number of CPUs 120d). FIG. 4 shows another example SoW 100b that includes a different combination and arrangement of computing components 120. For instance, the partial cross-section of FIG. 3B shows the SoW 100a may include a HBM computing component 120f and multiple GPUs 120e. In comparison, FIG. 4 shows the SoW 100b may include an VO computing component 120a, a HBM computing component 120f, a GPU 120e, and a CPU 120d along the same partial cross-section as the SoW 100a.
[0204] FIG. 4 further shows the BEOL 110 of the SoW 100b may have a similar design as the BEOL 110 of the SoW 100a. As shown, the RDL 102b of the SoW 100b may be modified compared to the RDL 102b of the SoW 100a to accommodate the different combination of computing components 120. However, the bridge layer 102c, the RDL 102d, and the bump layer 102e of the SoW 100b may be the same as the bridge layer 102c, the RDL 102d, and the bump layer 102e of the SoW 100a, respectively. The RDL 102b may be modified, for example, to facilitate interconnections between the computing components 120 of the SoW 100b (e.g., shorter interconnection pathways) in the event the programmable logic circuits of theAttorney Docket No. MTSH-152WO01computing components 120 are insufficient. In this example, the programmable logic circuits 170 of the computing components 120 may still decrease the time to redesign the BEOL 110 by reducing the extent of the changes that need to be made to the network of interconnections 130. Moreover, by keeping the respective bump layers 102e the same between the SoWs, subsequent integration of auxiliary components, such as the connector 190 or the power module 191, may be kept the same. For instance, the modifications to the BEOL 110 may be limited only to the traces 132 in the RDL 102b while the traces 142 in the bridge layer 102c and / or the traces 152 in the RDL 102d remain unchanged. It should be appreciated that, in some implementations, the network of interconnections 130 of the SoWs 100a and 100b may be the same.
[0205] As described above, the SoWs disclosed herein may generally include one or more computing components 120 with at least one programmable logic circuit 170. In some implementations, each computing component 120 in the SoW 100a may include a programmable logic circuit 170 as shown in FIGS. 3A and 3B. However, it should be appreciated that, in some implementations, the SoW 100a may include a combination of computing components that include a programmable logic circuit 170 and computing components that do not include any programmable logic circuit. For the computing components 120 that do not include a programmable logic circuit 170, their connections to the network of interconnections 130 may be fixed.
[0206] The programmable logic circuit 170 may control communication between the computing component 120 and the BEOL 110. For instance, the computing components 120 disclosed herein may generally include a functional circuit block. The programmable logic circuit 170 may be communicatively coupled to the functional circuit block and at least a subset of the interconnection pathways provided by the network of interconnections 130. The programmable logic circuit 170 may control communications with the BEOL 110 by selectively activating or deactivating various interconnection pathways in the subset. For example, the programmable logic circuit 170 may redirect transmission of a signal from a first interconnection pathway to a second interconnection pathway by activating the second interconnection pathway and deactivating the first interconnection pathway.
[0207] The programmable logic circuit 170 may include a component capable of actively redirecting a signal from a first interconnection pathway to a second interconnection pathway. In some implementations, the programmable logic circuit 170 may be electronically controllable. For example, the programmable logic circuit 170 may include one or moreAttorney Docket No. MTSH-152WO01transistors (e.g., millions of transistors). In another example, the programmable logic circuit 170 may include one or more gate arrays, e.g., formed from one or more transistors. In yet another example, the programmable logic circuit 170 may include a field-programmable gate array (FPGA). In some implementations, the programmable logic circuit 170 may include a micro-electromechanical systems (MEMS) based interconnect. For example, a MEMS-based switch may provide a way to physically activate or deactivate interconnection pathways connected to a particular computing component 120 with a programmable logic circuit 170.(0208] The computing components 120 disclosed herein may include one or more programmable logic circuits 170. The number of programmable logic circuits 170 for a given computing component 120 may depend, in part, on the function of the computing component 120. For example, some computing components 120 may require more interconnection pathways for operation, such as a CPU or a GPU. Accordingly, CPUs and GPUs implemented as respective computing components 120 may include multiple programmable logic circuits 170 to accommodate the greater number of interconnection pathways used by these computing components 120. In some implementations, the overall size of the programmable logic circuits 170 may vary between computing components 120. Referring to the example of CPUs and GPUs implemented as respective computing components 120, these computing components 120 may incorporate relatively large programmable logic circuits 170 (e.g., programmable logic circuits 170 that span the width of the substrate).(0209] It should also be appreciated that the number of interconnection pathways connected to the programmable logic circuits 170 of different computing components 120 may vary. For example, a first computing component 120 (e.g., a CPU) may be connected to a first number of interconnection pathways in the network 130 and a second computing component 120 (e.g., a switch) may be connected to a second number of interconnection pathways in the network 130. In this example, the first number and the second number may be different.(0210] The computing components 120 disclosed herein may generally be connected to at least a subset of the interconnection pathways provided by the network 130. In some implementations, the programmable logic circuit 170 may be connected to all the interconnection pathways in the subset. In other words, any signal transmitted to or from the computing component 120 goes through the programmable logic circuit 170. In some implementations, the programmable logic circuit 170 may not be connected to all the interconnection pathways in the subset. Said another way, the programmable logic circuit 170 may be connected to a fraction of the subset. In some implementations, the computingAttorney Docket No. MTSH-152WO01component 120 may include connections that are not managed by the programmable logic circuit 170. For example, the computing component 120 may include some connections that connect the functional circuit block directly to one or more of the interconnection pathways in the subset.[02111 The computing component 120 containing the programmable logic circuit 170 may have a square or rectangular shape. In some implementations, the computing component 120 containing the programmable logic circuit 170 may have dimensions (e.g., a length, a width) ranging from about 10 mm to about 30 mm, including all sub-ranges and values in between.
[0212] The computing components 120 of the SoW 100a may be fabricated separately and then integrated into the SoW 100a. For example, the computing components 120 disclosed herein may include a substrate and the programmable logic circuit 170 may be disposed on the substrate. In some implementations, the substrate may be made of a semiconductor including, but not limited to, silicon, gallium arsenide, sapphire, silicon carbine, indium phosphide, gallium nitride, germanium, and the like. The computing component 120 may further include a functional circuit block that provides one or more functions associated with that computing component 120. The functional circuit block may also be disposed on the substrate. In some implementations, the programmable logic circuit 170 and the functional circuit block may be formed on the substrate together. For example, the programmable logic circuit 170 and the functional circuit block may be fabricated simultaneously using one or more semiconductor device manufacturing processes, e.g., deposition, lithography, etching, and so on.[02.13] Once fabricated, the computing components 120 may be mounted directly to the BEOL 110. This mounting may be accomplished, for example, using an epoxy material. For example, FIG. 3A shows the computing components 120 may be secured to the BEOL 110 via an epoxy material 122 (or a reflow material), which is shown disposed between the computing components 120.
[0214] The separate fabrication and integration of the computing components 120 may provide more flexibility in the design and manufacture of the computing components 120. For example, the computing components 120 may be fabricated using different process nodes. In another example, the BEOL 110 may be fabricated using different process nodes compared to the computing components 120. A process node may generally refer to a semiconductor manufacturing process with proprietary design rules and is often classified based on the size of the smallest features that can be manufactured (e.g., a 3 nm process node, a 5 nm process node,Attorney Docket No. MTSH-152WO01and so on).
[0215] Additionally, the separate manufacture of the computing components 120 in the SoW 100a may appreciably improve the yield of the SoW 100a. The manufacture of electronic components is often an imperfect process. Manufacturing defects can occur resulting in defective electronic components. Although careful control of process conditions can appreciably reduce the failure rate and increase yield, it is often challenging to fully eliminate manufacturing defects. For large format computing systems, such as SoWs, monolithic fabrication is often challenging due to their sensitivity to manufacturing defects. Accordingly, the separate fabrication of the computing components 120 with the programmable logic circuit 170 provides an opportunity to screen and select non-defective computing components 120 for integration into the SoW 100a, thus increasing yield.
[0021] In some implementations, the BEOL 110 may be formed in the shape of a wafer. For example, the BEOL 110 may be fabricated on a sacrificial wafer and retain the same geometry as the wafer when the wafer is removed. The shape and size of the BEOL 110 may determine the shape and size of the SoW 100a. Generally, the BEOL 110 may have various geometries including, but not limited to, a circle, an oval, a triangle, a square, a rectangle, a polygon, and any combinations of the foregoing. In some implementations, the characteristic width, £>, of the BEOL 110 may be equal to about 300 millimeters (mm). More generally, the characteristic width, £>, of the BEOL 110 may generally range from about 25 mm to about 450 mm, including any sub-ranges and values in between. In some implementations, the BEOL 110 may have a thickness, / , equal to about 1 mm. More generally, the thickness, / , of the BEOL 110 may generally range from about 0.3 mm to about 2 mm, including any sub-ranges and values in between.
[0217] It should be appreciated that the computing components 120 may be disposed onto only a portion of the BEOL 110, particularly when the BEOL 110 is shaped as a wafer. Said another way, the computing components 120 may not be distributed uniformly across the BEOL 110. In some implementations, the network of interconnections 130 may be localized to where the computing components 120 are located. Thus, the outer edges of the BEOL 110 may primarily comprise an electrical insulator with few or, in some instances, no interconnections. In some implementations, the network of interconnections 130 may extend across the width of the BEOL 110 even when the computing components 120 are localized to a portion of the BEOL 110. Said another way, the network of interconnections 130 may extend to the outer edges of the BEOL 110. The extension of the network 130 to the outer edges of the BEOL 110 mayAttorney Docket No. MTSH-152WO01facilitate connection with, for example, physical connectors (see, e.g., the connectors 190).
[0218] As described above, the network 130 may be formed from the traces in the RDLs 102b and 102d, the bridge layer 102c, and the bump layer 102e and / or LSI 146 in the bridge layer 102c. As shown in FIG. 3B, the traces may be surrounded by an electrical insulator providing electrical isolation between the traces, thus reducing interference. In the SoW 100a, the traces 132 may be surrounded by an electrical insulator 134 in the RDL 102b. The traces 142 and the LSI 146 may be surrounded by an electrical insulator 144 in the bridge layer 102c. The traces 152 may be surrounded by an electrical insulator 154 in the RDL 102d. The contact pads 162 may be surrounded by an electrical insulator 164 in the bump layer 102e.[0219| The traces 132, 142, and 152 and the contact pads 162 may be formed from an electrically conductive material including, but not limited to, aluminum, copper, gold, and the like. The electrical insulators 134, 144, 154, and 164 may be formed from an electrically insulating material including, but not limited to, silicon dioxide, aluminum oxide, hafnium oxide, a polymer, and the like. In some implementations, the traces 132, 142, and 152 and the contact pads 162 may be electrically connected via electrical vias. In some implementations, the RDLs 102b and 102d and the bridge layer 102c may each include multiple layers of traces connected via vias.
[0220] In some implementations, the traces may increase in width from the RDL 102b to the bump layer 102e to accommodate higher electrical currents, e.g., to facilitate signal transmission across the SoW 100a. In some implementations, the traces may follow a Manhattan routing scheme where respective layers of traces may alternate between a horizontal orientation and a vertical orientation when the SoW 100a is viewed from a top-down view.
[0221] The LSI 146 in the bridge layer 102c may provide high-density and high-bandwidth interconnections. The LSI 146 may be formed as a chip with high density traces designed to facilitate interconnections between the computing components 120 in the SoW 100a. Accordingly, the LSI 146 may include its own contact pads and / or solder balls to facilitate connection with the traces 132, 142, and 152. In some implementations, the electrical insulator 144 of the bridge layer 102c may be an epoxy material similar to the epoxy material 122.
[0222] In some implementations, the SoW may include one or more LSIs to facilitate connections between the computing components 120. For example, the LSIs 146 may be arranged in the bridge layer 102c to provide interconnection pathways that communicatively couple computing components 120 disposed adjacent to one another. As an illustrativeAttorney Docket No. MTSH-152WO01example, FIG. 5 shows a SoW 100c that includes the same computing components 120 as the SoW 100a, but with LSIs 146 arranged between various computing components 120. In this example, the LSIs 146 may replace the traces communicatively coupled to the computing components 120 in the SoW 100a as shown in FIG. 3A. Compared to the traces used in the SoW 100a, the LSIs 146 may provide a higher density of interconnections (e.g., a greater number of interconnections per unit volume or per unit area).
[0223] The contact pads 162 and the solder balls 166 of the bump layer 102e may facilitate communication between the SoW 100a and an external electronic device. In some implementations, the network of interconnections 130 may include or define an Integrated Fan-Out (InFO) packaging connection. In one example, the external electronic device may include a PCB with other computing components. In another example, the SoW 100a may include one or more auxiliary components mounted onto the bump layer 102e and electrically coupled to the solder balls 166. For example, the auxiliary components may include a power module 191 with a voltage converter to module an external supply voltage to power some or all of the computing components 120. In another example, the auxiliary components may include a chip connector 190 to facilitate connection with another electronic device (e.g., another SoW) via cable.
[0224] In some implementations, the SoW 100a may include one or more thermal modules 192 disposed on the computing component layer 102a. The thermal modules 192 may facilitate the removal and dissipation of heat generated, for example, by the computing components 120 in the computing component layer 102a.1.3. An Example Method for Making a SoW with Programmable Interconnects
[0225] The SoWs disclosed herein (e.g., SoWs lOOa-lOOc) may be manufactured using various packaging processes including, but not limited to, a chip-first manufacturing process and a chip-last manufacturing process.
[0226] In a chip-first manufacturing process, the computing components 120 may first be mounted to a wafer and secured to the wafer via, for example, the epoxy material 122. The epoxy material 122 may thereafter be planarized. As described above, the computing components 120 and the epoxy material 122 may form the computing component layer 102a. The subsequent layers of the SoW (e.g., the RDLs 102b and 102d, the bridge layer 102c, and the bump layer 102e) may then be formed directly on the computing component layer 102a. For example, the traces in the network of interconnections 130 and / or the surrounding electricalAttorney Docket No. MTSH-152WO01insulators may be formed, in part, using various semiconductor manufacturing processes (e.g., chemical vapor deposition, physical vapor deposition, photolithography, dry etching, wet etching). Once the network 130 is formed, the solder balls 166 may be mounted onto corresponding electrical contact pads 162. Then, one or more auxiliary components may be connected to one or more of the solder balls 166. Lastly, the wafer supporting the computing components 120 may be removed.10227] FIG. 6 shows a non-limiting example of a chip-last manufacturing method 200 to form the SoWs disclosed herein. As shown, the layers of the SoW forming the network of interconnections 130 (e.g., the RDLs 102b and 102d, the bridge layer 102c, and the bump layer 102e) may be formed on a sacrificial wafer first at step 202. For example, the traces in the network of interconnections 130 and / or the surrounding electrical insulators may be formed, in part, using various semiconductor manufacturing processes (e.g., chemical vapor deposition, physical vapor deposition, photolithography, dry etching, wet etching).
[0228] At step 204, the computing components 120 may be mounted to the RDL 102b of the BEOL 110. In some implementations, the computing components 120 may be connected to the traces 132 in the RDL 102b via bump bonding and / or hybrid bonding during this step. At step 206, the epoxy material 122 may be applied to securely couple the computing components 120 to the BEOL 110 and / or to protect the computing components 120. Once the epoxy material 122 is applied, a planarization process is applied to flatten the top surface of the epoxy material 122. At step 208, the sacrificial carrier wafer may be removed (e.g., via chemical etching, dry etching), thus exposing, for example, the electrical contact pads 162. The exposure of the contact pads 162, in turn, may allow solder balls 166 to be mounted onto the electrical contact pads 162 and, thereafter, any auxiliary components for the SoW.
[0229] It should be appreciated that the method 200 is a non-limiting example of a chip-last manufacturing process and alterations to the method 200 are readily contemplated herein. For example, the steps of method 200 may be performed in any order. For example, the bridge layer 102c, the RDL 102d, and the bump layer 102e may be formed on a first sacrificial wafer. However, the computing components 120 may be mounted to a second sacrificial wafer and the RDL 102b may be formed directly thereon. Thereafter, the second sacrificial wafer (with the computing components 120 and the RDL 102b) may be connected to the bridge layer 102c disposed on the first sacrificial carrier wafer (e.g., via bump bonding or the like). The sacrificial carrier wafers may be removed thereafter.Attorney Docket No. MTSH-152WO01
[0230] The sacrificial wafer discussed above may generally provide mechanical support for the other components of the SoWs disclosed herein (e.g., SoW 100a, 100b, 100c). In some implementations, the wafer may provide no electrical function. Said another way, the wafer may not transmit any signal within the SoW 100a. The wafer may be formed from a semiconductor including, but not limited to, silicon, gallium arsenide, sapphire, silicon carbine, indium phosphide, gallium nitride, germanium, and / or another semiconducting material.2. An Example System-on- A- Wafer (SOW) with Programmable Interconnects
[0231] The Inventors have also recognized and appreciated traditional computing architectures are unable to scale with the large amounts of data processed in modern Al models due to limitations on power efficiency and memory bandwidth. For example, computing systems based on the von Neumann architecture typically include a processor (e.g., central processing unit (CPU), a graphics processing unit (GPU)) and memory that communicate via relatively long electrical interconnects integrated in a printed circuit board (PCB). The transmission of large amounts of data in these systems is limited by the power required to transmit data and the latency of transmission. Furthermore, the PCB -integrated format often results in computing systems with a relatively large physical footprint, which limits the computational density attainable for the overall system. For example, the PCB of a server may have an area that is two times larger than the area of the constituent computing components mounted to the PCB.
[0232] To address these foregoing limitations, the Inventors have invented computing systems with heterogeneous computing components (e.g., GPUs, central processing units (CPUs), switches, network interface cards (NICs), memory, and the like) integrated onto a back end of line (BEOL). In some implementations, the BEOL may be shaped like a semiconductor wafer. These computing systems are referred to herein as a SoW. The SoW architecture offers a convenient way to design and build computing systems in a highly modular manner. For example, an SoW may include different combinations of computing components to meet the requirements of different applications (e.g., high memory bandwidth, low latency, high compute density, low power consumption, and the like). Moreover, the SoW architecture facilitates the integration of computing components fabricated using different process nodes, thus reducing manufacturing cost.
[0233] As an illustrative example, FIGS. 7A and 7B show a SoW 70a, which includes a BEOL 71 and computing components 72 (e.g., GPUs 72a and switches 72b) disposed on the BEOL 71. It should be noted that a semiconductor wafer may be used to facilitate fabrication of theAttorney Docket No. MTSH-152WO01BEOL 71 and / or assembly of the computing components 72 onto the BEOL 71. Once the BEOL 71 and computing components 72 are assembled, the wafer may be removed. Herein, a computing component is also referred to as a “chip” or a “chiplet.” As shown, the computing components 72 may be communicatively coupled via a combination of traces (also referred to as “electrical wiring”) in the BEOL 71 (see, for example, the traces 73a, 74a, and 75a). The traces may be disposed in multiple redistribution layers (RDLs) and a bridge layer and / or local silicon interconnects (LSI) 76 in the bridge layer. The integration of multiple computing components directly onto a wafer-shaped BEOL provides a way to significantly reduce the physical distance between computing components, especially processing and memory, resulting in higher density, improved power efficiency, and greater memory bandwidth. SoWs which may be used in the present technology are described in PCT Application No. PCT / US2024 / 059364 filed on December 10, 2024, and entitled “WAFER-LEVEL HETEROGENEOUS INTEGRATION FOR ARTIFICIAL INTELLIGENCE ACCELERATORS AND SYSTEMS,” which is incorporated herein by reference in its entirety.
[0234] The SoW architecture offers a convenient way to design and build computing systems in a highly modular manner. For example, a SoW may include different combinations of computing components to meet the requirements of different applications (e.g., high memory bandwidth, low latency, high compute density, low power consumption, and the like). Moreover, the SoW architecture facilitates the integration of computing components fabricated using different process nodes, thus reducing manufacturing cost.
[0235] However, the Inventors have also recognized the design of the computing components and the manner in which they electrically interface with the electrical wiring of the SoW are typically different for computing components that perform different functions and often change over time as improvements are made to the design of the computing components. Accordingly, one challenge with SoWs is providing an electrical interface that can readily accommodate different computing components while maintaining efficient routing of interconnection pathways and interconnectivity between the computing components. Conventional interconnections, such as the traces in RDLs and LSI, are passive, i.e., the interconnections are static and fixed over time. As a result, if a computing component in the SoW is replaced with another computing component having a different design, significant modifications to the network of interconnections within the SoW are often required. The lack of flexibility in the interconnections can result in increased design time, greater complexity, and potentialAttorney Docket No. MTSH-152WO01performance bottlenecks, especially with dynamically changing workloads commonly encountered in Al training and Al inference.(0236] As an illustrative example, FIG. 7B shows a cross-section of the SoW 70a, which includes multiple GPUs 72a communicatively coupled via a combination of interconnects 77 and LSI 76. FIG. 8 shows a cross-section of a SoW 70b along the same plane as the SoW 70a in FIG. 7B. As shown, the SoW 70b includes a different combination of computing components 72 compared to the SoW 70a, such as a GPU 72a, an input / output (I / O) computing component 72c, a high bandwidth memory (HBM) computing component 72d, and a CPU 72e. As shown, the traces in the first RDL, the bridge layer, and a portion of the second RDL (see, for example, the traces 73b, 74b, and 75b) and the arrangement of the LSI 76 in the bridge layer are changed to accommodate the different combination of computing components in the SoW 70b.(0237] However, the Inventors have also recognized the design of the computing components and the manner in which they electrically interface with the electrical wiring of the SoW are typically different for computing components that perform different functions and often change over time as improvements are made to the design of the computing components. Accordingly, one challenge with SoWs is providing an electrical interface that can readily accommodate different computing components while maintaining efficient routing of interconnection pathways and interconnectivity between the computing components. Conventional interconnections, such as the traces in RDLs and LSI, are passive, i.e., the interconnections are static and fixed over time. As a result, if a computing component in the SoW is replaced with another computing component having a different design, significant modifications to the network of interconnections within the SoW are often required. The lack of flexibility in the interconnections can result in increased design time, greater complexity, and potential performance bottlenecks, especially with dynamically changing workloads commonly encountered in Al training and Al inference.(0238] In view of the foregoing limitations of passive interconnections in SoWs, the present disclosure is directed to various inventive implementations of a SoW that incorporates one or more programmable interconnects. The programmable interconnects provide a way to dynamically reconfigure the interconnection pathways within the SoW. For example, the network of interconnections in the SoW may be adjusted to accommodate different computing component configurations without requiring a physical change to the arrangement and / or layout of the interconnection pathways. In another example, the network of interconnections in the SoW may be adjusted to allocate computational resources to different computingAttorney Docket No. MTSH-152WO01components (e.g., a particular CPU or GPU) in response to dynamically changing workloads managed by the SoW. In this manner, the programmable interconnects may provide greater flexibility, performance, and / or adaptability to facilitate the use of different combinations of computing components and / or support dynamic workloads without requiring extensive redesign of the interconnection network.(0239] In one aspect, the programmable interconnects may provide a way to selectively transmit a signal along one interconnection pathway out of multiple available interconnection pathways within the SoW. In some implementations, the programmable interconnects may include logic circuitry. For example, the programmable interconnects may include a transistor. In another example, the programmable interconnects may include a gate array. In yet another example, the programmable interconnects may include a field programmable gate array (FPGA).(0240] In another aspect, the SoW may have a layered structure. The layers may include, but are not limited to, a computing component layer containing one or more computing components, a RDL layer containing traces, a bridge layer containing traces and / or LSI, and a bump layer containing solder balls (also referred to as “solder bumps”). The programmable interconnects may be incorporated into one or more layers of the SoW. In one example, the programmable interconnects may be implemented as a computing component that is incorporated within the computing component layer. In another example, the programmable interconnects may be implemented in the bridge layer. In some implementations, the bridge layer may further include LSI with the programmable interconnects. In yet another example, the programmable interconnects may be implemented in the bump layer as a separate device mounted to one or more solder balls in the bump layer. In yet another example, the programmable interconnects may be incorporated in a combination of the foregoing layers.[02411 In yet another aspect, the SoWs disclosed herein may be manufactured via a chip-last manufacturing process (also referred to as a “chip-on-wafer process”). In the chip-last manufacturing process, the SoW may be manufactured starting from the bump layer and ending at the computing component layer using multiple sacrificial wafers to support the fabrication of the bump layer, the RDLs, and / or the bridge layer. Once the computing component layer is formed, a carrier wafer may be bonded to the computing component layer. Thereafter, the sacrificial wafers may be removed to provide, for example, access to the bump layer, e.g., for electrical connection with a PCB.Attorney Docket No. MTSH-152WO01
[0242] In one example implementation, a system-on-a-wafer (SoW) includes a back end of the line (BEOL), including a plurality of interconnects; a plurality of computing components, disposed on the BEOL, each computing component of the plurality of computing components including a substrate; and a circuit disposed on the substrate; and at least one programmable interconnect communicatively coupled to the plurality of interconnects, wherein the plurality of interconnects and the at least one programmable interconnect provide a plurality of interconnection pathways communicatively coupled to respective computing components of the plurality of computing components; and the at least one programmable interconnect is configured to change at least one interconnection pathway of the plurality of interconnection pathways.
[0243] The BEOL may be shaped as a semiconductor wafer. The plurality of computing components may include at least one first computing component; and at least one second computing component different from the at least one first computing component. The plurality of computing components may include at least one of a graphics processing unit (GPU); a central processing unit (CPU); a memory module; a switch; an input / output (I / O) controller; or a network interface controller (NIC). The at least one programmable interconnect may include a transistor. The at least one programmable interconnect may include a gate array. The at least one programmable interconnect may include a field programmable gate array (FPGA). The at least one programmable interconnect may include a micro-electromechanical systems (MEMS) switch. The plurality of computing components may include a first computing component containing the at least one programmable interconnect. The plurality of interconnects may include a first plurality of interconnects disposed in a first redistribution layer of the BEOL and communicatively coupled to the plurality of computing components; a second plurality of interconnects disposed in a bridge layer of the BEOL and communicatively coupled to the first plurality of interconnects; and a third plurality of interconnects disposed in a second redistribution layer of the BEOL and communicatively coupled to the second plurality of interconnects. The at least one programmable interconnect may be disposed in the bridge layer of the BEOL. At least one local silicon interconnect (LSI) may be further disposed in the bridge layer of the BEOL. The SoW may further include a plurality of solder balls disposed in a bump layer of the BEOL, wherein the plurality of solder balls is communicatively coupled to the third plurality of interconnects. The at least one programmable interconnect may be disposed in the bump layer of the BEOL. The plurality of computing components may be configured to operate as an artificial intelligence (Al) accelerator.Attorney Docket No. MTSH-152WO01
[0244] In another example implementation, a SoW includes a computing component layer, the computing component layer including a plurality of computing components, each computing component of the plurality of computing components including a substrate; and a circuit integrally formed on the substrate; a back end of the line (BEOL), including a first redistribution layer disposed on the computing component layer, the first redistribution layer including a first plurality of interconnects communicatively coupled to the plurality of computing components; a bridge layer disposed on the first redistribution layer, the bridge layer including a second plurality of interconnects communicatively coupled to the first plurality of interconnects; a second redistribution layer disposed on the bridge layer, the second redistribution layer including a third plurality of interconnects communicatively coupled to the second plurality of interconnects; and a bump layer disposed on the second redistribution layer, the bump layer including a plurality of solder balls communicatively coupled to the third plurality of interconnects; and at least one programmable interconnect disposed in at least one of the computing component layer, the bridge layer, or the bump layer, wherein the BEOL is shaped as a semiconductor wafer.
[0245] The bridge layer may further include at least one local silicon interconnect (LSI).
[0246] Following below are various inventive implementations of a SoW incorporating one or more programmable interconnects. The programmable interconnects are integrated into a network of interconnections that communicatively coupled to a plurality of computing components in the SoW. The network of interconnections provides interconnection pathways to transmit signals between different computing components and / or between a computing component in the SoW and an external computing component, e.g., communicatively coupled to the SoW via a ball grid array. In addition to the programmable interconnects, the network may generally include passive interconnects and / or local silicon interconnects (LSI).
[0247] The programmable interconnects provide a way to dynamically reconfigure the interconnection pathways within the SoW without requiring physical modifications to the entire network of interconnections. Thus, the same network of interconnections may be readily adapted to support different arrangements of computing components and / or different combinations of computing components in the SoW (collectively referred to herein as the “configuration of the computing components”). In this manner, the programmable interconnects can efficiently connect different configurations of computing components while appreciably reducing the time to redesign the interconnection network. Additionally, theAttorney Docket No. MTSH-152WO01programmable interconnect architecture is readily scalable to SoWs of varying size and complexity.(0248] The programmable interconnects may also improve the operation of the SoW. For instance, the allocation of resources to different computing components (e.g., CPUs, GPUs) may dynamically change as the workloads performed by the SoW change over time. The programmable interconnects also provide a way to dynamically and rapidly adjust the allocation of computational resources and / or reduce latency in data transmission for the computing components in the SoW.
[0249] In some implementations, the SoWs disclosed herein may be provided with an embedded instruction set to dynamically reconfigure interconnection pathways based on changes to the configuration of computing components used and / or changes in workload requirements. This may be accomplished, for example, by the instruction set including a setup of firmware parameters that modifies the settings of the programmable interconnects to reconfigure the interconnection pathways. In this manner, the SoW may not require a dedicated processor or dedicated memory to facilitate the reconfiguration of the interconnection pathways.
[0250] In some implementations, the SoW may be designed as a layered structure. For example, the SoW may include, but is not limited to, a computing component layer, one or more redistribution layers (RDLs), a bridge layer, and a bump layer. The computing component layer contains various computing components (e.g., chips) of the SoW. The RDL may contain traces of varying size. The bridge layer may contain traces and / or LSI to communicatively couple respective RDLs. The bump layer may include a plurality of solder balls to facilitate connection of the SoW to another electronic component of a computing system, e.g., a PCB with one or more other computing components.
[0251] Generally, the programmable interconnects may be integrated in one or more layers of the SoWs disclosed herein. For example, the programmable interconnects may be implemented as a computing component disposed within the computing component layer to facilitate direct communication with other computing components to provide finer control over specific interconnection pathways. In another example, the programmable interconnects may be integrated in the bridge layer to provide localized data flow between computing components. It should be appreciated that the programmable interconnects may be integrated in an RDL. In this example, however, the architecture is similar or, in some instances, the same as when theAttorney Docket No. MTSH-152WO01programmable interconnects are integrated in the bridge layer. In yet another example, the programmable interconnects may be integrated in the bump layer to provide comprehensive routing control over the SoW.
[0252] Following below are several examples of SoWs where the programmable interconnects are integrated into one or more layers of the SoW. It should be appreciated that the following examples are non-limiting and that the programmable interconnects disclosed in one example SoW may be readily integrated into other example SoWs with programmable interconnects.2.1 Example Programmable Interconnects as a Computing Component
[0253] FIG. 9 shows an example SoW 900a with programmable interconnects 970a (abbreviated “PI”) implemented as a computing component and integrated in a computing component layer 902a of the SoW 900a. As shown, the SoW 900a may include a layered structure with computing components 920 and a network of interconnections 930 (also referred to herein as a “network 930”) communicatively coupled to the computing components 920. The layered structure may include the computing component layer 902a, aRDL 902b disposed on the computing component layer 902a, a bridge layer 902c disposed on the RDL 902b, a RDL 902d disposed on the bridge layer 902c, and a bump layer 902e disposed on the RDL 902d. The RDL 902b, the bridge layer 902c, the RDL 902d, and the bump layer 902e together form a BEOL 910. It should be appreciated that one or more wafers may be used to facilitate fabrication of the SoW 900a (see, for example, the method 1400 in FIG. 14). However, the wafer may be removed from the final SoW 900a.
[0254] The computing components 920 are integrated into the computing component layer 902a with the programmable interconnects 970a. The network of interconnections 930 is formed from traces distributed in the RDLs 902b and 902d and the bridge layer 902c. Specifically, the network of interconnections 930 may include traces 932 disposed in the RDL 902b, traces 942 disposed in the bridge layer connected to the traces 932, traces 952 disposed in the RDL 902d connected to the traces 942, contact pads 962 disposed in the bump layer 902e connected to the traces 952 and solder balls 966. The solder balls 966 may provide an electrical connection to an external electronic device, such as a PCB with other computing components in a computing system (e.g., a server) or an auxiliary component.
[0255] The programmable interconnects 970a may include a component capable of actively redirecting a signal from a first interconnection pathway to a second interconnection pathway. In some implementations, the programmable interconnects 970a may be electronicallyAttorney Docket No. MTSH-152WO01controllable. For example, the programmable interconnects 970a may include one or more transistors (e.g., millions of transistors). In another example, the programmable interconnects 970a may include one or more gate arrays, e.g., formed from one or more transistors. In yet another example, the programmable interconnects 970a may include a field-programmable gate array (FPGA). In some implementations, the programmable interconnects 970a may include a micro-electromechanical systems (MEMS) based interconnect. For example, a MEMS-based switch may provide a way to physically change the interconnection pathways in the SoW.
[0256] In this example, the programmable interconnects 970a may be fabricated as a chip (or, alternatively, a chiplet). Thereafter, the chip may be mounted to a wafer during assembly and manufacture of the SoW 900a and, ultimately, mounted to the BEOL 910. The BEOL 910 may be separately supported by another wafer. In this manner, the programmable interconnects 970a may be designed and manufactured with fewer constraints imposed by the SoW 900a. For example, the programmable interconnects 970a may be fabricated using a different process node from the network of interconnections 930, thus providing greater flexibility in the design and manufacture of the programmable interconnects 970a. A process node may generally refer to a semiconductor manufacturing process with proprietary design rules and is often classified based on the size of the smallest features that can be manufactured (e.g., a 3 nm process node, a 5 nm process node, and so on).
[0257] Additionally, the separate manufacture of the programmable interconnects 970a from other components in the SoW 900a may appreciably improve the yield of the SoW 900a. The manufacture of electronic components is often an imperfect process. Manufacturing defects can occur resulting in defective electronic components. Although careful control of process conditions can appreciably reduce the failure rate and increase yield, it is often challenging to fully eliminate manufacturing defects. For large format computing systems, such as SoWs, monolithic fabrication is often challenging due to their sensitivity to manufacturing defects. Accordingly, the separate fabrication of the programmable interconnects 970a in the SoW 900a provides an opportunity to screen and select non-defective programmable interconnects 970a for integration into the SoW 900a, thus increasing yield.
[0258] The chip containing the programmable interconnects 970a may have a square or rectangular shape. In some implementations, the chip containing the programmable interconnects 970a may have dimensions (e.g., a length, a width) ranging from about 10 mm to about 30 mm, including all sub-ranges and values in between.Attorney Docket No. MTSH-152WO01
[0259] In some implementations, the programmable interconnects 970a may be subject to constraints related to its size, shape, and / or layout of electrical contacts to improve compatibility with the network of interconnections 930. For example, the size and / or shape of the chip containing the programmable interconnects 970a may be limited to provide sufficient space for other computing components 920 in the computing component layer 902a. In another example, the arrangement of electrical contacts on the chip may be fixed. In this manner, chips with different designs, but with the same arrangement of electrical contacts may readily connect to the network of interconnections 930 with little or, in some instances, no modifications to the network of interconnections 930. In this manner, these constraints may provide more flexibility in the selection of different computing components 920 integrated into the SoW 900a.
[0260] It should be appreciated, however, that for the SoWs disclosed herein, the network of interconnections 930 does not have to be fixed. In some implementations, the network of interconnections 930 may be modified to accommodate different chips with programmable interconnects 970a. In these implementations, the programmable interconnects 970a may still decrease redesign time by reducing the extent of the changes that need to be made to the network of interconnections 930. For example, modifications may be limited only to the traces 932 in the RDL 902b. Thus, the traces 942 in the bridge layer 902c and / or the traces 952 in the RDL 902d may remain unchanged.[02611 It should also be appreciated that, in some implementations, the SoW 900a may include multiple chips containing the programmable interconnects 970a distributed throughout the computing component layer 902a. In some implementations, the chips containing the programmable interconnects 970a may be uniformly distributed across the computing component layer 902a such that, on average, the separation distance between nearest neighbor chips is approximately constant. For example, the chips containing the programmable interconnects 970a may be distributed as an in-line array or a staggered array across the computing component layer 902a.
[0262] In some implementations, the BEOL 910 may be formed in the shape of a wafer. For example, the BEOL 910 may be fabricated on a sacrificial wafer and retain the same geometry as the wafer when the wafer is removed. The shape and size of the BEOL 910 may determine the shape and size of the SoW 900a. Generally, the BEOL 910 may have various geometries including, but not limited to, a circle, an oval, a triangle, a square, a rectangle, a polygon, andAttorney Docket No. MTSH-152WO01any combinations of the foregoing. The characteristic width of the BEOL 910 may generally range from about 25 mm to about 450 mm, including any sub-ranges and values in between.(0263] It should be appreciated that the computing components 920 may be disposed only onto a portion of the BEOL 910, particularly when the BEOL 910 is shaped as a wafer. In some implementations, the network of interconnections 930 may be localized to where the computing components 920 are located. Thus, the outer edges of the BEOL 910 may primarily comprise an electrical insulator with few or, in some instances, no interconnections. In some implementations, the network of interconnections 930 may extend across the width of the BEOL 910 even when the computing components 920 are disposed only onto a portion of the BEOL 910. Said another way, the network of interconnections 930 may extend to the outer edges of the BEOL 910. This may facilitate connection with, for example, physical connectors.(0264] The computing components 920 (also referred to as “chips 920” or “chiplets 920”) may be communicatively coupled to the chips containing the programmable interconnects 970a via the network of interconnections 930. During operation, the chip containing the programmable interconnects 970a may redirect signals between the computing components 920, for example, by directing the signal to different electrical contacts on the chip, which, in turn, are electrically connected to different computing components 920. Given the close proximity of the programmable interconnects 970a to the computing components 920, the programmable interconnects 970a may primarily reconfigure interconnection pathways for computing components 920 in close proximity to the programmable interconnects 970a. For example, the programmable interconnects 970a may primarily reconfigure interconnection pathways formed within the RDL 902b.(0265] The computing components 920 may generally provide one or more functions during operation of the SoW 900a. For example, the computing components 920 may include, but are not limited to, a logic chip, a switch, a memory module (e.g., high-bandwidth memory (HBM)), a network interface controller (NIC), an input / output (I / O) controller, and the like. Examples of a logic chip include, but are not limited to, a graphics processing unit (GPU), a central processing unit (CPU), a data processing unit (DPU), a tensor processing unit (TPU), a systems on a chip (SoC), and the like. A memory module may include logic circuitry configured to store data. Examples of memory module includes, but are not limited to, a dynamic random access memory (DRAM) module, a static random access memory (SRAM) module, a flash memory module, a solid-state drive (SSD), a non-volatile random access memory (NVRAM) module, a read-only memory (ROM) module (e.g., a floating-gate ROM, erasableAttorney Docket No. MTSH-152WO01programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), one-time programmable ROM (OTPROM), and the like), and any other suitable form of memory.
[0266] The computing components 920 may be selected in any combination to provide an overall function or architecture for the SoW 900a. In this example, the computing components 920 may include an input / output (I / O) computing component 920a, a GPU 920b, and a high bandwidth memory (HBM) computing component 920c. However, it should be appreciated that this is a non-limiting example. In some implementations, the computing components 920 may be selected such that the SoW 900a forms a server or at least a portion of a server. Examples of servers incorporating the SoW 900a include, but are not limited to, an artificial intelligence (Al) accelerator (e.g., for an Al training server, or an Al inference server), a web server, a database server, an email server, a web proxy server, a domain name server (DNS), an application server, and the like.
[0267] As discussed above with respect to the programmable interconnects 970a, the computing components 920 may be fabricated separately and then integrated into the SoW 900a. For example, the computing components 920 may generally include a substrate (e.g., a silicon substrate) and a circuit formed on the substrate (e.g., the computing component 920 may be an integrated circuit). This provides more flexibility in the design and manufacture of the computing components 920 and can further increase the yield of the SoW 900a. The computing components 920 may be mounted directly to the BEOL 910. This may be accomplished, for example, using an epoxy material. For example, FIG. 9 shows the computing components 920 may be secured to the BEOL 910 via an epoxy material 922 (or a reflow material), which is shown disposed between the computing components 920.
[0268] As described above, the interconnection network 930 may be formed from traces in the RDLs 902b and 902d, the bridge layer 902c, and the bump layer 902e and / or LSI 946 in the bridge layer 902c. As shown in FIG. 9, the traces may be surrounded by an electrical insulator providing electrical isolation between the traces, thus reducing interference. In the SoW 900a, the traces 932 may be surrounded by an electrical insulator 934 in the RDL 902b. The traces 942 and the LSI 946 may be surrounded by an electrical insulator 944 in the bridge layer 902c. The traces 952 may be surrounded by an electrical insulator 954 in the RDL 902d. The contact pads 962 may be surrounded by an electrical insulator 964 in the bump layer 902e.Attorney Docket No. MTSH-152WO01
[0269] The traces 932, 942, and 952 and the contact pads 962 may be formed from an electrically conductive material including, but not limited to, aluminum, copper, gold, and the like. The electrical insulators 934, 944, 954, and 964 may be formed from an electrically insulating material including, but not limited to, silicon dioxide, aluminum oxide, hafnium oxide, a polymer, and the like. In some implementations, the traces 932, 942, and 952 and the contact pads 962 may be electrically connected via electrical vias. In some implementations, the RDLs 902b and 902d and the bridge layer 902c may each include multiple layers of traces connected via vias.
[0270] In some implementations, the traces may increase in width from the RDL 902b to the bump layer 902e to accommodate higher electrical currents, e.g., to facilitate signal transmission across the SoW 900a. In some implementations, the traces may follow a Manhattan routing scheme where respective layers of traces may alternate between a horizontal orientation and a vertical orientation when the SoW 900a is viewed from a top-down view.
[0271] The LSI 946 in the bridge layer 902c may provide high-density and high-bandwidth interconnections. The LSI 946 may be formed as a chip with high density traces designed to facilitate interconnections between the computing components 920 in the SoW 900a. Accordingly, the LSI 946 may include its own contact pads and / or solder balls to facilitate connection with the traces 932, 942, and 952. In some implementations, the electrical insulator 944 of the bridge layer 902c may be an epoxy material similar to the epoxy material 922.
[0272] The contact pads 962 and the solder balls 966 of the bump layer 902e may facilitate communication between the SoW 900a and an external electronic device. In some implementations, the network of interconnections 930 may include or define an Integrated Fan-Out (InFO) packaging connection. In one example, the external electronic device may include a PCB with other computing components. In another example, the SoW 900a may include one or more auxiliary components (not shown) mounted onto the bump layer 902e and electrically coupled to the solder balls 966. For example, the auxiliary components may include a power module with a voltage converter to module an external supply voltage to power some or all of the computing components 920. In another example, the auxiliary components may include a chip connector to facilitate connection with another electronic device (e.g., another SoW) via cable. Examples of auxiliary components, such as power modules and chip connectors, are described in PCT Application No. PCT / US2024 / 059364 filed on December 10, 2024, and entitled “WAFER-LEVEL HETEROGENEOUS INTEGRATION FOR ARTIFICIALAttorney Docket No. MTSH-152WO01INTELLIGENCE ACCELERATORS AND SYSTEMS,” which is incorporated herein by reference in its entirety.2.2 Example Programmable Interconnects in the Bridge Layer(0273] FIG. 10 shows an example SoW 900b with programmable interconnects 970b integrated in the bridge layer 902c of the SoW 900b. The SoW 900b may include several of the same components as the SoW 900a. The SoW 900b may further operate in a similar manner as the SoW 900a. For brevity, repeated discussions of these components and / or features may not be provided below.(0274] In this example, the computing components 920 (e.g., an input / output (I / O) computing component 920a, a GPU 920b, a high bandwidth memory (HBM) computing component 920c, and a CPU 920d) may not include any programmable interconnects. Additionally, the bridge layer 902c may not include the LSI 946. Thus, the bridge layer 902c may only include the traces 942 and the programmable interconnects 970b. As shown in FIG. 10, the programmable interconnects 970b may connect to the traces 932 in the RDL 902b and the traces 952 in the RDL 902d, e.g., via electrical vias, contact pads, solder balls, and the like. Compared to the programmable interconnects 970a, the programmable interconnects 970b may facilitate the reconfiguration of interconnection pathways covering a larger portion of the SoW 900b and, thus, more computing components 920 albeit with less fine control over specific interconnection pathways near the computing components 920.(0275] In some implementations, the programmable interconnects 970b may be fabricated with the traces in the interconnection network 930 as part of the same semiconductor fabrication process. The programmable interconnects 970b may be formed as spatially localized blocks of logic circuitry within the bridge layer 902c. However, it should be appreciated that the programmable interconnects 970b may be uniformly mixed with the traces 942. In some implementations, the programmable interconnects 970b may be distributed across the bridge layer 902c (i.e., along the in-plane direction of the BEOL 910 spanning the width of the SoW 900b). In some implementations, the programmable interconnects 970b may be localized to only a portion of the bridge layer 902c. For instance, the programmable interconnects 970b may be disposed only near GPUs or CPUs in the computing component layer 902a.(0276| In some implementations, the programmable interconnects 970b may be fabricated separately as a computing component (e.g., a chip, a chiplet) and then subsequently integrated into the SoW 900b, similar to the programmable interconnects 970a of the SoW 900a. ForAttorney Docket No. MTSH-152WO01example, one or more chips containing the programmable interconnects 970b may be distributed across the bridge layer 902c. In some implementations, the chips may take the place of the LSI 946 in the SoW 900a.
[0277] However, it should be appreciated that both programmable interconnects and LSI may be incorporated into the bridge layer 902c. For example, FIG. 11 shows a SoW 900c with programmable interconnects 970b integrated in the bridge layer 902c of the SoW 900c with the LSI 946. The SoW 900c may include several of the same components as the SoWs 900a and 900b. The SoW 900c may further operate in a similar manner as the SoWs 900a and 900b.2.3 Example Programmable Interconnects in the Bump Layer
[0278] FIG. 12 shows an example SoW 900d with programmable interconnects 970c integrated in the bump layer 902e of the SoW 900d. The SoW 900d may include several of the same components as the SoWs 900a-900c. The SoW 900d may further operate in a similar manner as the SoWs 900a-900c. For brevity, repeated discussions of these components and / or features may not be provided below.
[0279] In this example, the programmable interconnects 970c may be an auxiliary component that connects to one or more solder balls 966 of the bump layer 902e. Thus, the programmable interconnects 970c may redirect signals transmitted into and out of the SoW 900d through the contact pads 962. Compared to the programmable interconnects 970a and 970b, the programmable interconnects 970c may facilitate the reconfiguration of interconnection pathways spanning a larger portion and, in some instances, the entirety of the SoW 900d. In this manner, the programmable interconnects 970c may provide more comprehensive routing control of interconnection pathways across the SoW 900b.
[0280] In some implementations, the programmable interconnects 970c may be fabricated separately as a computing component (e.g., a chip, a chiplet), similar to the programmable interconnects 970a of the SoW 900a, and then subsequently connected to the solder balls 966 in the bump layer 902e. In some implementations, the addition of the programmable interconnects 970c may occur after other components of the SoW 900d are fabricated. In other words, the various layers of the SoW 900d may be designed in a similar manner as a conventional SoW with the difference being the addition of programmable interconnects 970c on at least some of the solder balls 966.Attorney Docket No. MTSH-152WO012.4 Example Combinations of Programmable Interconnects
[0281] It should be appreciated that the programmable interconnects disclosed above in the SoWs 900a-900d are not generally limited to a single layer. Rather, programmable interconnects may generally be disposed in more than one layer of the SoW. FIG. 13 shows an example SoW 900e that includes programmable interconnects 970a, 970b, and 970c integrated in the computing component layer, the bridge layer, and the bump layer, respectively. The SoW 900e may include several of the same components as the SoWs 900a-900d. The SoW 900e may further operate in a similar manner as the SoWs 900a-900d. For brevity, repeated discussions of these components and / or features may not be provided below.
[0282] Together, the programmable interconnects 970a, 970b, and 970c may provide a way to comprehensively reconfigure the interconnection pathways throughout the network of interconnections 930 while also providing control over specific interconnection pathways for each computing component 920.2.5. An Example Method for Making a SoW with Programmable Interconnects
[0283] The SoWs disclosed herein (e.g., SoWs 900a-900d) may be manufactured using various packing processes including, but not limited to, a chip-first manufacturing process and a chip-last manufacturing process.
[0284] In a chip-first manufacturing process, the computing components 920 may first be mounted to a wafer and secured to the wafer via, for example, the epoxy material 922. The epoxy material 922 may thereafter be planarized. As described above, the computing components 920 and the epoxy material 922 may form the computing components layer 902a. The subsequent layers of the SoW (e.g., the RDLs 902b and 902d, the bridge layer 902c, and the bump layer 902e) may then be formed directly on the computing component layer 902a. For example, the traces in the network of interconnections 930 and / or the surrounding electrical insulators may be formed, in part, using various semiconductor manufacturing processes (e.g., chemical vapor deposition, physical vapor deposition, photolithography, dry etching, wet etching). Once the network 930 is formed, the solder balls 966 may be mounted onto corresponding electrical contact pads 962. Then, one or more auxiliary components may be connected to one or more of the solder balls 966. Lastly, the wafer supporting the computing components 920 may be removed.
[0285] FIG. 14 shows a non-limiting example of a chip-last manufacturing method 1400 to form the SoWs disclosed herein (e.g., SoWs 900a-900d). As shown, the layers of the SoWAttorney Docket No. MTSH-152WO01forming the network of interconnections 930 (e.g., the RDLs 902b and 902d, the bridge layer 902c, and the bump layer 902e) may be formed on a sacrificial wafer first at step 1402. For example, the traces in the network of interconnections 930 and / or the surrounding electrical insulators may be formed, in part, using various semiconductor manufacturing processes (e.g., chemical vapor deposition, physical vapor deposition, photolithography, dry etching, wet etching). This approach may also provide an opportunity to integrate computing components fabricated separately. For example, if the programmable interconnects 970b and / or the LSI 946 in the bridge layer 902c are fabricated as separate computing components, they may be mounted to the RDL 902d during this step.J0286| At step 1404, the computing components 920 may be mounted to the RDL 902b of the BEOL 910. In some implementations, the computing components 920 may be connected to the traces 932 in the RDL 902b via bump bonding and / or hybrid bonding during this step. At step 1406, the epoxy material 922 may be applied to securely couple the computing components 920 to the BEOL 910 and / or to protect the computing components 920. Once the epoxy material 922 is applied, a planarization process is applied to flatten the top surface of the epoxy material 922. At step 1408, the sacrificial carrier wafer may be removed (e.g., via chemical etching, dry etching), thus exposing, for example, the electrical contact pads 962. This, in turn may allow solder balls 966 to be mounted onto the electrical contact pads 962 and, thereafter, any auxiliary components for the SoW.|0287] It should be appreciated that the method 1400 is a non-limiting example of a chip-last manufacturing process and alterations to the method 1400 are readily contemplated herein. For example, the steps of method 1400 may be performed in any order. For example, the bridge layer 902c, the RDL 902d, and the bump layer 902e may be formed on a first sacrificial wafer. However, the computing components 920 may be mounted to a second sacrificial wafer and the RDL 902b may be formed directly thereon. Thereafter, the second sacrificial wafer (with the computing components 920 and the RDL 902b) may be connected to the bridge layer 902c disposed on the first sacrificial carrier wafer (e.g., via bump bonding or the like). The sacrificial carrier wafers may be removed thereafter.|0288| The sacrificial wafer discussed above may generally provide mechanical support for the other components of the SoWs disclosed herein (e.g., SoWs 900a-900e). In some implementations, the wafer may provide no electrical function. Said another way, the wafer may not transmit any signal within the SoW 900a. The wafer may be formed from aAttorney Docket No. MTSH-152WO01semiconductor including, but not limited to, silicon, gallium arsenide, sapphire, silicon carbine, indium phosphide, gallium nitride, germanium, and / or another semiconducting material.3. An Example SOW Module and Computing Systems Including the Same(0289] Traditional computing architectures are unable to scale with the large amounts of data processed in modem Al models due to limitations on power efficiency and memory bandwidth. For example, computing systems based on the von Neumann architecture typically include a processor (e.g., central processing unit (CPU), a graphics processing unit (GPU)) and memory that communicate via relatively long electrical interconnects integrated in a printed circuit board (PCB). The transmission of large amounts of data in these systems is limited by the power required to transmit data and the latency of transmission. Furthermore, the PCB-integrated format often results in computing systems with a relatively large physical footprint, which limits the compute density attainable for the overall system. For example, the PCB of a server may have an area that is two times larger than the area of the constituent computing components mounted to the PCB.(0290] To address some or all of the foregoing limitations, computing systems may include heterogeneous computing components integrated onto an individual semiconductor wafer. The computing components may include, but are not limited to, a GPU, a central processing unit (CPU), a switch, an input / output (VO) computing component, a network interface controller (NIC), a memory module, and the like. A single wafer with a plurality of computing components integrated thereon is referred to herein as a system-on-a-wafer (SoW).
[0291] The integration of multiple computing components directly onto a wafer provides a way to significantly reduce the physical distance between computing components, especially processing and memory, resulting in higher density, improved power efficiency, and greater memory bandwidth. For example, a SoW may include different combinations of computing components for different applications (e.g., high memory bandwidth, low latency, high compute density, low power consumption, and the like). Moreover, the SoW architecture facilitates the integration of computing components fabricated using different process nodes, thus reducing manufacturing cost.
[0292] Despite the benefits of the SoW architecture, the scalability and flexibility of conventional SoWs are limited in computing systems that use multiple SoWs to process a workload. Conventional SoWs typically include a combination of different computing components that provide different functionalities to take advantage of the benefits ofAttorney Docket No. MTSH-152WO01heterogenous integration as described above. However, for computing systems that use computational resources distributed across multiple SoWs, heterogenous integration can make it more difficult to build and / or adapt these computing systems to perform different workloads. Generally, the compute time for different workloads can vary depending on the computational resources allocated to the workload. For example, some workloads are more memory intensive. For these workloads, the compute time can be reduced by using a computing system that provides and allocates more memory to these workloads. In another example, some workloads are more GPU-intensive. For these workloads, the compute time can be reduced by using a computing system that provides and allocates more GPUs to those workloads.[0293 By combining together different computing components, conventional SoWs make it more challenging to build and / or modify a computing system to provide sufficient computational resources to process different workloads without adding redundant and / or unnecessary computing components. As an illustrative example, a computing system with a single conventional SoW may not have sufficient memory to process a particular workload. The addition of a second conventional SoW may address the shortage of memory, but introduce additional computing components (e.g., more CPUs) that are not needed to process the workload. These additional computing components may go underutilized or, in some instances, may not be used at all, thus wasting space within the computing system and increasing component costs. In some instances, these unused computing components may still consume power, thus increasing power consumption of the computing system and operating costs. Although the SoW can be redesigned to include a different combination of computing components that reduces the quantity of unnecessary computational resources added to a computing system, the process of redesigning a SoW is often a time-consuming and costly endeavor and, thus, impractical to do for every workload.
[0294] In view of the foregoing limitations of conventional SoWs, the present disclosure is directed to various inventive implementations of SoW modules that separate and disaggregate various computing components conventionally integrated into a single conventional SoW. The present disclosure is further directed to computing systems incorporating one or more SoW modules.
[0295] In one aspect, the SoW modules disclosed herein may predominantly include a single type of functional computing component. For example, a SoW module may include a plurality of functional computing components of which one type of functional computing component may constitute at least about 90% of the functional computing components in the SoW module.Attorney Docket No. MTSH-152WO01Herein, a functional computing component is a computing component that provides one or more functionalities to facilitate operation of a computing system. A functional computing component may be distinguished from other functional computing components based, in part, on the functionalities they provide. In some implementations, a functional computing component may be distinguished using common conventions to classify different integrated circuits in a computing system. For instance, a system-on-a-chip (SoC) typically includes a GPU, a CPU, a switch, a memory module (e.g., high-bandwidth memory), and the like. Accordingly, a functional computing component may be categorized according to the foregoing computing components included in the SoC.
[0296] In some implementations, the SoW module may only include functional computing components that are functionally identical to one another. In one example, a SoW module may only include GPUs (i.e., a GPU-only SoW module) as the functional computing component. In another example, a SoW module may only include CPUs (i.e., a CPU-only SoW module). In yet another example, a SoW module may only include switches (i.e., a switch-only SoW module). In addition to functional computing components, the SoW modules disclosed herein may include other components that are not functional computing components as defined herein. For example, the SoW module may include components to facilitate communication between the SoW module and other computing components (e.g., other SoW modules), such as input / output (VO) computing components, NICs, and / or connectors.
[0297] The disaggregated architecture of the SoW modules disclosed herein may provide appreciable improvement in scalability and flexibility compared to conventional SoWs. For instance, the SoW modules may improve the ease of building and / or modifying computing systems to provide sufficient computational resources to process certain workloads while appreciably reducing or, in some instances, eliminating redundant and / or unnecessary computing components in the computing system. For example, a computing system that includes a plurality of SoW modules may not have sufficient memory to process a particular workload. Accordingly, a memory-only SoW module (i.e., a SoW module with memory modules as the functional computing components) may be added to address the shortage of memory. In this example, other computing components that are not needed to process the workload, such as additional CPUs or GPUs, may optionally be omitted from the computing system, resulting in fewer unused computing components and reducing the amount of space allocated to these unused computing components.
[0298] Moreover, the SoW modules may be readily removed and replaced with the same SoWAttorney Docket No. MTSH-152WO01module or a different SoW module. This modular approach provides a way to reconfigure a computing system to accommodate different workloads, e.g., by increasing and / or decreasing different computational resources according to the demands of the workload. For example, if a workload is more memory intensive, memory-only SoW modules may be readily added to the computing system. In another example, if a workload is more GPU-intensive, GPU-only SoW modules may be readily added to the computing system.
[0299] The SoW modules of a computing system may be communicatively coupled to each other to facilitate operation of the computing system. In some implementations, the SoW modules may be mounted onto a printed circuit board (PCB), and the PCB may provide electrical interconnections to transmit electrical signals between the SoW modules. In some implementations, communication between SoW modules may be accomplished through a cable connection. For example, each SoW module may include a connector communicatively coupled to one or more VO computing components, which, in turn, are communicatively coupled to a plurality of functional computing components. The connector may support a connection to a cable configured to carry an electrical signal to and / or from the SoW module. In some implementations, the combination of the connector and the I / O computing components may provide a common communication interface amongst different SoW modules, e.g., SoW modules with different types of functional computing components.[0300| In some implementations, the communication interface may support various communication protocols including, but not limited to, a die-to-die protocol (e.g., Universal Chiplet Interconnect Express (UCIe)), peripheral component interconnect express (PCI Express), Ultra Accelerator Link (UALink), Ethernet, InfiniBand, Ultra Ethernet, Optical I / O, and the like. For example, PCI Express may be used to facilitate communication between a GPU-only SoW module and a CPU-only SoW module. In another example, a UALink may be used to facilitate communication between a pair of GPU-only SoW modules for GPU-to-GPU scaling.
[0301] In one example implementation, a computing system comprises one or more system-on-a- wafer (SoW) modules where each SoW module of the one or more SoW modules includes a back end of line (BEOL); and a plurality of computing components electrically coupled to the BEOL. The plurality of computing components consists of at least one of an input / output (I / O) computing component or a network interface controller (NIC); and a plurality of functional computing components where each functional computing component of the plurality of functional computing components is functionally identical to every other functionalAttorney Docket No. MTSH-152WO01computing component of the plurality of functional computing components.
[0302] The one or more SoW modules may include a SoW module where the SoW module includes a plurality of central processing units (CPUs) as the plurality of functional computing components of the SoW module. The one or more SoW modules may include a SoW module where the SoW module includes a plurality of graphics processing units (GPUs) as the plurality of functional computing components of the SoW module. The one or more SoW modules may include a SoW module where the SoW module includes a plurality of switches as the plurality of functional computing components of the SoW module. The one or more SoW modules may include a SoW module where the SoW module includes a plurality of memory modules as the plurality of functional computing components of the SoW module. The one or more SoW modules may include a first SoW module including a plurality of central processing units (CPUs) as the plurality of functional computing components of the first SoW module; a second SoW module including a plurality of graphics processing units (GPUs) as the plurality of functional computing components of the second SoW module; a third SoW module including a plurality of switches as the plurality of functional computing components of the third SoW module; and a fourth SoW module including a plurality of memory modules as the plurality of functional computing components of the fourth SoW module.
[0303] Each SoW module of the one or more SoW modules may be communicatively coupled to every other SoW module of the one or more SoW modules. Each SoW module of the one or more SoW modules may be communicatively coupled to up to four other SoW modules of the one or more SoW modules. The computing system may further include a plurality of cables, wherein each cable of the plurality of cables is communicatively coupled to a pair of SoW modules of the one or more SoW modules and is configured to transmit one or more signals between the pair of SoW modules. Each SoW module of the one or more SoW modules may further include a connector having a first end and a second end, wherein the first end is communicatively coupled to the at least one of the I / O computing component or the NIC and the second end is communicatively coupled to one cable of the plurality of cables.
[0304] The computing system may further include a printed circuit board (PCB) to communicatively couple each SoW module of the one or more SoW modules to every other SoW module of the one or more SoW modules. The one or more SoW modules may include a first SoW module; a second SoW module communicatively coupled to the first SoW module via a first connection standard; and a third SoW module communicatively coupled to the first SoW module via a second connection standard different from the first connection standard.Attorney Docket No. MTSH-152WO01The at least one of the I / O computing component or the NIC may support at least one of a Universal Chiplet Interconnect Express (UCIe) connection, a Peripheral Component Express (PCIe) connection, an Ultra Accelerator Link (UALink) connection, an Ethernet connection, an Ultra Ethernet connection, an InfiniBand connection, or an optical input / output (EO) connection.(0305] In another example implementation, a computing system includes one or more system-on-a- wafer (SoW) modules where each SoW module of the one or more SoW modules includes a back end of line (BEOL); and a plurality of computing components electrically coupled to the BEOL. The plurality of computing components includes at least one of an input / output (EO) computing component or a network interface controller (NIC); and a plurality of functional computing components predominantly including a first type of functional computing component.(0306] A percentage of the plurality of functional computing components corresponding to the first type of functional computing component may be greater than or equal to 90%. A quantity of the plurality of functional computing corresponding to one or more other type of functional computing components different from the first type of functional computing components may be less than or equal to 150.(0307] Conventional SoWs typically include a combination of different computing components (see, e.g., the computing components 12 in FIG. IB) to take advantage of the benefits of heterogenous integration with respect to a single wafer. However, conventional SoWs, such as the SoW 10a, are limited in terms of their scalability and flexibility, particularly when used in computing systems that utilize computational resources from multiple SoWs to process a workload. By combining together different computing components into a single package, conventional SoWs do not readily provide a way to selectively increase or decrease specific computational resources associated with certain functional computing components (e.g., CPUs, GPUs, memory modules, switches, and the like). Rather, the addition (or removal) of conventional SoWs to a computing system affects computational resources associated with multiple functional computing components, thus making it difficult to configure a computing system to efficiently provide computational resources to process various workloads.(0308] As an illustrative example, a computing system supporting multiple SoWs may have a surplus of switches 12c and a shortage of CPUs 12d. Although the shortage of CPUs 12d can be rectified by the addition of more SoWs 10a to the computing system, this approach can haveAttorney Docket No. MTSH-152WO01several drawbacks. First, a computing system only supports a limited quantity of SoWs. Thus, it may not be possible to add more SoWs to address shortages in certain computational resources. Second, the addition of more SoWs 10a increases the capital costs of the computing system. The capital costs may further include the cost of surplus computing components that may be underutilized or, in some instances, not used at all. The introduction of surplus computing components may also adversely affect operation of the computing system by increasing power consumption and, hence, operating costs of the computing system.(0309] Following below are more detailed descriptions of various concepts related to, and implementations of, a SoW module that predominantly includes functional computing components that are functionally identical to facilitate a disaggregated architecture that provides greater flexibility and scalability compared to conventional SoWs. The present disclosure is also directed to computing systems that include one or more of the SoW modules disclosed herein. It should be appreciated that various concepts introduced above and discussed in greater detail below may be implemented in multiple ways. Examples of specific implementations and applications are provided primarily for illustrative purposes so as to enable those skilled in the art to practice the implementations and alternatives apparent to those skilled in the art.
[0310] The figures and example implementations described below are not meant to limit the scope of the present implementations to a single embodiment. Other implementations are possible by way of interchange of some or all of the described or illustrated elements. Moreover, where certain elements of the disclosed example implementations may be partially or fully implemented using known components, in some instances only those portions of such known components that are necessary for an understanding of the present implementations are described, and detailed descriptions of other portions of such known components are omitted so as not to obscure the present implementations.
[0311] In the discussion below, various examples of a SoW module are provided, wherein a given example or set of examples showcases a functional computing component, an VO computing component, a connector, a cable, or a PCB supporting the SoW modules. It should be appreciated that one or more features discussed in connection with a given example of a SoW module may be employed in other examples of SoW modules, respectively, according to the present disclosure, such that the various features disclosed herein may be readily combined in a given SoW module according to the present disclosure (provided that respective features are not mutually inconsistent).Attorney Docket No. MTSH-152WO01
[0312] Certain dimensions and features of the SoW modules are described herein using the terms “approximately,” “about,” “substantially,” and / or “similar.” As used herein, the terms “approximately,” “about,” “substantially,” and / or “similar” indicates that each of the described dimensions or features is not a strict boundary or parameter and does not exclude functionally similar variations therefrom. Unless context or the description indicates otherwise, the use of the terms “approximately,” “about,” “substantially,” and / or “similar” in connection with a numerical parameter indicates that the numerical parameter includes variations that, using mathematical and industrial principles accepted in the art (e.g., rounding, measurement or other systematic errors, manufacturing tolerances, etc.), would not vary the least significant digit.3.1 Example SoW Modules
[0313] To address some or all of the limitations of conventional SoWs, a system-on-a-wafer (SoW) module is disclosed herein to facilitate the separation and disaggregation of multiple functionalities typically included in a single conventional SoW. For example, the SoW modules disclosed herein may predominantly include functional computing components that are functionally identical to one another. For instance, the SoW modules disclosed herein may include a plurality of functional computing components of which a first type of functional computing component (e.g., a CPU, a GPU, a switch, a memory module, and so on) constitutes a substantial proportion of the plurality of functional computing components.
[0314] In some implementations, a substantial proportion may be expressed as the percentage of functional computing components that correspond to the first type of functional computing component, i.e., the ratio of the quantity of functional computing components of the first type divided by the total quantity of functional computing components in the SoW module where the ratio is expressed as a percentage. The percentage of functional computing components that correspond to the first type of functional computing component may be greater than or equal to about 90%, greater than or equal to about 95%, or greater than or equal to about 99%. The term “about,” when used to describe the percentage of functional computing components that correspond to the first type of functional computing component is intended to cover rounding of the percentage to the nearest integer when the percentage is expressed as a percentage normalized to a maximum of 100%. For example, “about 90%” may correspond to the following ranges: 89.55% to 90.45%, 89.1% to 90.9%, 88.65% to 91.35%, 88.2% to 91.8%, or 88% to 92% including all values and sub-ranges in between.(0315] In some implementations, a substantial proportion may be expressed by the quantity ofAttorney Docket No. MTSH-152WO01functional computing components that correspond to one or more other type of functional computing components, i.e., functional computing components that are different from the first type of functional computing component. The quantity of functional computing components that correspond to one or more other types of functional computing components may be less than or equal to 150, less than or equal to 100, less than or equal to 50, less than or equal to 25, less than or equal to 10, less than or equal to 5, or less than or equal to 1.[03.16] In some implementations, the SoW modules disclosed herein may only include functional computing components that are functionally identical to one another. A computing system may include multiple SoW modules that each provide a particular type of functional computing component. For example, a computing system may include one or more CPU-only SoW modules (i.e., a SoW module that only includes CPUs as the functional computing component), one or more GPU-only SoW modules (i.e., a SoW module that only includes GPUs as the functional computing component), one or more memory-only SoW modules (i.e., a SoW module that only includes memory modules as the functional computing component), one or more switch-only SoW modules (i.e., a SoW module that only includes switches as the functional computing component), and so on.
[0317] Limiting each SoW module to predominantly or, in some instances, only a single set of functionalities associated with a particular functional computing component (e.g., a CPU, a GPU, and so on) provides a way to build and customize a computing system to provide sufficient computational resources for different workloads while reducing or, in some instances, eliminating redundant and / or unnecessary computing components in the computing system. Additionally, the SoW modules disclosed herein may support various communication protocols, thus allowing one SoW module (e.g., a GPU-only SoW module) in a computing system to be readily replaced with another SoW module (e.g., a memory-only SoW module) in a modular manner. Compared to conventional SoWs, the SoW modules disclosed herein provide greater flexibility to combine and scale specific computational resources. Referring to the example scenario above where a computing system has a surplus of switches and a shortage of CPUs, the SoW modules may provide a way to efficiently address this imbalance in computational resources. If the computing system is built using SoW modules including, for example, a switch-only SoW module, the switch-only SoW module in the computing system may be replaced with a CPU-only SoW module. This, in turn, may reduce the surplus of switches while addressing the shortage of CPUs.
[0318] In some implementations, the functional computing components in a given SoWAttorney Docket No. MTSH-152WO01module may be identical in terms of their design. For example, the functional computing components of a SoW module may be fabricated from the same IC layout design. It should be appreciated, however, that the functional computing components in a given SoW module, despite sharing the same design, may not be physically identical due to variabilities in manufacture. For instance, the dimensions and / or placement of various features in the functional computing components are generally subject to variation due to manufacturing tolerances associated with various processes used in the fabrication of ICs.(0319] In some implementations, the functional computing components in a SoW module may not be identical in design but may nevertheless be considered functionally identical. For instance, the functional computing components in a given SoW module may provide the same or similar functionalities but may have different performance specifications. For example, a CPU-only SoW module may include a combination of (i) one or more first CPUs that provide more compute power at the expense of greater power consumption and (ii) one or more second CPUs that consume less power at the expense of providing less compute power. The inclusion of functional computing components with different performance specifications may provide greater flexibility when allocating computational resources to different workloads. Referring to the example CPU-only SoW module described above, some workloads may require less CPU compute power and, hence, may be allocated the second CPUs that consume less power. Likewise, some workloads may require more CPU compute power and, hence, may be allocated the first CPUs that provide more compute power.(0320] The computing systems disclosed herein may include one or more SoW modules. In some implementations, the only wafer-based computing devices in the computing system may be SoW modules. Said another way, the computing system may not include a SoW that incorporates multiple, different types of functional computing components (e.g., the SoW 10a). In some implementations, the SoW modules may operate in combination with SoWs incorporating multiple, different types of functional computing components (e.g., the SoW 10a). For example, the computing system may include the SoW 10a and one or more GPU-only SoW modules and / or memory-only SoW modules to selectively increase the GPUs and memory, respectively, available in the computing system.(0321] FIGS. 15A-15D show several non-limiting examples of SoW modules. FIG. 15A shows an example SoW module 1500a that includes only GPUs 1520b as the functional computing components. Accordingly, the SoW module 1500a is also referred to herein as a GPU-only SoW module 1500a. FIG. 15B shows an example SoW module 1500b that includesAttorney Docket No. MTSH-152WO01only memory modules 1520c as the functional computing components. Accordingly, the SoW module 1500b is al so referred to herein as a memory-only SoW module 1500b. FIG. 15C shows an example SoW module 1500c that includes only switches 1520d as the functional computing components. Accordingly, the SoW module 1500c is also referred to herein as a switch-only SoW module 1500c. FIG. 15D shows an example SoW module 1500d that includes only CPUs 1520e as the functional computing components. Accordingly, the SoW module 1500d is also referred to herein as a CPU-only SoW module 1500d. The SoW modules 1500a-1500d may include several of the same components as the SoW 10a. For brevity, repeated discussions of these components and / or features may not be provided below.[03221 As described above, functional computing components may be distinguished by the functionalities they provide to a computing system. Different functionalities typically require different functional circuit blocks in the computing component. Herein, a functional circuit block refers to electronic circuity designed to perform one or more specific functions associated with the computing component. Thus, functional computing components may also be distinguished by the functional circuit blocks contained therein. Following below are several non-limiting examples of functional circuit blocks often used in different functional computing components.
[0323] For the GPU 1520b, the functional circuit block may include electronic circuity that facilitates the execution of one or more functions typically associated with GPUs. The electronic circuitry in this example may include, but is not limited to, a multiprocessor, a register, shared memory, and the like.
[0324] For the memory module 1520c, the functional circuit block may include electronic circuity that facilitates the execution of one or more functions typically associated with memory modules. The electronic circuitry in this example may include, but is not limited to, a memory cell, and the like. Examples of memory modules 1520c also include, but are not limited to, a dynamic random access memory (DRAM), a static random access memory (SRAM) module, a flash memory module, a solid-state drive (SSD), a non-volatile random access memory (NVRAM) module, a read-only memory (ROM) module (e.g., a floating-gate ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), one-time programmable ROM (OTPROM), and the like), and any other suitable form of memory.
[0325] For the switch 1520d, the functional circuit block may include electronic circuity thatAttorney Docket No. MTSH-152WO01facilitates the execution of one or more functions typically associated with switches. The electronic circuitry in this example may include, but is not limited to, a switching element (e.g., a metal-oxide-semiconductor field-effect transistor (MOSFET)), control logic, a data multiplexer and / or demultiplexer, a clock and synchronization circuit, and the like.[0326| For the CPU 1520e, the functional circuit blocks may include electronic circuitry that facilitates the execution of one or more functions typically associated with CPUs. The electronic circuitry in this example may include, but is not limited to, an arithmetic logic unit (ALU), a control unit (CU), a register, and the like.
[0327] As shown in FIGS. 15A-15D, the SoW modules 1500a-1500d may share several common features and / or components. Each SoW module 1500a-1500d may include a BEOL 1510 to provide mechanical support for the functional computing components 1520. The BEOL 1510 may further provide a network of interconnection pathways to receive and / or transmit data signals (e.g., signals containing information), to receive and / or transmit power, and / or to electrically ground the SoW module. To simplify discussion of the interconnection pathways herein, the interconnection pathways are described as being able to carry a “signal.” The term “signal” is intended to cover any electrical signal transmitted by the interconnection pathways including, but not limited to, a data signal, a power signal, a ground signal, and the like. The network of interconnection pathways may thus facilitate the transmission of signals between the computing components 1520 of the SoW module and the transmission of signals to or from external computing components (e.g., other SoW modules).
[0328] In some implementations, the BEOL 1510 may be formed in the shape of a wafer. For example, the BEOL 1510 may be fabricated on a sacrificial wafer and retain the same geometry as the wafer when the wafer is removed. The shape and size of the BEOL 1510 may determine the shape and size of the SoW modules 1500a-1500d. Generally, the BEOL 1510 may have various geometries including, but not limited to, a circle, an oval, a triangle, a square, a rectangle, a polygon, and any combinations of the foregoing. In some implementations, the characteristic width, £>, of the BEOL 1510 may be equal to about 300 millimeters (mm). More generally, the characteristic width, £>, of the BEOL 1510 may generally range from about 25 mm to about 450 mm, including any sub-ranges and values in between. In some implementations, the BEOL 1510 may have a thickness, / , equal to about 1 mm. More generally, the thickness, / , of the BEOL 1510 may generally range from about 0.3 mm to about 2 mm, including any sub-ranges and values in between.Attorney Docket No. MTSH-152WO01(0329] It should be appreciated that the computing components 1520 may be disposed onto only a portion of the BEOL 1510, particularly when the BEOL 1510 is shaped as a wafer. Said another way, the computing components 1520 may not be distributed uniformly across the BEOL 1510. For example, FIGS. 15A-15D show the computing components 1520 are disposed onto a central portion of the BEOL 1510. In some implementations, no computing components 1520 may be disposed onto the outer portions of the BEOL 1510. In some implementations, the network of interconnections of the BEOL 1510 may be localized to where the computing components 1520 are located. Thus, the outer edges of the BEOL 1510 may primarily comprise an electrical insulator with few or, in some instances, no interconnections. In some implementations, the network of interconnections may extend across the width of the BEOL 1510 even when the computing components 1520 are localized to a portion of the BEOL 1510. Said another way, the network of interconnections may extend to the outer edges of the BEOL 1510. This may facilitate connection with, for example, physical connectors (see, e.g., the connectors 1590).
[0330] FIGS. 15A-15D further show the SoW modules 1500a-1500d may include other components in addition to the functional computing components described above. For example, each of the SoW modules 1500a-1500d may include one or more EO computing components 1520a and one or more connectors 1590. The I / O computing components 1520a and the connectors 1590 are used to transmit signals between the functional computing components of the SoW module and external computing components (e.g., other SoW modules).(0331] The EO computing components 1520a may provide various functionalities including, but not limited to, the transmission and routing of signals to and from each functional computing component in the SoW modules 1500a-1500d, compliance with one or more communication protocols, signal buffering, interrupt handling, and the like. In some implementations, the SoW modules disclosed herein (e.g., SoW modules 1500a-1500d) may include a network interface controller (NIC) (not shown) to provide relatively higher speed communication between SoW modules compared to the EO computing component. In some implementations, the NIC may be included in combination with the I / O computing components 1520a. In some implementations, the EO computing components 1520a may be replaced by the NICs. The connectors 1590 may provide a physical interface to transmit signals from the SoW module and / or receive signals for the SoW module. In some implementations, the connector 1590 may be a bus supporting a connection to a cable (see, e.g., the cable 1610 inAttorney Docket No. MTSH-152WO01FIG. 16).
[0332] The VO computing component 1520a and the connectors 1590 may support one or more communication protocols including, but not limited to, Universal Chiplet Interconnect Express (UCIe), Peripheral Component Express (PCIe), Ultra Accelerator Link (UALink), an Ethernet connection, an Ultra Ethernet connection, an InfiniBand connection, and an optical I / O connection.
[0333] In some implementations, the I / O computing components 1520a and the connectors 1590 may support multiple communication protocols, in part, to facilitate interoperability between different SoW modules (e.g., the SoW modules 1500a-1500d). The support of multiple communication protocols (also referred to herein as a Unified I / O chiplet interface) may allow SoW modules to be installed and / or replaced in a modular manner. For instance, a SoW module supporting a first type of function computing component (e.g., a CPU, see, e.g., CPU-only SoW module 1500d) may be readily replaced with another SoW module supporting a second type of function computing component (e.g., a GPU, see, e.g., GPU-only SoW module 1500a) without requiring any other modifications to the hardware of the computing system. This modular approach, in turn, may readily allow different combinations of SoW modules to be installed into a computing system. Additionally, different SoW modules requiring different communication protocols may utilize the same I / O computing components 1520a and connectors 1590, thus reducing the time and cost of designing different SoW modules.
[0334] As an illustrative example, a computing system may include a CPU-only SoW module 1500d, a first GPU-only SoW module 1500a, and a second GPU-only SoW module 1500a. The first GPU-only SoW module 1500a may be communicatively coupled to the CPU-only SoW module 1500d using a first connection standard, e.g., a PCIe communication protocol. The first GPU-only SoW module 1500a may be communicatively coupled to the second GPU-only SoW module 1500a using a second connection standard, e.g., a UALink communication protocol. In this example, the I / O computing components 1520a and connectors 1590 in each SoW module may be identical. Thus, the SoW modules may be connected together in a plug-and-play manner.
[0335] As shown, the I / O computing component 1520a may be disposed on the BEOL 1510 between the connector 1590 and the functional computing components. In some implementations, the computing components 1520 may be arranged such that the functional computing components (e.g., the GPUs 1520b, the memory modules 1520c, the switchesAttorney Docket No. MTSH-152WO011520d, or the CPUs 1520e) are grouped together and disposed onto a central portion of the BEOL 1510. The I / O computing components 1520a, in turn, may be disposed along the periphery of the group of functional computing components, thus surrounding the functional computing components. The connectors 1590 may be disposed near the outer edges of the BEOL 1510 in close proximity to one or more I / O computing components 1520a. The foregoing arrangement of the computing components 1520 and the connectors 1590 may provide higher compute densities, in part, by reducing the distance between neighboring computing components, thus increasing the packing density of similarly shaped computing components.
[0336] It should be appreciated that some functional computing components may not be disposed directly adjacent to an I / O computing component 1520a, but these functional computing components may nevertheless be electrically connected directly to the I / O computing component 1520a via the network of interconnection pathways within the BEOL 1510. It should also be appreciated that the foregoing arrangement of the computing components 1520 and the connectors 1590 is a non-limiting example. More generally, the computing components 1520 may be distributed across the BEOL 1510 in different ways. In another example, the computing components 1520 may be arranged such that each functional computing component is disposed directly adjacent to an I / O computing component 1520a.[0337| It should be appreciated that other components not shown in FIGS. 15A-15D may be integrated into the SoW modules disclosed herein (e.g., the SoW modules 1500a-1500d). For example, the SoW modules disclosed herein may include a power module to regulate electrical power supplied to the SoW modules (see, e.g., the power module 19 in FIG. IB). In another example, the SoW modules disclosed herein may include a thermal module to dissipate heat generated by the SoW module during operation (see, e.g., the thermal module 17 in FIG. IB). Examples of thermal modules, which may be used in the present technology, are described in PCT Application No. PCT / US2023 / 067058 filed on May 16, 2023 and entitled, “Electronic Package Construction for Immersion Cooling of Integrated Circuits,” and PCT Application No. PCT / US2024 / 055034 filed on November 8, 2024 and entitled, “Systems for Thermal and Power Management in Three-Dimensional Integrated Circuits.” Each of the foregoing applications are incorporated herein by reference in their entirety for all purposes.
[0338] Additional examples and features of SoWs, which may be used in the present technology, are described in PCT Application No. PCT / US2024 / 059364 filed on December 10, 2024, and entitled “WAFER-LEVEL HETEROGENEOUS INTEGRATION FORAttorney Docket No. MTSH-152WO01ARTIFICIAL INTELLIGENCE ACCELERATORS AND SYSTEMS,” which is incorporated herein by reference in its entirety.(0339] As described above, a computing system may incorporate one or more SoW modules (e.g., the SoW modules 1500a-1500d). In some implementations, the combination of SoW modules may be selected so that the computing system functions as a server (e.g., a server in a datacenter) or at least a portion of a server. Examples of servers incorporating the SoW modules disclosed herein include, but are not limited to, an artificial intelligence (Al) accelerator (e.g., for an Al training server, or an Al inference server), a web server, a database server, an email server, a web proxy server, a domain name server (DNS), an application server, and the like.[0340| The disaggregated architecture provided by the SoW modules 1500a-1500d may be especially beneficial when processing Al-related workloads (e.g., Al training, Al inference). The hardware demands for Al-related workloads can vary significantly. For example, some Al-related workloads may be more GPU-intensive; hence, the compute time for these workloads can be reduced by increasing the quantity of GPUs available in the computing system. In another example, some Al-related workloads may be more memory-intensive; hence, the compute time for these workloads can be reduced by increasing the quantity of memory modules available in the computing system. The SoW modules 1500a-1500d disclosed herein may address these different hardware demands by providing a way to selectively add or remove GPUs and / or memory modules to the computing system, e.g., by installing and / or uninstalling GPU-only SoW modules and / or memory-only SoW modules. In this manner, the SoW modules 1500a-1500d disclosed herein may appreciably enhance the scalability and adaptability of computing systems used to process Al-related workloads.3.2. Example Computing Systems with SoW Modules
[0341] The computing systems disclosed herein may include one or more SoW modules (e.g., SoW modules 1500a-1500d). For computing systems that include multiple SoW modules, each SoW module may be electrically connected to one or more of the other SoW modules. In some implementations, each SoW module may be electrically connected to every other SoW module in the computing system. This may be accomplished in several ways.
[0342] In one non-limiting example, the SoW modules in the computing system may be electrically connected together via one or more cables. For instance, a cable may be connected to the connector 1590. Moreover, each SoW module may include multiple connectors 1590 to facilitate connection with multiple SoW modules. FIG. 16 shows an example computingAttorney Docket No. MTSH-152WO01system 1600a that includes the SoW modules 1500a, 1500b, 1500c, and 1500d. As shown, each of the SoW modules 1500a-1500d is connected to two other SoW modules via respective cables 1610. Each cable 1610 is communicatively coupled to a pair of SoW modules. Specifically, the SoW module 1500a is connected to the SoW modules 1500b and 1500c, the SoW module 1500b is connected to the SoW modules 1500a and 1500d, and so on. In this example, each of the SoW modules 1500a-1500d includes four connectors 1590. Thus, each of the SoW modules 1500a-1500d may be connected to two additional SoW modules. For instance, the SoW module 1500a may be connected to the SoW module 1500d or another SoW module (not shown) in the computing system 1600a.[0343 The quantity of SoW modules that can be connected to each SoW module may be limited by the quantity of connectors available in the SoW module. Generally, the SoW modules 1500a-1500d disclosed herein may include less than four connectors or greater than four connectors. The quantity of connectors in the SoW modules 1500a-1500d may be limited by the size of the connectors 1590 and the size and / or space available on the BEOL 1510 to support the connectors 1590.
[0344] In another non-limiting example, the SoW modules in the computing system (e.g., SoW modules 1500a-1500d) may be electrically connected together via a printed circuit board (PCB). For instance, each SoW module may be electrically connected to the PCB via a ball grid array (BGA). The BGA may be electrically connected, for example, to a plurality of electrical contacts formed on the BEOL of each SoW module. The PCB, in turn, may include a plurality of traces that provide interconnection pathways to transmit signals between the SoW modules mounted to the PCB. FIG. 17 shows an example computing system 1600b that includes SoW modules 1500a’, 1500b’, 1500c’, and 1500d’ mounted to a PCB 1620. The SoW modules 1500a’, 1500b’, 1500c’, and 1500d’ do not include connectors 1590, but are otherwise identical to the SoW modules 1500a, 1500b, 1500c, and 1500d. As shown, the PCB 1620 may include a plurality of traces 1622 that provide interconnection pathways to electrically connect the SoW modules 1500a’, 1500b’, 1500c’, and 1500d’ to one another.
[0345] FIG. 17 further shows each of the SoW modules 1500a’-1500d’ is connected to two other SoW modules. It should be appreciated that these connections are a non-limiting example and that each SoW module may be connected to only one other SoW module or more than two SoW modules. Furthermore, in this example, the quantity of SoW modules that can be connected to each SoW module may only be limited by the density and arrangement of the traces 1622 in the PCB 1620.Attorney Docket No. MTSH-152WO01
[0346] The computing systems disclosed herein may be in various forms including, but not limited to a tower server, a rack server, a blade server, and the like. For example, FIG. 18 shows a computing system 1600c embodied as a rack server. As shown, the computing system 1600c may include a frame 1630 supporting multiple racks. The SoW modules 1500a, 1500b, and 1500c are installed in the frame 1630 of the computing system 1600c as different racks.3.3. An Example Method for Making a Computing System with SoW Modules
[0347] The SoWs disclosed herein (e.g., SoWs 1500a-1500d and / or 1500a’-1500d’) may be manufactured using various packaging processes including, but not limited to, a chip-first manufacturing process and a chip-last manufacturing process, which are described and illustrated in greater detail above in Sections 1 and 2.
[0348] As described above in Sections 1 and 2, in a chip-first manufacturing process, the computing components 1520 may first be mounted to a wafer and secured to the wafer via, for example, an epoxy material (e.g., epoxy material 122 / 922). The epoxy material may thereafter be planarized. The computing components 1520 and the epoxy material may form a computing component layer. The subsequent layers of the SoW (e.g., an RDL(s), a bridge layer, and / or a bump layer) may then be formed directly on the computing component layer. For example, the traces in the network of interconnections and / or the surrounding electrical insulators may be formed, in part, using various semiconductor manufacturing processes (e.g., chemical vapor deposition, physical vapor deposition, photolithography, dry etching, wet etching). Once the network of interconnections is formed, solder balls may be mounted onto corresponding electrical contact pads. Then, one or more auxiliary components may be connected to one or more of the solder balls. Lastly, the wafer supporting the computing components 1520 may be removed.
[0349] Additionally, as described above in Sections 1 and 2, in a chip-last manufacturing process, the layers of the SoW forming the network of interconnections (e.g., an RDL(s), a bridge layer, and / or a bump layer) may be formed on a sacrificial wafer first. For example, the traces in the network of interconnections and / or the surrounding electrical insulators may be formed, in part, using various semiconductor manufacturing processes (e.g., chemical vapor deposition, physical vapor deposition, photolithography, dry etching, wet etching). The computing components 1520 may be mounted to an RDL of the BEOL. In some implementations, the computing components 1520 may be connected to traces in the RDL via bump bonding and / or hybrid bonding during this step. Then, an epoxy material may be appliedAttorney Docket No. MTSH-152WO01to securely couple the computing components 1520 to the BEOL and / or to protect the computing components 1520. Once the epoxy material is applied, a planarization process is applied to flatten the top surface of the epoxy material. Finally, the sacrificial carrier wafer may be removed (e.g., via chemical etching, dry etching), thus exposing, for example, electrical contact pads. The exposure of the contact pads, in turn, may allow solder balls to be mounted onto the electrical contact pads and, thereafter, any auxiliary components for the SoW.
[0350] It should be appreciated that this is a non-limiting example of a chip-last manufacturing process and alterations to this process are readily contemplated herein. For example, a bridge layer, an RDL, and a bump layer may be formed on a first sacrificial wafer. However, the computing components 1520 may be mounted to a second sacrificial wafer and a RDL may be formed directly thereon. Thereafter, the second sacrificial wafer (with the computing components 1520 and an RDL) may be connected to the bridge layer disposed on the first sacrificial carrier wafer (e.g., via bump bonding or the like). The sacrificial carrier wafers may be removed thereafter.
[0351] The sacrificial wafer discussed above may generally provide mechanical support for the other components of the SoWs disclosed herein (e.g., SoWs 1500a-1500d and / or 1500a’-1500d’). In some implementations, the wafer may provide no electrical function. Said another way, the wafer may not transmit any signal within the SoWs disclosed herein (e.g., SoWs 1500a-1500d and / or 1500a’-1500d’). The wafer may be formed from a semiconductor including, but not limited to, silicon, gallium arsenide, sapphire, silicon carbine, indium phosphide, gallium nitride, germanium, and / or another semiconducting material.
[0352] FIG. 19 shows a non-limiting example of a manufacturing method 1900 to form a computing system (e.g., 1600a-1600c) including the SoWs disclosed herein (e.g., SoWs 1500a-1500d and / or 1500a’-1500d’). The method 1900 begins at steps 1902a-1902d, which may be performed sequentially or at the same time. At step 1902a, a first SoW, including a plurality of first computing components mounted to the first SoW, is manufactured using any of the chip manufacturing processes disclosed herein. At step 1902b, a second SoW, including a plurality of second computing components mounted to the second SoW, is manufactured using any of the chip manufacturing processes disclosed herein. At step 1902c, a third SoW, including a plurality of third computing components mounted to the third SoW, is manufactured using any of the chip manufacturing processes disclosed herein. At step 1902d, a fourth SoW, including a plurality of fourth computing components mounted to the fourth SoW, is manufactured using any of the chip manufacturing processes disclosed herein.Attorney Docket No. MTSH-152WO01
[0353] Then, at step 1904, the first, second, third, and / or fourth SoWs may be electrically connected to form a computing system. As described above, each SoW module may be electrically connected to one or more of the other SoW modules. In some implementations, each SoW module may be electrically connected to every other SoW module in the computing system. The electrical connection may be accomplished in several ways as outlined above, including, but not limited to, one or more cables and / or via a PCB.
[0354] It should be appreciated that this is a non-limiting example of a manufacturing method to form a computing system and alterations to this process are readily contemplated herein. For example, the steps of method 1900 may be performed in any order. For example, while method 1900 illustrates four SoWs, the method 1900 can be performed with any number of SoWs depending on the needs of the computing system. As outlined above, the first SoW module may be a CPU-only SoW module, the second SoW module may be a GPU-only SoW module, the third SoW module may be a memory-only SoW module, and the fourth SoW module may be a switch-only SoW module. However, the computing system may be formed with any combination of functional computing components disclosed herein in any order.4. An Example Memory Architecture in SOW Al Accelerator
[0355] Some implementations of the present disclosure relate to modular and configurable memory architectures and hierarchies designed for SoW technology. Some implementations of the present disclosure include the integration of different memory modules types — such as on-chip SRAM, stacked memory, and side memory — into a cohesive and flexible memory hierarchy. By carefully designing the memory configuration and SoW layout, some implementations of the present technology allow for the modulation of memory types and their arrangement to meet various computing requirements, such as Al model training and inference, high-performance tasks demanding rapid data access, and large-scale models needing substantial storage capacity.
[0356] Following below are more detailed descriptions of various concepts related to, and implementations of, modular and configurable memory architectures and hierarchies for SoW and servers including SoWs. In particular, the present technology includes the integration of different memory module types — such as on-chip SRAM, stacked memory, and side memory — into a cohesive and flexible memory hierarchy. By carefully designing the memory configuration and SoW layout, the present technology allows for the modulation of memory types and their arrangement to meet various computing requirements, such as Al model trainingAttorney Docket No. MTSH-152WO01and inference, high-performance tasks demanding rapid data access, and large-scale models needing substantial storage capacity. It should be appreciated that various concepts introduced above and discussed in greater detail below may be implemented in multiple ways. Examples of specific implementations and applications are provided primarily for illustrative purposes so as to enable those skilled in the art to practice the implementations and alternatives apparent to those skilled in the art.
[0357] The figures and example implementations described below are not meant to limit the scope of the present implementations to a single embodiment. Other implementations are possible by way of interchange of some or all of the described or illustrated elements. Moreover, where certain elements of the disclosed example implementations may be partially or fully implemented using known components, in some instances only those portions of such known components that are necessary for an understanding of the present implementations are described, and detailed descriptions of other portions of such known components are omitted so as not to obscure the present implementations.4.1. An Example Memory Architecture in SOW Al Accelerator
[0358] FIG. 20A illustrates a top view of an SoW 2000 in accordance with the present technology. SoW 2000 may include a plurality of systems-on-a-chip (SOCs 2020) grouped in unit blocks 2002, discussed in greater detail with respect to FIG. 20B. SoW 2000 may include a substrate 2010, which may be a silicon wafer in an aspect. Substrate 2010 may include one or more redistribution layers (RDLs), local silicon interconnects (LSIs), or other arrangements of traces within substrate 2010. Substrate 2010 may be circular and have a diameter of about 300 mm. In an aspect, substrate 2010 may be a panel wafer having a square profile of about 300 mm in length by about 300 mm in width. Further examples of wafer substrates and SoWs having both circular and square / panel shapes are disclosed in PCT Application No. PCT / US2025 / 018252, titled “SYSTEMS AND METHODS FOR SYSTEMS-ON-A-WAFER,” filed March 4, 2025, and incorporated herein by reference in its entirety for all purposes. SoWs described herein, including SoW 2000, may include a plurality of logic ICs and / or memory modules disposed in a rectilinear pattern (e.g., in rows and columns as illustrated in FIG. 20A); however, any suitable pattern is contemplated.[0359'] Each SOC of SOCs 2020 may be an IC that performs a wide variety of different functions, including computation (e.g., CPU or GPU calculations), data routing and management, memory, communication, power management, hardware acceleration, etc. SOCsAttorney Docket No. MTSH-152WO012020 may be directly or indirectly communicatively coupled with one or more input / output chips (I / O chips 2030), e.g., through one or more universal bridges (equivalently, interconnects; not shown). Each universal bridge may be configured to interface with a plurality of ICs through functional communication blocks disposed within each IC. These functional blocks 2036 (illustrated in FIG. 20B and FIG. 20C), also referred to as die-to-die (or D2D) blocks, may provide a universal interface through with different types of ICs may be communicatively coupled to a given IC. An SOC 2020 may have nominal dimensions of about 15 mm by about 25 mm, about 10 mm by about 15 mm, about 5 mm by about 5 mm; however, any suitable footprint is contemplated herein.|0360] ICs used in accordance with the present technology may include, but are not limited to, switches, memory modules (e.g., high-bandwidth memory (HBM)), network interface controllers (NICs), input / output (I / O) controllers, logic ICs, and / or other chips. Examples of logic ICs include graphics processing units (GPUs), central processing units (CPUs), data processing units (DPUs), tensor processing units (TPUs), and the like.|0361] A memory module may be an IC configured to store data and may include a dynamic random access memory (DRAM) module, a static random access memory (SRAM) module, a flash memory module, a solid-state drive (SSD), a non-volatile random access memory (NVRAM) module, a read-only memory (ROM) module (such as a floating-gate ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), one-time programmable ROM (OTPROM), or the like), or any suitable type of memory module.
[0362] For example, FIG. 20A illustrates a plurality of SOCs 2020 distributed in rows and columns on substrate 2010. A group of SOCs 2020 may be organized as unit blocks 2002, described in greater detail with respect to FIGS. 20B and 20C. SOCs 2020 of a unit block may have any suitable dimensions, for example a group of four SOCs 2020 may have a footprint of about 47 mm wide by about 43 mm long. SOCs 2020 in a unit block may have a footprint of about 10 mm in width, about 15 mm in width, about 20 mm in width, about 25 mm in width, about 30 mm in width, about 35 mm in width, about 40 mm in width, about 45 mm in width, about 50 mm in width, about 55 mm in width, about 60 mm in width, about 65 mm in width, about 70 mm in width, about 75 mm in width, or between any two aforementioned values. SOCs 2020 in a unit block may have a footprint of about 10 mm in length, about 15 mm in length, about 20 mm in length, about 25 mm in length, about 30 mm in length, about 35 mm in length, about 40 mm in length, about 45 mm in length, about 50 mm in length, about 55 mmAttorney Docket No. MTSH-152WO01in length, about 60 mm in length, about 65 mm in length, about 70 mm in length, about 75 mm in length, or between any two aforementioned values.(0363] SoW 2000 may include any suitable number of SOCs 2020, for example 50 SOCs, 80 SOCs, 96 SOCs, 100 SOCs, 112 SOCs, 120 SOCs, 150 SOCs, 200 SOCs, between about 50 SOCs and about 100 SOCs, between about 80 SOCs and about 150 SOCs, between about 100 SOCs and about 200 SOCs, or any suitable number of SOCs. SOCs disposed on an SoW in accordance with the present technology may have any suitable footprint across a substrate as illustrated in FIG. 20A, for example about 188 mm by about 215 mm, about 175 mm by about 200 mm, about 200 mm by about 200 mm, or any suitable footprint.(0364| SoW 2000 may include a plurality of I / O chips 2030, which may be configured to transfer data to and from components of SoW 2000 including SOCs 2020, memory modules (which may be integrated into SOCs 2020), wafer connector 2040, link chiplets (not shown), and / or switches, NICs, CPUs, GPUs, or other suitable components. I / O chips 2030 may provide low-latency, high-bandwidth data transfer between components, provide remote memory access for components (e.g., allowing a first SOC 2020 to access memory of a second SOC 2020 in a different portion of the SoW 2000, for example using non-uniform memory access (NUMA)), allowing load balancing, task division, thread and core coordination within and between SOCs 2020, coordinating memory sharing and performance optimization for workloads, etc. I / O chips may further enable data transfer required for hardware and Al acceleration.(0365] I / O chips 2030 may include one or more PCIe connectors or connectors for any suitable protocol (including compute express link (CXL)) enabling data transfer from a first component, through I / O chip 2030, to a second component. I / O chips 2030 may be communicatively coupled to one or more RDLs, LSIs, or other suitable arrangements of traces disposed on or in substrate 2010. I / O chips 2030 may provide a communications path between a first component external to SoW 2000 and a second component disposed on SoW 2000 (e.g., an SOC 2020). I / O chips 2030 may be temporarily connected to a data transfer path, e.g., through one or more cables, or may be permanently or semi -permanently connected to a data transfer path, e.g., through one or more bonds to leads, traces, pads, etc.|0366| SoW 2000 may include one or more wafer connectors 2040, which may be an electrical connection point to an external electrical component and / or to another SoW, and may provide communicative coupling between SoW 2000 and the external electrical component and / orAttorney Docket No. MTSH-152WO01other SoW. Wafer connectors 2040 may provide connections for data and / or power to the SoW 2000, and may be coupled to one or more RDLs, LSIs, voltage regulator modules (VRMs), electrical components (such as capacitors, resistors, etc.), traces, leads, pads, bonds, etc.
[0367] SoW 2000 may further include one or more link chiplets (not shown). Link chiplets may provide connections over longer distances than universal bridges. Link chiplets may be analogous to universal bridges, for example they may include a substrate made from a suitable material such as PCB, silicon, or other semiconductor. Link chiplets may include a plurality of traces as described herein for communicatively coupling a first component to a second component. In an aspect, link chiplets may be physically and / or communicatively coupled to one or more universal bridges. For example, a first SOC may be communicatively coupled to a link chiplet through two universal bridges. The link chiplet may then be communicatively coupled to a second SOC through two more universal bridges. Accordingly, link chiplets may include one or more patterns of exposed traces that mirror corresponding patterns of exposed traces on a universal bridge and therefore interface with one another when positioned accordingly. This may enable link chiplets to communicatively couple any two components of SoW 2000 that are compatible with one or more universal bridges.
[0368] SoW 2000 may further include a plurality of passthroughs 2060, which may include screw holes or other types of openings for fasteners such as screws, bolts, or the like, in substrate 2010. Passthroughs 2060 may accept fasteners that secure SoW 2000 to a PCB or other suitable structure such as a server or computing rack. Plurality of passthroughs 2060 may further be utilized for feeding through wiring or other communications pathways from a first side of substrate 2010 to the opposite side of substrate 2010. Substrate 2010 may include any suitable number of passthroughs 2060, for example 5 passthroughs, 15 passthroughs, 20 passthroughs, 22 passthroughs, 25 passthroughs, 40 passthroughs, or any suitable number of passthroughs.
[0369] FIGS. 20A-20C further illustrate an exemplary unit block 2002. Unit block 2002 may include a grouping of components including SOCs 2020 (for example four SOCs 2020, e.g., 2020a, 2020b, 2020c, and 2020d), I / O chips 2030 (e.g., two I / O chips 2030, four I / O chips 2030, six I / O chips 2030, zero I / O chips 2030, or any suitable number of I / O chips 2030), link chiplets (not shown; e.g., two link chiplets, four link chiplets, or a suitable number of link chiplets), and universal bridges (not shown; e.g., two or more universal bridges, four or more universal bridges, eight or more universal bridges, 16 or more universal bridges, 20 or more universal bridges, 24 or more universal bridges, or any suitable number of universal bridges).Attorney Docket No. MTSH-152WO01Unit blocks described herein may be modular, allowing for a unit block to be modified, repaired, or replaced without affecting other unit blocks in a given SoW.(0370] Unit blocks 2002 may further include a voltage regulator module (VRM 2070) which may be electrically coupled to SOCs 2020 in unit block 2002 in order to provide stable and regulated voltage to SOCs 2020. VRM 2070 may be configured to convert an input voltage (e.g., 12 V) down to a lower voltage required by logic ICs and / or memory that make up SOCs 2020 (e.g., 1.2 V). VRM 2070 may be configured to provide a constant voltage output even if the input voltage or SOC load changes and may improve power efficiency by modifying voltage in response to changes in load. These improvements to power efficiency may reduce waste heat generated by SOCs 2020 by generating less waste heat in the first place.(0371] VRM 2070 may include one or more metal oxide semiconductor field-effect transistors (MOSFETs), one or more inductors, one or more capacitors, one or more pulse width modulation (PWM) controllers, or other suitable components. A VRM 2070 used in accordance with SoW 2000 may be a linear voltage regulator, a switching voltage regulator (e.g., a buck converter), a multi-phase VRM, or any suitable type of VRM. A VRM 2070 may function to protect components of unit block 2002 from power fluctuations, increase stability and performance, allow for overclocking of logic ICs of SOCs 2020 by providing stable voltage under varying and increased loads, among other benefits.(0372] Unit block 2002 further provides an illustration of functional blocks 2036, which is shown in each of SOCs 2020 and I / O chips 2030. Functional blocks 2036 may include a suitable substrate, including any suitable substrate described herein. Each functional block 2036 may include a plurality of traces configured to provide a corresponding plurality of communications paths to a plurality of leads exposed on one or more surfaces of the functional block 2036.(0373] At least some of the plurality of traces may be communicatively coupled to one or more ICs such as field-programmable gate arrays (FPGAs) disposed on or in one or more universal bridges and / or functional blocks 2036. These universal bridge and functional block ICs may control routing of signals transmitted along traces of functional block 2036. For example, a functional block IC may route signals between a first lead and second lead along a first trace. Subsequently, the functional block IC may reroute signals transmitted from the first lead along a second trace to a third lead. The functional block IC may control some or all communications along traces of a functional block and route or reroute signals from a first communications pathAttorney Docket No. MTSH-152WO01to a second communications path. In an aspect, a communications path may include at least one trace and at least one lead. This reconfiguration of interconnections and routing paths may provide a benefit of enabling a functional block to effectively adapt to changes in components to which the functional block is connected, e.g., if a component is redesigned.
[0374] This improves routing flexibility and resilience to hardware changes for use cases such as Al workloads, which may result in frequent changes to workflow and data flow configurations. Further examples of inter-chip communications including link chiplets, functional blocks, and universal bridges are described herein in Section 5.
[0375] FIG. 20C is a side view of unit block 2002 illustrating SOCs 2020a and 2020b disposed on substrate 2010, I / O chip 2030 disposed adjacent to SOC 2020a, and functional blocks 2036 disposed within each of VO chip 2030 and SOCs 2020a and 2020b. FIG. 20C further illustrates a side view of VRM 2070 disposed on a side of substrate 2010 opposite SOCs 2020a and 2020b. VRM 2070 may be electrically and / or communicatively coupled to SOCs 2020a and / or 2020b and / or I / O chips 2030 through one or more RDLs or LSIs disposed within substrate 2010.
[0376] FIG. 21A illustrates an SoW 2100 in accordance with the present technology. SoW 2100 may be analogous to SoW 2000 in some aspects; SoW 2100 may include a substrate 2010, I / O chips 2030, wafer connectors 2040, plurality of passthroughs 2060, and VRMs 2070, each of which may be analogous to the corresponding components of SoW 2000. SoW 2100 may further include a plurality of three-dimensional integrated circuit (3DIC) stacks (3DIC stacks 2124), each of which may include an SOC (e.g., SOC 2120a and / or 2120b) with memory module (e.g., memory module 2122a and / or 2122b).
[0377] Memory modules (e.g., memory module 2122a and / or 2122b) may be arranged in layers, e.g., two, four, six, eight, ten, 12, 16, 20, or any suitable number of layers. Each layer may include a same amount of data storage capacity as the other layers, or the storage capacities of each layer may vary by layer.
[0378] 3DIC stacks 2124 may provide an advantage of higher bandwidth and lower latency for data transfer between logic ICs (e.g., CPUs, GPUs, SOCs, etc.) and data storage (e.g., memory modules 2122a / 2122b). Placing memory modules 2122a / 2122b closer to SOCs 2120a / 2120b may additionally reduce a signal path length between data storage and computational logic, reducing power losses associated with signal transfer along a communications path. 3DIC stacks 2124 may further enable heterogeneous integration of varying functionalities into aAttorney Docket No. MTSH-152WO01single structure, which may result in greater design flexibility and more adaptation to a fast pace of change in technologies such as artificial intelligence (Al) models, large language models (LLMs), etc. 3DIC stacks 2124 may further benefit from mixed process nodes, where leading-edge component manufacturing processes may be combined with older, more cost-effective processes to reduce total costs where advantageous.
[0379] FIG. 21 A shows unit block 2102, which may be analogous to unit block 2002, but with different components. FIGS. 21B and 21C illustrate unit block 2102 in greater detail. Unit block 2102 may include a grouping of components repeated across the SoW 2100 and may provide the advantage of giving system designers a consistent grouping of components with which to work and design SoW 2100. Unit block 2102 may include a number of 3DIC stacks 2124, e.g., four, which are spaced in pattern such as 2x2. As with unit block 2002, unit block 2102 may have any suitable footprint, such as about 47 mm by about 43 mm as shown in FIG.21B.
[0380] Similarly to FIG. 20B, FIG. 2 IB shows a top view of unit block 2102, including 3DIC stacks 2124 and VRM 2070. VRM 2070 may supply regulated power to each 3DIC stack 2124 in unit block 2102. Substrate 2010 and a plurality of passthroughs 2060 are shown but are not considered a part of any unit block disclosed herein, including unit block 2102. Each 3DIC stack of 3DIC stacks 2124 may be communicatively coupled with a neighboring (e.g., immediately adjacent) 3DIC stack through one or more functional blocks 2036, which may be communicatively coupled with a corresponding functional block disposed in a neighboring 3DIC stack. Either or both of SOCs 2120a / 2120b and memory modules 2122a / 2122b in a given 3DIC stack 2124 may include one or more functional blocks 2036, and a first 3DIC stack may be communicatively coupled with a second, neighboring 3DIC stack through a plurality of functional blocks 2036.
[0381] For example, a first 3DIC stack 2124a may include a first SOC 2120a and a first memory module 2122a. A second 3DIC stack 2124b may include a second SOC 2120b and a second memory module 2122b. A third 3DIC stack 2124c may include a third SOC and a second memory module. A fourth 3DIC stack 2124d may include a fourth SOC and a fourth memory module. First 3DIC stack 2124a may be communicatively coupled to second 3DIC stack 2124b through some, any, or all of the following: a first functional block 2036a disposed in SOC 2120a may be communicatively coupled to a second functional block 2036b disposed in SOC 2120b; a third functional block 2036c disposed in memory module 2122a may be communicatively coupled to a fourth functional block disposed in memory module 2122b; aAttorney Docket No. MTSH-152WO01plurality of functional blocks 2036 disposed in both SOC 2120a and memory module 2122a may be communicatively coupled to a corresponding second plurality of functional blocks disposed in both SOC 2120b and memory module 2122b.
[0382] FIG. 22A illustrates an SoW 2200 analogous to SoW 2000 and / or SoW 200 but having a plurality of 3DIC stacks 2224 with a corresponding plurality of memory modules 2222 disposed laterally to one or more 3DIC stacks 2224. The memory modules 2222 may be communicatively coupled with one or more neighboring 3DIC stacks 2224. In an aspect, two 3DIC stacks 2224 may utilize the same memory module 2222 for data storage and transfer. This arrangement may provide improved data transfer and storage capabilities for logic ICs disposed in 3DIC stacks 2224 but may reduce an overall heat generated by SoW 2200 due to relatively fewer logic ICs in SOCs 2220 as compared to, e.g., SoW 2000. SOCs and memory modules included in SoW 2200 may be grouped in unit blocks 2202 analogous to unit blocks 2002 and / or unit blocks 2102.
[0383] FIGS. 22B and 22C illustrate top and side views, respectively, of a unit block 2202 in accordance with the present technology. Unit block 2202 may include a plurality of SOCs 2220, for example two SOCs 2220, each SOC 2220 having a memory module 2222 disposed on top of SOC 2220. This combination of SOC and memory module may be referred to as a 3DIC stack 2224. Unit block 2202 may additionally include a plurality of memory modules 2222 disposed laterally to each 3DIC stack 2224. Unit block 2202 may further include a VRM 2070. VRM 2070 may supply regulated power to each 3DIC stack 2224 in unit block 2202. Substrate 2010 and a plurality of passthroughs 2060 are shown but are not considered a part of any unit block disclosed herein, including unit block 2202.
[0384] Each 3DIC stack of 3DIC stacks 2224 may be communicatively coupled with a neighboring (e.g., immediately adjacent) 3DIC stack through one or more functional blocks 2036, which may be communicatively coupled with a corresponding functional block disposed in a neighboring 3DIC stack. Either or both of SOCs 2220 and memory modules 1 in a given 3DIC stack 2224 may include one or more functional blocks 2036, and a first 3DIC stack may be communicatively coupled with a second, neighboring 3DIC stack through a plurality of functional blocks 2036.
[0385] For example, a first 3DIC stack 2224a may include a first SOC (e.g., SOC 2220) and a first memory module (e.g., memory module 2222). A second 3DIC stack 2224b may include a second SOC and a second memory module. First 3DIC stack 2224a may be communicativelyAttorney Docket No. MTSH-152WO01coupled to second 3DIC stack 2224b through some, any, or all of the following: a first functional block 2036a disposed in SOC 2220 may be communicatively coupled to a second functional block 2036b disposed in SOC; a third functional block 2036c disposed in memory module 2222 may be communicatively coupled to a fourth functional block 2036d disposed in memory module; a plurality of functional blocks 2036 disposed in both SOC 2220 and memory module 2222 may be communicatively coupled to a corresponding second plurality of functional blocks disposed in both the second SOC and the second memory module.(0386] One or more additional functional blocks 2036 may communicatively couple 3DIC stack 2224a with memory module 2222c and with memory module 2222d, while one or more other functional blocks 2036 may communicatively couple 3DIC stack 2224b with memory module 2222e and with memory module 2222f. As described above, substrate 2010 may include one or more RDLs or LSIs that provide communications paths between components disposed on SoW 2200, including components of unit blocks 2202. For example, memory module 2222c may be communicatively coupled to memory module 2222f via an RDL disposed in substrate 2010.(0387] FIGS. 23A-23C illustrate an SoW 2300 and unit block 2302, which are largely analogous to SoW 2200 and unit block 2202 and may include all of the respective components of SoW 2200 and / or unit block 2202, but further including a plurality of auxiliary memory modules 2326 disposed in gaps between 3DIC stacks 2324. 3DIC stacks 2324 may include at least an SOC and a memory module and may be similar to or the same as 3DIC stacks 2224. These auxiliary memory modules 2326 may serve to provide more efficient use of surface area available on substrate 2010. Auxiliary memory modules 2326 may include one or more functional blocks 2036 through which the auxiliary memory modules 2326 are communicatively coupled to 3DIC stacks 2324 through SOCs 2320a and / or 2320b and / or memory modules 2322. SOCs 2320a and 2320b are not directly visible in FIGS. 23A-23C as they are blocked by memory modules 2322 and / or memory modules 2326 but occupy the same relative position in 3DIC stacks 2324 as SOCs 2220 in 3DIC stacks 2224.(0388] As with unit block 2302, each 3DIC stack of 3DIC stacks 2324 may be communicatively coupled with a neighboring (e.g., immediately adjacent) 3DIC stack through one or more functional blocks 2036, which may be communicatively coupled with a corresponding functional block disposed in a neighboring 3DIC stack. Either or both of SOCs 2320a and / or 2320b and memory modules 2322 in a given 3DIC stack 2324 may include oneAttorney Docket No. MTSH-152WO01or more functional blocks 2036, and a first 3DIC stack may be communicatively coupled with a second, neighboring 3DIC stack through a plurality of functional blocks 2036.(0389] Again, in unit block 2302 as with unit block 2202, a first 3DIC stack 2324a may include a first SOC 2320a and a first memory module 2322a. A second 3DIC stack 2324b may include a second SOC 2320b and a second memory module. First 3DIC stack 2324a may be communicatively coupled to second 3DIC stack 2324b through some, any, or all of the following: a first functional block 2036a disposed in SOC 2320a may be communicatively coupled to a second functional block 2036b disposed in SOC 2320b; a third functional block 2036c disposed in memory module 2322a may be communicatively coupled to a fourth functional block 2036d disposed in the second memory module; a plurality of functional blocks 2036 disposed in both SOC 2320a and memory module 2322a may be communicatively coupled to a corresponding second plurality of functional blocks disposed in both SOC 2320b and the second memory module.
[0390] One or more additional functional blocks 2036 may communicatively couple 3DIC stack 2324a with memory module 2322c and with memory module 2322d while one or more other functional blocks 2036 may communicatively couple 3DIC stack 2324b with memory module 2322e and with memory module 2322f. As described above, substrate 2010 may include one or more RDLs or LSIs that provide communications paths between components disposed on SoW 2200, including components of unit blocks 2302. For example, memory module 2322c may be communicatively coupled to memory module 2322f via an RDL disposed in substrate 2010.4.2. An Example Method for Making a Memory Architecture in SOW
[0391] The SoWs disclosed herein may be manufactured using various packaging processes including, but not limited to, a chip-first manufacturing process and a chip-last manufacturing process as described above in Sections 1-3.
[0392] FIG. 24 shows a non-limiting example of a manufacturing method 2400 to form a memory architecture in a SoW (e.g., SoW 2000, 2100, 2200, and / or 2300). Method 2400 begins at step 2402, where a plurality of components, including SOCs are disposed on a substrate to form a unit block. Then, at step 2404, a plurality of unit blocks are arranged on the substrate to form a SoW.
[0393] It should be appreciated that the method 2400 is a non-limiting example of a manufacturing process to form a memory architecture in a SoW and alterations to the methodAttorney Docket No. MTSH-152WO012400 are readily contemplated herein. For example, the steps of method 2400 may be performed in any order. For example, as outlined above, the SOCs may be directly or indirectly communicatively coupled with one or more VO chips through one or more universal bridges, which may be configured to interface with a plurality of ICs through functional communication blocks disposed within each IC. A group of SOCs may be organized as unit blocks and the SoW may include any suitable number of SOCs. As outlined above, the unit blocks may be distributed in rows and columns on the substrate. However, the SOCs may have any suitable footprint across the substrate.5. An Example Universal Bridge for Chiplet Interface and Communication in SOW
[0394] One challenge in SoW-based Al accelerators is accommodating evolving IC configurations and functionalities while maintaining efficient routing and interconnectivity. Traditional passive interconnections (equivalently, interconnects), such as redistribution layers (RDL) and local silicon interconnects (LSI), may require modifications to an entire hardware routing network if an IC layout changes. This lack of flexibility can result in increased design time, greater complexity, and potential performance bottlenecks, especially in Al applications with dynamically changing workloads and configurations. To address some or all of these issues, some implementations of the present disclosure include improved interconnect solutions.
[0395] Some implementations of the present disclosure relate to flexible, programmable interconnect layers within an SoW structure. Where traditional designs utilize passive interconnections, the present technology includes active interconnection components (such as FPGAs) that can be programmed to adapt routing for inter-component communication as needed. The present technology may include some or all of the following benefits:
[0396] Programmable Active Interconnection: The active interconnect layer may allow for reconfiguration, providing adaptable routing paths without requiring physical modifications to the entire RDL network. This flexibility enables rapid adjustments to accommodate varying chiplet configurations and workload demands, providing benefits for Al workloads where configurations can change frequently.
[0397] Enhanced Routing Flexibility: By using active interconnects, the present technology can support dynamic and modular routing adjustments within the SoW, enabling optimized and enhanced pathways based on the specific workload and chiplet arrangements. This flexibility improves data flow efficiency and reduces latency for Al tasks.Attorney Docket No. MTSH-152WO01
[0398] Strategic Placement of Active Interconnects: The active interconnect components can be strategically placed at different locations within the SoW architecture, e.g., within chiplets, within an RDL / LSI, within a substrate, within components, or elsewhere as is suitable.
[0399] This active interconnection approach reduces the need for extensive re-routing when an SoW configuration changes, enhancing adaptability, reducing redesign time, and improving performance for diverse Al workloads. Active interconnection provides flexible, programmable routing pathways that can adapt to changes in SoW configuration, offering significant advantages over traditional passive interconnects. These active interconnections enable an adaptable routing network that can be customized based on specific chiplet requirements and workload characteristics. Further, embedded instruction sets within these active interconnections allow for the generation of dynamic, ad-hoc paths within a fixed hardware network as workload requirements and / or IC designs change, which may enable such a fixed hardware network to continue providing optimized performance where a traditional passive network could be rendered obsolete.
[0400] Some implementations of the present disclosure relate to active interconnects that are scalable and support a broad range of SoW configurations and updated IC designs, allowing hardware to keep pace with fast-evolving Al model and software changes.
[0401] Following below are more detailed descriptions of various concepts related to, and implementations of, a SoW and servers including SoWs. In particular, the present technology includes improvements to communicative coupling between chips, e.g., between GPUs and memory modules. It should be appreciated that various concepts introduced above and discussed in greater detail below may be implemented in multiple ways. Examples of specific implementations and applications are provided primarily for illustrative purposes so as to enable those skilled in the art to practice the implementations and alternatives apparent to those skilled in the art.
[0402] The figures and example implementations described below are not meant to limit the scope of the present implementations to a single embodiment. Other implementations are possible by way of interchange of some or all of the described or illustrated elements. Moreover, where certain elements of the disclosed example implementations may be partially or fully implemented using known components, in some instances only those portions of such known components that are necessary for an understanding of the present implementations areAttorney Docket No. MTSH-152WO01described, and detailed descriptions of other portions of such known components are omitted so as not to obscure the present implementations.5.1. An Example Universal Bridge for Chiplet Interface and Communication in SOW (0403] FIG. 25 is a top view of a SoW 2500 according to an embodiment. The SoW 2500 includes at least a semiconductor wafer 2502 and a plurality of chips 2510.[0404 The semiconductor wafer 2502 may include one or more semiconducting materials such as silicon, silicon oxide, silicon dioxide, silicon nitride, gallium arsenide, sapphire, silicon carbine, indium phosphide, gallium nitride, germanium, and / or another semiconducting material.
[0405] The chips 2510 may be mounted on or above the semiconductor wafer 2502. The chips 2510 can be multi-functional and can include logic chips (chips may be equivalently referred to as integrated circuits (ICs)), switches, memory modules (e.g., high-bandwidth memory (HBM)), network interface controllers (NICs), input / output (I / O) controllers, and / or other chips. Examples of logic chips include GPUs, central processing units (CPUs), data processing units (DPUs), tensor processing units (TPUs), systems on a chip (SoCs), and the like.
[0406] In an aspect, chips 2510 may include a plurality of different chips configured to perform respective functions. For example, chips 2510 may include one or more GPUs 2520, one or more memory modules 2522, one or more CPUs 2530, one or more input / output communications chiplets (VO chiplets 2542), one or more switch chiplets 2544, one or more NICs 2546, or the like. Each chip of chips 2510 may be configured to perform substantially the same or substantially a different function as the remaining chips of chips 2510. For example, GPUs 2520 may be configured to execute massively parallel algorithms but may not be configured to (and / or do not have the capability to) perform packet transmission and reception, while NICs 2546 may be configured to perform packet transmission and reception but may not be configured to execute massively parallel algorithms.[0407| A memory module may be an IC configured to store data and may include a dynamic random access memory (DRAM) module, a static random access memory (SRAM) module, a flash memory module, a solid-state drive (SSD), a non-volatile random access memory (NVRAM) module, a read-only memory (ROM) module (such as a floating-gate ROM, erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), one-time programmable ROM (OTPROM), or the like), or any suitable type of memory module.Attorney Docket No. MTSH-152WO01|0408] The chips 2510 can be selected in any combination to provide an overall function or architecture for the SoW 2500. For example, the chips 2510 can be selected such that the SoW 2500 can function as server or a portion of a server. Examples of servers that the SoW 2500 can function as can include an artificial intelligence (Al) training server, an Al inference server, a web server, a database server, an email server, a web proxy server, a domain name server (DNS), an application server, and / or another server.(0409] Each chip of chips 2510 may include chip-specific connectors that do not match chipspecific connectors in different types of chips. For example, GPUs 2520 may each have a first chip-specific connector, CPUs 2530 may each have a second chip-specific connector different from the first chip-specific connector, I / O chiplets 2542 may each have a third chip-specific connector different from the first and second chip-specific connectors, etc. This may make communicative coupling between different types of chips difficult, particularly if the design of a particular chip changes over time.
[0410] The chips 2510 are electrically coupled to one or more modules or structures that are mounted on or above the semiconductor wafer 2502 and that are electrically coupled to the chips 2510. The modules / structures can include a power module 2560 and / or a SoW connector 2540. The power module 2560 can include a voltage converter that can modulate an external supply voltage to a chip supply voltage that is at an appropriate level to power some or all of the chips 2510. The SoW connector 2540 provides an electrical connection point to an external electrical component and / or to another SoW. One or more of the modules can be located directly above one or more of the chips 2510 but are not illustrated in FIG. 25 for illustration purposes only so as to not obscure the chips 2510.
[0411] SoW 2500 may be analogous to or the same as SoW 2500 except that the chips 2510 in SoW 2500 include CPUs, GPUs, switches, NICs, I / O chips, and memory (e.g., HBM). The VO chips and the memory modules are represented as 2542 and 2522, respectively. This configuration / architecture can function as an Al training server or a portion of an Al training server. In a specific example, the SoW 2500 can include 8 GPU chips, 2 CPU chips, 9 NIC chips, 2 switch chips, 18 memory modules, and 16 VO chips. The configuration in this specific example can replace (e.g., function as) a conventional Peripheral Component Interconnect Express (PCIE) card for a conventional Al training server, where the PCIE card includes the same number of GPU, CPU, and switch chips (i.e., 8 GPU chips, 2 CPU chips, 9 NIC chips, 2 switch chips, 18 memory modules, and 16 VO chips). In some embodiments, the SoW 2500 can include 16 GPU chips, 4 CPU chips, 18 NIC chips, 4 switch chips, 16 memory modules,Attorney Docket No. MTSH-152WO01and 32 I / O chips in which case the SoW 2500 can replace (e.g., function as) two or more PCIE cards (for example, six, eight, or more PCIe cards) for a conventional Al training server. Additionally, or alternatively, the SoW 2500 can include memory (e.g., HBM) chips, NIC chips, and / or other chips. Additional or fewer numbers of each type of chip 2510 can be included in other embodiments.
[0412] A technical advantage of the configuration / architecture of SoW 2500 is that the memory modules 2522 are located in close physical proximity (e.g., at the millimeter scale and / or at the micron scale) to the GPU chips, which improves processing speed and / or reduces latency.[04131 FIG. 26 illustrates a cross-sectional view of a SoW 2600 in accordance with the present technology. SoW 2600 may include communication layers 2511. Communication layers 2511 may include first redistribution layer (first RDL 2512), second redistribution layer (second RDL 2514), and local silicon interconnect layer (LSIs 2516; including bridge 2516-1, bridge 2516-2, and bridge 2516-3). Communication layers 2511 may further include traces 2518, which may provide a plurality of communications paths through communication layers 2511 to and from chips 2510, e.g., between GPUs 2520, memory modules 2522, CPUs 2530, and I / O chiplets 2542, or from one or more of those components to one or more additional components either on SoW 2600 or remote from SoW 2600.
[0414] Chips included in SoW 2600 may include one or more GPUs 2520, one or more memory modules 2522, one or more CPUs 2530, one or more I / O chiplets 2542, or any suitable chips. Each chip may be communicatively coupled to each other and / or other components through first RDL 2512, second RDL 2514, and LSIs 2516 by way of traces 2518 embedded in each layer.
[0415] First RDL 2512, second RDL 2514, and LSI 2516, may enable fan-out packaging in which traces 2518 communicatively coupling an individual chip spread out across a greater area than that of the area in contact with a substrate on which it is disposed, as well as routing the traces 2518 around other components. This may enable smaller chips to remain connected to a large number of other components as well as provide additional routes for data and signal transfer, increasing bandwidth.
[0416] Traces 2518 may be made of a metal such as copper, gold, aluminum, or any suitable material. Traces 2518 may be embedded in respective substrates for each of first RDL 2512, second RDL 2514, and / or LSIs 2516. Substrates of first RDL 2512, second RDL 2514, and / orAttorney Docket No. MTSH-152WO01LSIs 2516 may include printed circuit boards (PCBs), silicon, silicon dioxide, silicon nitride, or other suitable material.(0417] SoW 2600 may include bump layer 2570, which may include a plurality of bump bonds 2572 communicatively coupling chips or other components of SoW 2600 with SoW connector 2540, power module 2560, or any suitable components. Plurality of bump bonds 2572 may include bump bonds, micro-bump bonds, controlled collapse of chip connection (C4) bonds, solder bonds, hybrid bonds, wire bonds, or any suitable type of bonds.(0418] SoW 2600 may further include a thermal module 2580, which may provide improved heat transfer from one or more GPUs 2520, one or more memory modules 2522, one or more CPUs 2530, and / or I / O chiplets 2542. Thermal module 2580 may include one or more thermal interface materials (TIMs), one or more thermal plates or heat spreaders (e.g., a copper plate for enhancing and evening out heat transfer), one or more boiling enhancement coatings (BECs), or other components for improving heat transfer from chips in SoW 2600.(0419] FIG. 27A illustrates a top view of a SoW 2700 in accordance with the present technology. SoW 2700 may include a plurality of unit blocks 2702, discussed in greater detail with respect to FIG. 27B. SoW 2700 may include a substrate 2710, which may be a silicon wafer in an aspect. Substrate 2710 may include one or more RDLs, LSIs, or other arrangements of traces within substrate 2710.(0420] SoW 2700 may include a plurality of systems-on-a-chip (SOCs 2720). Each SOC of SOCs 2720 may be an IC that performs a wide variety of different functions, including computation (e.g., CPU or GPU calculations), data routing and management, memory, communication, power management, hardware acceleration, etc. SOCs 2720 may be directly or indirectly communicatively coupled with one or more I / O chips 2730, e.g., through one or more universal bridges 2732 (equivalently, interconnects). Each universal bridge 2732 may be configured to interface with a plurality of ICs through functional communication blocks disposed within each IC. These functional blocks 2736 (illustrated in FIG. 27B), also referred to as die-to-die (or D2D) blocks, may provide a universal interface through with different types of ICs may be communicatively coupled to a given IC. An SOC 2720 may have nominal dimensions of about 15 mm by about 25 mm, about 10 mm by about 15 mm, about 5 mm by about 5 mm; however, any suitable footprint is contemplated herein.(0421] For example, FIG. 27A illustrates a plurality of SOCs 2720 communicatively coupled to one another via one or more universal bridges 2732. A universal bridge may include aAttorney Docket No. MTSH-152WO01substrate made of a suitable material such as PCB or semiconductor such as silicon. The substrate may further include a plurality of traces, which may be gold, copper, aluminum, etc., as appropriate. These traces terminate in leads or pads exposed on a plurality of surfaces of the universal bridge 2732. The exposed leads or pads may be arranged in a pattern that mirrors a second pattern on exposed leads of a functional block for an IC such as an SOC 2720.
[0422] In an aspect, a universal bridge may include at least two surfaces having exposed leads or pads that have the same pattern. The universal bridge may be rotationally symmetrical, e.g., if the universal bridge is divided by a plane parallel to a first surface and a second surface of the universal bridge and the universal bridge is flipped 180° such that each surface is rotated to the other side of the plane relative to the one it started on, the leads of the first and second surfaces would be the same for the first surface and the second surface when each surface is on the same side of the plane. This type of symmetry may be referred to as glide symmetry or glide symmetric.
[0423] SoW 2700 may include a plurality of I / O chips 2730, which may be configured to transfer data to and from components of SoW 2700 including SOCs 2720, memory modules (which may be integrated into SOCs 2720), wafer connector 2740, link chiplet 2750, and / or switches, NICs, CPUs, GPUs, or other suitable components. I / O chips 2730 may provide low-latency, high-bandwidth data transfer between components, provide remote memory access for components (e.g., allowing a first SOC 2720 to access memory of a second SOC 2720 in a different portion of the SoW 2700, for example using non-uniform memory access (NUMA)), allowing load balancing, task division, thread and core coordination within and between SOCs 2720, coordinating memory sharing and performance optimization for workloads, etc. I / O chips may further enable data transfer required for hardware and Al acceleration.
[0424] I / O chips 2730 may include one or more PCIe connectors or connectors for any suitable protocol enabling data transfer from a first component, through I / O chip 2730, to a second component. I / O chips 2730 may be communicatively coupled to one or more RDLs, LSIs, or other suitable arrangements of traces disposed on or in substrate 2710. I / O chips 2730 may provide a communications path between a first component external to SoW 2700 and a second component disposed on SoW 2700 (e.g., an SOC 2720). I / O chips 2730 may be temporarily connected to a data transfer path, e.g., through one or more cables, or may be permanently or semi-permanently connected to a data transfer path, e.g., through one or more bonds to leads, traces, pads, etc.Attorney Docket No. MTSH-152WO01
[0425] SoW 2700 may include one or more wafer connectors 2740, which may be an electrical connection point to an external electrical component and / or to another SoW, and may provide communicative coupling between SoW 2700 and the external electrical component and / or other SoW. Wafer connectors 2740 may provide connections for data and / or power to the SoW 2700, and may be coupled to one or more RDLs, LSIs, voltage regulator modules (VRMs), electrical components (such as capacitors, resistors, etc.), traces, leads, pads, bonds, etc. J0426] SoW 2700 may further include link chiplets 2750. Link chiplets 2750 may provide connections over longer distances than universal bridges 2732. Link chiplets 2750 may be analogous to universal bridges 2732, for example they may include a substrate made from a suitable material such as PCB, silicon, or other semiconductor. Link chiplets 2750 may include a plurality of traces as described herein for communicatively coupling a first component to a second component. In an aspect, link chiplets 2750 may be physically and / or communicatively coupled to one or more universal bridges 2732. For example, a first SOC may be communicatively coupled to a link chiplet 2750 through two universal bridges. The link chiplet 2750 may then be communicatively coupled to a second SOC through two more universal bridges. Accordingly, link chiplets 2750 may include one or more patterns of exposed traces that mirror corresponding patterns of exposed traces on a universal bridge 2732 and therefore interface with one another when positioned accordingly. This may enable link chiplets 2750 to communicatively couple any two components of SoW 2700 that are compatible with one or more universal bridges 2732.
[0427] SoW 2700 may further include a plurality of passthroughs 2760, which may include screw holes or other types of openings for fasteners such as screws, bolts, or the like, in substrate 2710. Passthroughs 2760 may accept fasteners that secure SoW 2700 to a PCB or other suitable structure such as a server or computing rack. Plurality of passthroughs 2760 may further be utilized for feeding through wiring or other communications pathways from a first side of substrate 2710 to the opposite side of substrate 2710. Substrate 2710 may include any suitable number of passthroughs 2760, for example 5 passthroughs, 15 passthroughs, 20 passthroughs, 22 passthroughs, 25 passthroughs, 40 passthroughs, or any suitable number of passthroughs.
[0428] FIG. 27A and FIG. 27B further illustrate an exemplary unit block 2702. Unit block 2702 may include a grouping of components including (SOCs 2720, for example four SOCs 2720), I / O chips 2730 (e.g., two I / O chips 2730, four I / O chips 2730, six I / O chips 2730, zero I / O chips 2730, or any suitable number of I / O chips 2730), link chiplets 2750 (e.g., two linkAttorney Docket No. MTSH-152WO01chiplets 2750, four link chiplets 2750, or a suitable number of link chiplets 2750), and universal bridges 2732 (e.g., two or more universal bridges 2732, four or more universal bridges 2732, eight or more universal bridges 2732, 16 or more universal bridges 2732, 20 or more universal bridges 2732, 24 or more universal bridges 2732, or any suitable number of universal bridges 2732).
[0429] Unit blocks 2702 may further include a voltage regulator module (VRM 2770) which may be electrically coupled to SOCs 2720 in unit block 2702 in order to provide stable and regulated voltage to SOCs 2720. VRM 2770 may be configured to convert an input voltage (e.g., 12 V) down to a lower voltage required by logic ICs and / or memory that make up SOCs 2720 (e.g., 1.2 V). VRM 2770 may be configured to provide a constant voltage output even if the input voltage or SOC load changes and may improve power efficiency by modifying voltage in response to changes in load. These improvements to power efficiency may reduce waste heat generated by SOCs 2720 by generating less waste heat in the first place.
[0430] VRM 2770 may include one or more metal oxide semiconductor field-effect transistors (MOSFETs), one or more inductors, one or more capacitors, one or more pulse width modulation (PWM) controllers, or other suitable components. A VRM 2770 used in accordance with SoW 2700 may be a linear voltage regulator, a switching voltage regulator (e.g., a buck converter), a multi-phase VRM, or any suitable type of VRM. A VRM 2770 may function to protect components of unit block 2702 from power fluctuations, increase stability and performance, allow for overclocking of logic ICs of SOCs 2720 by providing stable voltage under varying and increased loads, among other benefits.10431] Unit block 2702 further provides an illustration of functional blocks 2736, which is shown in each of SOCs 2720, link chiplets 2750, and VO chips 2730. Functional blocks 2736 may include a suitable substrate, including any suitable substrate described herein. Each functional block 2736 may include a plurality of traces configured to provide a corresponding plurality of communications paths to a plurality of leads exposed on one or more surfaces of the functional block 2736.
[0432] At least some of the plurality of traces may be communicatively coupled to one or more ICs such as field-programmable gate arrays (FPGAs) disposed on or in one or more universal bridges 2732 and / or functional blocks 2736. These universal bridge and functional block ICs may control a routing of signals transmitted along traces of functional block 2736. For example, a functional block IC may route signals between a first lead and second lead along a first trace.Attorney Docket No. MTSH-152WO01Subsequently, the functional block IC may reroute signals transmitted to the first lead along a second trace to a third lead. The functional block IC may control some or all communications along traces of a functional block and route or reroute signals from a first communications path to a second communications path. In an aspect, a communications path may include at least one trace and at least one lead. This reconfiguration of interconnections and routing paths may provide a benefit of enabling a functional block to effectively adapt to changes in components to which the functional block is connected, e.g., if a component is redesigned.(0433] This improves routing flexibility and resilience to hardware changes for use cases such as Al workloads, which may result in frequent changes to workflow and data flow configurations.(0434] FIG. 28A illustrates an SoW 2800 in accordance with the present technology. SoW 2800 may include a plurality of components analogous to SoW 2700, such as substrate 2810, SOCs 2820, link chiplets 2850, universal bridges 2832, and passthroughs 2860. SoW 2800 may further include one or more three-dimensional integrated circuit (3DIC) stacks that include an SOC 2820 as well as one or more memory modules 2822. Memory modules 2822 may be placed laterally to SOC 2820 on substrate 2810 or may be disposed on top of one or more SOCs 2820.(0435] Additionally, or alternatively, memory modules 2822 may be arranged in layers, e.g., two, four, six, eight, ten, 12, 16, 20, or any suitable number of layers. Each layer may include a same amount of data storage capacity as the other layers, or the storage capacities of each layer may vary by layer.(0436] In an aspect, SoW 2800 may exclude I / O chips analogous to VO chips 2730 and wafer connectors analogous to wafer connector 2740. Alternatively, SoW 2800 may include I / O chips analogous to VO chips 2730 and wafer connectors analogous to wafer connector 2740, but these components may be disposed on a side of substrate 2810 opposite the side shown in FIG.28A.(0437] Analogously to SoW 2700, components disposed on SoW 2800 may be grouped into unit blocks 2802. An exemplary unit block 2802 is illustrated in FIG. 28B. Unit block 2802 may include one or more SOCs 2820, one or more universal bridges 2832, one or more link chiplets 2850, and a VRM 2870. Unit blocks 2802 may further include one or more memory modules 2822, which may be disposed on substrate 2810 laterally to SOC 2820, or may beAttorney Docket No. MTSH-152WO01disposed on top of an SOC 2820 as part of a 3DIC stack 2880. Such a 3DIC stack 2880 may include a memory module 2822, base layer 2824, and one or more functional blocks 2836. (0438] FIG. 29A illustrates a SoW 2900 in accordance with the present technology. SoW 2900 may include components analogous to SoW 2700 and / or SoW 2800, including substrate 2910 (which may be analogous to substrate 2710 and / or substrate 2810), SOCs 2920 (which may be analogous to SOCs 2720 and / or SOCs 2820), memory modules 2922 (which may be analogous to one or more memory modules 2822), universal bridges 2932 (which may be analogous to one or more universal bridges 2732 and / or universal bridges 2832), and link chiplets 2950 (which may be analogous to link chiplets 2750 and / or link chiplets 2850).[0439| SoW 2900 may include unit block 2902, which may be analogous to unit blocks 2702 and / or unit blocks 2802, or which may differ. FIG. 29B illustrates an exemplary unit block 2902. For example, unit block 2902 may include SOCs 2920, memory modules 2922, universal bridges 2932, one or more link chiplets 2950, and VRM 2970. Unit block 2902 may include modular hub 2980. Modular hub 2980 may be a 3DIC including an SOC and at least one memory module 2922 disposed on the SOC, or modular hub 2980 may exclude an SOC. Modular hub 2980 may further include base layer 2924, one or more communications interfaces 2934 (which may be a PCIe or other suitable interface), and one or more functional blocks 2936.5.2. An Example Method for Making an Universal Bridge for Chiplet Interface and Communication in SOW|0440) The SoWs disclosed herein (e.g., SoWs 2500, 2700, 2800, 2900) may be manufactured using various packaging processes including, but not limited to, a chip-first manufacturing process and a chip-last manufacturing process as outlined above in Sections 1-3.
[0441] FIG. 30 shows a non-limiting example of a manufacturing method 3000 to form a universal bridge for chiplet interface and communication in an SoW (e.g., SoW 2500, 2700, 2800, and / or 2900). Method 3000 begins at step 3002, where a plurality of components, including SOCs are disposed on a substrate to form a unit block. Then, at step 3004, a plurality of unit blocks are arranged on the substrate to form an SoW. Then, at step 3006, at least one SOCs disposed on the substrate may be communicatively coupled to at least another SOC disposed on the substrate via one or more universal bridges.(04421 It should be appreciated that method 3000 is a non-limiting example of a manufacturing process to form a universal bridge for chiplet interface and communication in an SoW andAttorney Docket No. MTSH-152WO01alterations to the method 3000 are readily contemplated herein. For example, the steps of method 3000 may be performed in any order. For example, as outlined above, each SoW may also include link chiplets which may provide connections over longer distances than universal bridges. A group of SOCs may be organized as unit blocks and the SoW may include any suitable number of SOCs and universal bridges.6. Overview of Immersion-Cooling System
[0443] FIG. 31 depicts aspects of an immersion cooling system 500 for dissipating heat from one or more heat-generating components such as semiconductor die packages 505 via immersion cooling. Each package 505 can include one or more semiconductor dies that produce heat when the system is in operation. The immersion cooling system 500 in the illustrated example of FIG. 31 is a two-phase immersion cooling system, though the invention may also be implemented in a single-phase immersion cooling system.
[0444] One or more semiconductor die packages 505 may be 3DIC stacks in accordance with the present technology. For example, one or more semiconductor die packages 505 may include a logic IC and at least one memory module bonded to the logic IC using a hybrid bond or micro-bump bond. Additionally, or alternatively, the one or more semiconductor die packages 505 may be part of one or more SoWs, such as SoWs described herein (e.g., SoWs lOOa-lOOc, 900a-900d, 1500a-1500d, 1500a’-1500d’, 2000, 2100, 2200, 2300, 2500, 2700, 2800, and / or 2900).
[0445] Immersion cooling systems may provide particular advantage to 3DIC stacks and SoWs, such as those described herein, due to the lower surface area to volume ratio of a 3DIC stack or SoW compared to the individual components of the 3DIC stack or SoW (e.g., a bonded logic IC and memory module will have a lower surface area to volume ratio than the combined surface area to volume ratio of the physically separated logic IC and memory module) as well as the additional heat generated by state of the art logic ICs. This lower surface area to volume ratio means waste heat generated by the 3DIC stack may not be as efficiently dissipated and may require better cooling performance than air cooling can provide. Two-phase immersion cooling in particular can provide this additional heat removal required by 3DIC stacks and SoWs (e.g., SoWs lOOa-lOOc, 900a-900d, 1500a-1500d, 1500a’-1500d’, 2000, 2100, 2200, 2300, 2500, 2700, 2800, and / or 2900).
[0446] Immersion cooling system 500 includes a container such as tank 520 filled, at least in part, with immersion cooling liquid 564. The immersion cooling system 500 can further includeAttorney Docket No. MTSH-152WO01at least one chiller 580 that flows a heat-transfer fluid through at least one condenser tube 570 that is disposed in the tank 520 and headspace 508. Condenser tubes 570 and chiller 580 may be part of a heat exchanger. The packages 505 can be mounted on one or more printed circuit boards (PCBs) 557 that are immersed, at least in part, in the immersion cooling liquid 564. Immersion-cooling system 500 may further include a filter 575 disposed adjacent to the tank 520.
[0447] Filter 575 may include a filtration media, a housing, and a pump configured to force immersion cooling liquid 564 through filter 575 to remove contaminants, particulates, or other impurities that may be added to immersion cooling liquid 564 during use. Filter 575 may be housed outside of tank 520 while being in fluidic communication with immersion cooling liquid 564 in tank 520. Alternatively, filter 575 may be submerged within immersion cooling liquid 564 inside of tank 520.
[0448] Immersion cooling liquid 564 may be a hydrocarbon, a fluoroketone, an oil, or a similar dielectric liquid that will act as an insulator while simultaneously transferring heat from package 505 more efficiently than air. Examples of immersion cooling liquid 564 are Novec™ 649, Novec™ 7000, and Novec™ 7100 produced by 3M™. An exemplary immersion cooling liquid 564 used in accordance with embodiments of the present invention may have a dielectric constant baseline value of about 1.8-2 at a frequency of about 1 kHz.
[0449] In an embodiment of the invention, immersion cooling liquid 564 may be considered unacceptably contaminated if the dielectric constant and / or dielectric loss tangent of immersion cooling fluid being used in an immersion cooling system 500 differs by a threshold amount as compared to unused or pure immersion cooling liquid 564. For example, immersion cooling liquid 564 may be considered unacceptably contaminated or degraded if the dielectric constant and / or dielectric loss tangent differs by a threshold of 10% or more as compared to unused or pure immersion cooling liquid 564. In an embodiment, a dielectric constant and / or dielectric loss tangent variation threshold may be 20%, 15%, 5%, 3%, 1%, or any suitable threshold.
[0450] Contamination of the immersion cooling liquid 564 and resulting changes to dielectric constant and / or dielectric loss tangent may alter or negatively impact operation of components within immersion cooling liquid 564 including semiconductor die(s) 550. An altered dielectric constant and / or dielectric loss tangent may result in undesirable cross-talk between components on a PCB, additional noise or reduction in signal strength transmitted along exposed wires of a PCB or semiconductor die(s) 550 submerged in immersion fluid, and / or signal dissipationAttorney Docket No. MTSH-152WO01through the immersion cooling liquid 564. Signal loss may be severe enough that two elements may be effectively represented as being separated by an open circuit despite being physically connected. In an embodiment, a dielectric constant and / or dielectric loss tangent variation threshold may be selected based on an observed or inferred effect on one or more submerged semiconductor die(s) 550. For example, an increase in PCIe bit error rate above an error rate baseline may be correlated with an increase in dielectric constant and / or dielectric loss tangent above a dielectric constant and / or dielectric loss tangent baseline. Accordingly, operation of semiconductor die(s) 550 may be throttled or suspended when a dielectric constant and / or dielectric loss tangent of immersion cooling liquid 564 exceeds a predetermined threshold.[0451 Changes to dielectric constant and / or dielectric loss tangent may be caused by contaminants within immersion cooling liquid 564. In some cases, changes to dielectric constant and / or dielectric loss tangent may be reversed by filtering the contaminants from immersion cooling liquid 564. In some embodiments, upon detecting an increase in dielectric constant and / or dielectric loss tangent of immersion cooling liquid 564, controller 502 may instruct filter 575 to increase filtration throughput or notify a user that an immersion cooling liquid 564 filtration media may need to be replaced. If a dielectric constant and / or dielectric loss tangent exceeds a predetermined threshold, controller 502 may throttle or shut down one or more semiconductor die(s) 550, generate a notification that immersion cooling liquid 564 should be replaced, trigger an alarm, etc.
[0452] Further examples of sensors and methods for immersion cooling contamination monitoring may include probes for monitoring immersion cooling liquid parameters such as dielectric constant and dielectric loss tangent, and processors configured to identify trends in sensor data, model immersion cooling system behavior as a function of contamination, and alter operations of immersion cooling systems based on detected levels and / or states of contamination may be found in PCT Application PCT / US24 / 36936, filed July 5, 2024 and entitled “FLUID SENSOR SYSTEMS ,” the entirety of which is incorporated herein by reference.
[0453] The illustrated example of FIG. 31 is not intended to be to scale. The immersion cooling system 500 may house and provide immersion cooling liquid 564 to tens, hundreds, or even thousands of packages 505. In some cases, the immersion cooling system 500 can be small (e.g., the size of a floor unit air conditioner, approximately 1 meter high, 0.5 meter width, 0.5 meter depth or length). In some implementations, the immersion cooling system can be largeAttorney Docket No. MTSH-152WO01(e.g., the size of a van or larger, approximately 2.5 meters high, 2.5 meters width, 4 meters depth or length).(0454] The immersion cooling system 500 can also include a controller 502 (e.g., a microcontroller, programmable logic controller (PLC), microprocessor, field-programmable gate array, logic circuitry, memory, or some combination thereof) to manage system operation. Controller 502 can perform various system functions such as monitoring temperatures of system components, cooling fluid level, tank access, chiller operation etc. The controller 502 can further issue commands to control system operation such as executing a start-up sequence, executing a shut-down sequence, assigning workloads among the packages, changing cooling fluid level, changing the temperature of the heat-transfer fluid circulated by the chiller 580, etc. In some implementations, controller 502 can include (or itself be) a baseboard management controller (BMC) 504. That is, the BMC 504 may monitor and control all aspects of system operation for the immersion cooling system 500 in addition to monitoring and controlling workloads of the semiconductor dies 550 in the packages 505 cooled by the system. The immersion cooling system 500 can also include a network interface controller (NIC 503) to allow the system to communicate over a network, such as a local area network or wide area network. The immersion cooling system 500 can further include a fluid sensor array 590 having a plurality of fluid sensors 510. Fluid sensors 510 may include one or more leak detection sensors at least partially submerged in immersion cooling liquid 564.(0455] The semiconductor die(s) 550 and can be mounted on and attached to a printed circuit board (PCB) 555 (sometimes referred to as a substrate) in device package 505. The package 505 can be made commercially available as an off-the-shelf (OTS) product. The package 505 can be used for single-phase or two-phase immersion cooling of at least one semiconductor die 550, such as a microprocessor (e.g., a central processing unit (CPU) and / or graphics processing unit (GPU)), voltage regulator (VR), high bandwidth memory (HBM), a digital signal processing (DSP) die, an artificial intelligence (Al) accelerator, an application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), and / or other densely patterned semiconductor die.(0456| In the two-phase immersion cooling system 500 of FIG. 31, heat flows from the semiconductor die 550 where it is generated into the heat spreader 552. The heat spreader 552 is in thermal contact with an immersion cooling liquid 564 that can flow over and extract heat from the heat spreader 552. The amount of heat delivered by the heat spreader 552 to the immersion cooling liquid 564 is enough to boil the immersion cooling liquid 564 that contactsAttorney Docket No. MTSH-152WO01the heat spreader 552 (creating bubbles 565 and potentially creating froth 567 when bubbles 565 reach the surface of immersion cooling liquid 564). The vapor 566 from the boiled immersion cooling liquid 564 can be cooled and condensed back to liquid droplets 568, for example, by the condenser tube 570. The heat-transfer fluid, such as chilled water, from the chiller 580 can be circulated through the condenser tube 570 to lower the temperature of the condenser tube 570 below the condensation point in the headspace 508 of the tank 520. As a result, vapor 566 condenses on exterior surfaces of the condenser tube 570 and liquid droplets 568 from the condensed vapor can drip and / or flow back to the immersion cooling liquid 564. There may be a plurality of condenser tubes 570 in tank 520 to condense the vapor 566 into droplets. Some or all of the condenser tubes 570 may or may not be located directly over the PCBs 557. Instead, the condenser tube(s) 570 can be located near one or more walls of the tank 520, such that the condenser tube(s) 570 are not directly over the PCBs 557 on which the packages 505 are mounted.[0457| To improve thermal performance in two-phase immersion cooling system 500, the heat spreader 552 can include a boiling enhancement coating (BEC) on at least one surface. The BEC can be formed from copper or a copper alloy and can be porous, for example, though BECs can take various forms. In some cases, the BEC is a micro porous copper coating having a thickness from approximately or exactly 50 microns to 500 microns thick (which may be produced by electroplating and / or etching). In some implementations, the BEC comprises a mesh copper layer bonded (e.g., via resistance heating) to at least an outer surface of the heat spreader 552. In some cases, the BEC is applied as particulates to at least one smooth surface of the heat spreader 552 and then subsequently sintered to adhere to one another and to the heat spreader 552. The BEC provides an improved surface area to contact the immersion cooling liquid 564 and can increase the heat transfer coefficient from the heat spreader 552 to the immersion cooling liquid 564 by up to a factor of 15 versus a smooth surface on the heat spreader 552. Accordingly, BECs can increase thermal conductivity to, and accelerate the boiling of, the immersion cooling liquid 564.
[0458] Further implementations of boiling enhancement coatings and enclosures are possible. Additional arrangements, applications, and methods of use of boiling enhancement coatings and enclosures, including with semiconductor dies and 3DIC stacks, are described in the below U.S. Patent Applications.
[0459] U.S. Patent Application No. 18 / 327,615, filed June 1, 2023, and entitled "Boiler Enhancement Coatings with Active Boiling Management,” discloses heat spreader and boilingAttorney Docket No. MTSH-152WO01enhancement enclosure architectures thermally and / or mechanically coupled to one or more semiconductor dies or logic ICs that may be used for passive and / or active management of immersion cooling fluid boiling, including through the use of valves to control pressure of boiling immersion cooling fluid within a boiling enhancement chamber, particularly in paragraphs
[0018] -
[0039] and FIGS. 3-5B. The entirety of U.S. Patent Application No.18 / 327,615 is incorporated herein by reference.
[0460] U.S. Patent Application No. 18 / 460,091, filed September 1, 2023, and entitled “Direct to Chip Application of Boiling Enhancement Coating,” discloses BECs and methods for applying BECs to semiconductor dies, logic ICs, and / or 3DIC stacks in accordance with the present technology. In particular, paragraphs
[0024] -
[0046] and FIGS. 2A-5 disclose embodiments of BEC layers, adhesives, solders, sintering, laser ablation, meshes, and other BECs and BEC application methods. The entirety of U.S. Patent Application No. 18 / 460,091 is incorporated herein by reference.
[0461] PCT Application No. PCT / US2025 / 032315, filed June 4, 2025, and entitled “Systems and Methods For Cooling Heat Generating Components in Immersion-Cooling Systems,” discloses structures that may be thermally and / or mechanically coupled to computing hardware such as one or more semiconductor dies, logic ICs, and / or 3DIC stacks to enable the shedding of immersion cooling vapors generated from the boiling of immersion cooling fluid during operation of the computing hardware. The entirety of PCT Application No. PCT / US2025 / 032315 is incorporated herein by reference.
[0462] U.S. Non-Provisional Application No. 19 / 502,220, filed January 14, 2026, and entitled “APPARATUSES AND METHODS FOR MODIFYING SURFACES OF MICROELECTRONIC DEVICES TO PROMOTE BOILING OF A COOLANT LIQUID,” discloses methods for creating boiling enhancement modifications to surfaces such as the surfaces of computing hardware such as one or more semiconductor dies, logic ICs, and / or 3DIC stacks. For example, grooves, patterns, gouges, trenches, or other structures may be added to a surface or lid of a processor, semiconductor die, logic IC, 3DIC stack component, and / or BEC to encourage nucleation sites for bubbles of immersion cooling vapor to form during a cooling process, thus decreasing the thermal resistance between the processor, semiconductor die, logic IC, and / or 3DIC stack component and the surrounding immersion cooling fluid. The entirety of U.S. Non-Provisional Application No. 19 / 502,220 is incorporated herein by reference.Attorney Docket No. MTSH-152WO01
[0463] U.S. Non-Provisional Patent Application No. 19 / 502,226, filed January 14, 2026, and entitled “CONNECTOR FOR COMMUNICATION HAVING IMPROVED SIGNAL INTEGRITY IN IMMERSION COOLING ENVIRONMENTS,” discloses heaters for promoting boiling of immersion cooling fluid near electrical connectors such as connections between components of a 3DIC stack and enable improved impedances at those connectors. The entirety of U.S. Non-Provisional Patent Application No. 19 / 502,226 is incorporated herein by reference.7. Conclusion
[0464] All parameters, dimensions, materials, and configurations described herein are meant to be exemplary and the actual parameters, dimensions, materials, and / or configurations will depend upon the specific application or applications for which the inventive teachings is / are used. It is to be understood that the foregoing embodiments are presented primarily by way of example and that, within the scope of the appended claims and equivalents thereto, inventive embodiments may be practiced otherwise than as specifically described and claimed. Inventive embodiments of the present disclosure are directed to each individual feature, system, article, material, kit, and / or method described herein.
[0465] In addition, any combination of two or more such features, systems, articles, materials, kits, and / or methods, if such features, systems, articles, materials, kits, and / or methods are not mutually inconsistent, is included within the inventive scope of the present disclosure. Other substitutions, modifications, changes, and omissions may be made in the design, operating conditions and arrangement of respective elements of the example implementations without departing from the scope of the present disclosure. The use of a numerical range does not preclude equivalents that fall outside the range that fulfill the same function, in the same way, to produce the same result.[0466| Also, various inventive concepts may be embodied as one or more methods, of which an example has been provided. The acts performed as part of the method may be ordered in any suitable way. Accordingly, embodiments may be constructed in which acts are performed in an order different than illustrated, which may include performing some acts simultaneously, even though shown as sequential acts in illustrative embodiments.[0467| All publications, patent applications, patents, and other references mentioned herein are incorporated by reference in their entirety.Attorney Docket No. MTSH-152WO01
[0468] All definitions, as defined and used herein, should be understood to control over dictionary definitions, definitions in documents incorporated by reference, and / or ordinary meanings of the defined terms.
[0469] The indefinite articles “a” and “an,” as used herein in the specification and in the claims, unless clearly indicated to the contrary, should be understood to mean “at least one.”
[0470] The phrase “and / or,” as used herein in the specification and in the claims, should be understood to mean “either or both” of the elements so conjoined, i.e., elements that are conjunctively present in some cases and disjunctively present in other cases. Multiple elements listed with “and / or” should be construed in the same fashion, i.e., “one or more” of the elements so conjoined. Other elements may optionally be present other than the elements specifically identified by the “and / or” clause, whether related or unrelated to those elements specifically identified. Thus, as a non-limiting example, a reference to “A and / or B”, when used in conjunction with open-ended language such as “comprising” can refer, in one embodiment, to A only (optionally including elements other than B); in another embodiment, to B only (optionally including elements other than A); in yet another embodiment, to both A and B (optionally including other elements); etc.
[0471] As used herein in the specification and in the claims, “or” should be understood to have the same meaning as “and / or” as defined above. For example, when separating items in a list, “or” or “and / or” shall be interpreted as being inclusive, i.e., the inclusion of at least one, but also including more than one, of a number or list of elements, and, optionally, additional unlisted items. Only terms clearly indicated to the contrary, such as “only one of’ or “exactly one of,” or, when used in the claims, “consisting of,” will refer to the inclusion of exactly one element of a number or list of elements. In general, the term “or” as used herein shall only be interpreted as indicating exclusive alternatives (i.e., “one or the other but not both”) when preceded by terms of exclusivity, such as “either,” “one of,” “only one of,” or “exactly one of.” “Consisting essentially of,” when used in the claims, shall have its ordinary meaning as used in the field of patent law.
[0472] As used herein in the specification and in the claims, the phrase “at least one,” in reference to a list of one or more elements, should be understood to mean at least one element selected from any one or more of the elements in the list of elements, but not necessarily including at least one of each and every element specifically listed within the list of elements and not excluding any combinations of elements in the list of elements. This definition alsoAttorney Docket No. MTSH-152WO01allows that elements may optionally be present other than the elements specifically identified within the list of elements to which the phrase “at least one” refers, whether related or unrelated to those elements specifically identified. Thus, as a non-limiting example, “at least one of A and B” (or, equivalently, “at least one of A or B,” or, equivalently “at least one of A and / or B”) can refer, in one embodiment, to at least one, optionally including more than one, A, with no B present (and optionally including elements other than B); in another embodiment, to at least one, optionally including more than one, B, with no A present (and optionally including elements other than A); in yet another embodiment, to at least one, optionally including more than one, A, and at least one, optionally including more than one, B (and optionally including other elements); etc.
[0473] In the claims, as well as in the specification above, all transitional phrases such as “comprising,” “including,” “carrying,” “having,” “containing,” “involving,” “holding,” “composed of,” and the like are to be understood to be open-ended, i.e., to mean including but not limited to. Only the transitional phrases “consisting of’ and “consisting essentially of’ shall be closed or semi-closed transitional phrases, respectively, as set forth in the United States Patent Office Manual of Patent Examining Procedures, Section 2111.03.
Claims
Attorney Docket No. MTSH-152WO01CLAIMS1. A system-on-a-wafer (SoW) comprising:a back end of the line (BEOL) comprising a plurality of interconnects, the plurality of interconnects collectively providing a plurality of interconnection pathways; anda computing component disposed on the BEOL, the computing component comprising:a substrate;a functional circuit block disposed on the substrate; anda programmable logic circuit disposed on the substrate and electrically coupled to the plurality of interconnects,wherein:the programmable logic circuit is configured to redirect transmission of a signal from a first interconnection pathway of the plurality of interconnection pathways to a second interconnection pathway of the plurality of interconnection pathways; andthe first interconnection pathway and the second interconnection pathway are communicatively coupled to the functional circuit block via the programmable logic circuit.
2. The SoW of claim 1, wherein:the computing component is a first computing component; andthe SoW further comprises:a second computing component communicatively coupled to the first interconnection pathway; anda third computing component communicatively coupled to the second interconnection pathway.
3. The SoW of claim 1, wherein the programmable logic circuit is configured to, when redirecting transmission of the signal from the first interconnection pathway to the second interconnection pathway:activate the second interconnection pathway; anddeactivate the first interconnection pathway.
4. The SoW of claim 1, wherein the programmable logic circuit comprises a transistor.Attorney Docket No. MTSH-152WO015. The SoW of claim 1, wherein the programmable logic circuit comprises a gate array.
6. The SoW of claim 1, wherein the programmable logic circuit comprises a field programmable gate array (FPGA).
7. The SoW of claim 1, wherein the programmable logic circuit comprises a microelectromechanical systems (MEMS) switch.
8. The SoW of claim 1, wherein:the computing component is a first computing component;the substrate is a first substrate;the functional circuit block is a first functional circuit block;the programmable logic circuit is a first programmable logic circuit;the SoW further comprises:a second computing component, comprising:a second substrate;a second functional circuit block disposed on the second substrate; and a second programmable logic circuit disposed on the second substrate; the second programmable logic circuit is electrically coupled to the plurality of interconnects;the second programmable logic circuit is configured to redirect transmission of a signal from a third interconnection pathway of the plurality of interconnection pathways to a fourth interconnection pathway of the plurality of interconnection pathways; andthe third interconnection pathway and the fourth interconnection pathway are communicatively coupled to the second computing component.
9. The SoW of claim 8, wherein the first computing component is different from the second computing component.
10. The SoW of claim 9, wherein:the first programmable logic circuit is communicatively coupled to a first subset of interconnection pathways of the plurality of interconnection pathways;Attorney Docket No. MTSH-152WO01the second programmable logic circuit is communicatively coupled to a second subset of interconnection pathways of the plurality of interconnection pathways; andthe first subset of interconnection pathways and the second subset of interconnection pathways include a different number of interconnection pathways.
11. The SoW of claim 1, wherein:the computing component is a first computing component;the substrate is a first substrate;the functional circuit block is a first functional circuit block;the SoW further comprises:a second computing component, comprising:a second substrate; anda second functional circuit block disposed on the second substrate; and the second computing component does not include a programmable logic circuit to redirect a signal between different interconnection pathways of the plurality of interconnection pathways.
12. The SoW of claim 1, wherein:the computing component is communicatively coupled to a subset of interconnection pathways of the plurality of interconnection pathways; andthe programmable logic circuit is directly connected to all interconnection pathways of the subset of interconnection pathways.
13. The SoW of claim 1, wherein:the computing component is communicatively coupled to a subset of interconnection pathways of the plurality of interconnection pathways;the programmable logic circuit is directly connected to one or more interconnection pathways of the subset of interconnection pathways; andthe programmable logic circuit is not directly connected to all interconnection pathways of the subset of interconnection pathways.
14. The SoW of claim 1, wherein:the computing component is communicatively coupled to a subset of interconnection pathways of the plurality of interconnection pathways;Attorney Docket No. MTSH-152WO01one or more first interconnection pathways of the subset of interconnection pathways is directly connected to the programmable logic circuit; andone or more second interconnection pathways of the subset of interconnection pathways is directly connected to the functional circuit block.
15. The SoW of claim 1, wherein the BEOL is shaped as a semiconductor wafer.
16. The SoW of claim 1, wherein the computing component comprises at least one of a graphics processing unit (GPU);a central processing unit (CPU);a memory module;a switch;an input / output (VO) controller; ora network interface controller (NIC).
17. The SoW of claim 1, wherein the SoW is configured to operate as an artificial intelligence (Al) accelerator.
18. The SoW of claim 1, wherein the plurality of interconnects comprises at least one interconnect providing only one interconnection pathway of the plurality of interconnection pathways.
19. The SoW of claim 1, wherein the plurality of interconnects comprises at least one interconnect providing two or more interconnection pathways of the plurality of interconnection pathways.
20. A system-on-a-wafer (SoW), comprising:a back end of the line (BEOL), comprising a plurality of interconnects;a plurality of computing components, disposed on the BEOL, each computing component of the plurality of computing components comprising:a substrate; anda circuit disposed on the substrate; andat least one programmable interconnect communicatively coupled to the plurality of interconnects,wherein:Attorney Docket No. MTSH-152WO01the plurality of interconnects and the at least one programmable interconnect provide a plurality of interconnection pathways communicatively coupled to respective computing components of the plurality of computing components; andthe at least one programmable interconnect is configured to change at least one interconnection pathway of the plurality of interconnection pathways.
21. The SoW of claim 20, wherein the BEOL is shaped as a semiconductor wafer.
22. The SoW of claim 20, wherein the plurality of computing components comprises: at least one first computing component; andat least one second computing component different from the at least one first computing component.
23. The SoW of claim 20, wherein the plurality of computing components comprises at least one of:a graphics processing unit (GPU);a central processing unit (CPU);a memory module;a switch;an input / output (VO) controller; ora network interface controller (NIC).
24. The SoW of claim 20, wherein the at least one programmable interconnect comprises a transistor.
25. The SoW of claim 20, wherein the at least one programmable interconnect comprises a gate array.
26. The SoW of claim 20, wherein the at least one programmable interconnect comprises a field programmable gate array (FPGA).
27. The SoW of claim 20, wherein the at least one programmable interconnect comprises a micro-electromechanical systems (MEMS) switch.Attorney Docket No. MTSH-152WO0128. The SoW of claim 20, wherein the plurality of computing components comprises: a first computing component containing the at least one programmable interconnect.
29. The SoW of claim 20, wherein the plurality of interconnects comprises:a first plurality of interconnects disposed in a first redistribution layer of the BEOL and communicatively coupled to the plurality of computing components;a second plurality of interconnects disposed in a bridge layer of the BEOL and communicatively coupled to the first plurality of interconnects; anda third plurality of interconnects disposed in a second redistribution layer of the BEOL and communicatively coupled to the second plurality of interconnects.
30. The SoW of claim 29, wherein the at least one programmable interconnect is disposed in the bridge layer of the BEOL.
31. The SoW of claim 30, further comprising:at least one local silicon interconnect (LSI) disposed in the bridge layer of the BEOL.
32. The SoW of claim 29, further comprising:a plurality of solder balls disposed in a bump layer of the BEOL, wherein the plurality of solder balls is communicatively coupled to the third plurality of interconnects.
33. The SoW of claim 32, wherein the at least one programmable interconnect is disposed in the bump layer of the BEOL.
34. The SoW of claim 20, wherein the plurality of computing components is configured to operate as an artificial intelligence (Al) accelerator.
35. A system-on-a- wafer (SoW), comprising:a computing component layer, the computing component layer comprising a plurality of computing components, each computing component of the plurality of computing components comprising:a substrate; anda circuit integrally formed on the substrate;a back end of the line (BEOL), comprising:Attorney Docket No. MTSH-152WO01a first redistribution layer disposed on the computing component layer, the first redistribution layer comprising a first plurality of interconnects communicatively coupled to the plurality of computing components;a bridge layer disposed on the first redistribution layer, the bridge layer comprising a second plurality of interconnects communicatively coupled to the first plurality of interconnects;a second redistribution layer disposed on the bridge layer, the second redistribution layer comprising a third plurality of interconnects communicatively coupled to the second plurality of interconnects; anda bump layer disposed on the second redistribution layer, the bump layer comprising a plurality of solder balls communicatively coupled to the third plurality of interconnects; andat least one programmable interconnect disposed in at least one of the computing component layer, the bridge layer, or the bump layer,wherein the BEOL is shaped as a semiconductor wafer.
36. The SoW of claim 35, wherein the bridge layer further comprises at least one local silicon interconnect (LSI).
37. A computing system, comprising:one or more system-on-a-wafer (SoW) modules, each SoW module of the one or more SoW modules comprising:a back end of line (BEOL); anda plurality of computing components electrically coupled to the BEOL, the plurality of computing components consisting of:at least one of an input / output (I / O) computing component or a network interface controller (NIC); anda plurality of functional computing components, each functional computing component of the plurality of functional computing components being functionally identical to every other functional computing component of the plurality of functional computing components.Attorney Docket No. MTSH-152WO0138. The computing system of claim 37, wherein the one or more SoW modules comprises a SoW module, the SoW module comprising a plurality of central processing units (CPUs) as the plurality of functional computing components of the SoW module.
39. The computing system of claim 37, wherein the one or more SoW modules comprises a SoW module, the SoW module comprising a plurality of graphics processing units (GPUs) as the plurality of functional computing components of the SoW module.
40. The computing system of claim 37, wherein the one or more SoW modules comprises a SoW module, the SoW module comprising a plurality of switches as the plurality of functional computing components of the SoW module.
41. The computing system of claim 37, wherein the one or more SoW modules comprises a SoW module, the SoW module comprising a plurality of memory modules as the plurality of functional computing components of the SoW module.
42. The computing system of claim 37, wherein the one or more SoW modules comprises:a first SoW module comprising a plurality of central processing units (CPUs) as the plurality of functional computing components of the first SoW module;a second SoW module comprising a plurality of graphics processing units (GPUs) as the plurality of functional computing components of the second SoW module;a third SoW module comprising a plurality of switches as the plurality of functional computing components of the third SoW module; anda fourth SoW module comprising a plurality of memory modules as the plurality of functional computing components of the fourth SoW module.
43. The computing system of claim 37, wherein each SoW module of the one or more SoW modules is communicatively coupled to every other SoW module of the one or more SoW modules.
44. The computing system of claim 37, wherein each SoW module of the one or more SoW modules is communicatively coupled to up to four other SoW modules of the one or more SoW modules.Attorney Docket No. MTSH-152WO0145. The computing system of claim 37, further comprising:a plurality of cables, wherein each cable of the plurality of cables is communicatively coupled to a pair of SoW modules of the one or more SoW modules and is configured to transmit one or more signals between the pair of SoW modules.
46. The computing system of claim 45, wherein each SoW module of the one or more SoW modules further comprises:a connector having a first end and a second end;wherein:the first end is communicatively coupled to the at least one of the I / O computing component or the NIC; andthe second end is communicatively coupled to one cable of the plurality of cables.
47. The computing system of claim 37, further comprising:a printed circuit board (PCB) to communicatively couple each SoW module of the one or more SoW modules to every other SoW module of the one or more SoW modules.
48. The computing system of claim 37, wherein the one or more SoW modules comprises:a first SoW module;a second SoW module communicatively coupled to the first SoW module via a first connection standard; anda third SoW module communicatively coupled to the first SoW module via a second connection standard different from the first connection standard.
49. The computing system of claim 37, wherein the at least one of the I / O computing component or the NIC supports at least one of a Universal Chiplet Interconnect Express (UCIe) connection, a Peripheral Component Express (PCIe) connection, an Ultra Accelerator Link (UALink) connection, an Ethernet connection, an Ultra Ethernet connection, an InfiniBand connection, or an optical input / output (I / O) connection.
50. A computing system, comprising:Attorney Docket No. MTSH-152WO01one or more system-on-a-wafer (SoW) modules, each SoW module of the one or more SoW modules comprising:a back end of line (BEOL); anda plurality of computing components electrically coupled to the BEOL, the plurality of computing components comprising:at least one of an input / output (I / O) computing component or a network interface controller (NIC); anda plurality of functional computing components predominantly comprising a first type of functional computing component.
51. The computing system of claim 50, wherein a percentage of the plurality of functional computing components corresponding to the first type of functional computing component is greater than or equal to 90%.
52. The computing system of claim 50, wherein a quantity of the plurality of functional computing components corresponding to one or more other type of functional computing components different from the first type of functional computing components is less than or equal to 150.
53. A computing system, the computing system comprising:a substrate comprising a redistribution layer (RDL); anda first unit block disposed on a first surface of the substrate, wherein the first unit block comprises:a first system-on-a-chip (SOC) and a second SOC;a first memory module disposed on the first SOC;a second memory module disposed on the second SOC;a third memory module disposed on the substrate, the third memory module being adjacent to the first SOC; anda fourth memory module disposed on the substrate, the fourth memory module being adjacent to the second SOC;wherein the first SOC is communicatively coupled to the fourth memory module through at least one functional block; andwherein the first SOC and the second SOC each comprises at least one logic integrated circuit (IC).Attorney Docket No. MTSH-152WO0154. The computing system of claim 53, further comprising a second functional block disposed in the second SOC, wherein the second memory module is communicatively coupled to the third memory module through the second functional block.
55. The computing system of claim 53, wherein the third memory module and the fourth memory module are communicatively coupled through the RDL.
56. The computing system of claim 53, further comprising a third functional block disposed in the first memory module and a fourth functional block disposed in the second memory module, wherein the first memory module is communicatively coupled to the second memory module through the third functional block and the fourth functional block.
57. The computing system of claim 56, further comprising a fifth functional block disposed in the first SOC and a sixth functional block disposed in the second SOC, wherein the first memory module is further communicatively coupled to the second memory module through the fifth functional block and the sixth functional block.
58. The computing system of claim 53, further comprising a voltage regulation module (VRM) configured to provide regulated voltage to the first unit block.
59. The computing system of claim 58, wherein the VRM is disposed on a second surface of the substrate opposite the first surface of the substrate.
60. The computing system of claim 59, wherein the VRM is electrically coupled to the first unit block through the RDL.
61. The computing system of claim 53, further comprising a second unit block communicatively coupled to the first unit block through the RDL.
62. The computing system of claim 53, further comprising at least 50 logic ICs disposed on the substrate.
63. The computing system of claim 62, further comprising at least 100 logic ICs disposed on the substrate.Attorney Docket No. MTSH-152WO0164. The computing system of claim 63, further comprising at least 200 logic ICs disposed on the substrate.
65. The computing system of claim 53, further comprising:at least one input / output (I / O) chip communicatively coupled to the first unit block through the RDL; andat least one wafer connector communicatively coupled to the first unit block through the RDL.
66. A computing system, the computing system comprising:a substrate comprising a redistribution layer (RDL);a plurality of three-dimensional integrated circuit (3DIC) stacks disposed on a first surface of the substrate, wherein each 3DIC stack of the plurality of 3DIC stacks comprises:a logic IC;at least one memory module disposed on the logic IC; anda plurality of functional blocks communicatively coupling each 3DIC stack with a component; anda plurality of wafer connectors communicatively coupled to the RDL, wherein the plurality of wafer connectors is configured to transmit data between each 3DIC stacks and an external logic IC disposed externally to the computing system;wherein the plurality of 3DIC stacks are disposed on the first surface in a first pattern.
67. The computing system of claim 66, wherein the plurality of 3DIC stacks comprises at least 503DIC stacks.
68. The computing system of claim 67, wherein the plurality of 3DIC stacks comprises at least 1003DIC stacks.
69. The computing system of claim 68, wherein the plurality of 3DIC stacks comprises at least 2003DIC stacks.
70. The computing system of claim 66, wherein the plurality of 3DIC stacks is arranged in a corresponding plurality of unit blocks, each unit block communicatively coupled through the RDL to respective other unit blocks of the plurality of unit blocks.Attorney Docket No. MTSH-152WO0171. The computing system of claim 70, wherein:the plurality of unit blocks comprises a first unit block and a second unit block; the first unit block is communicatively coupled to the second unit block through: a first functional block disposed in a first logic IC; anda second functional block disposed in a second logic IC; andthe first unit block and the second unit block are communicatively coupled to an external logic IC through at least one wafer connector of the plurality of wafer connectors.
72. The computing system of claim 70, wherein:each unit block of the plurality of unit blocks further comprises a plurality of memory modules disposed on the first surface of the substrate; andeach memory module of the plurality of memory modules is communicatively coupled to a respective logic IC of the respective unit block through a second plurality of functional blocks.
73. A computing system, the computing system comprising:a substrate;a first component disposed on the substrate, wherein the first component comprises:first circuitry configured to perform a first function; anda first functional block having a first plurality of communications paths;a second component disposed on the substrate, wherein the second component comprises:second circuitry configured to perform a second function; anda second functional block having a second plurality of communications paths; and a universal bridge communicatively coupling the first component to the second component;wherein:the first functional block and the second functional block each comprises respective surfaces having respective exposed leads arranged in a primary pattern; and the universal bridge has a plurality of surfaces having exposed universal bridge leads arranged to interface with the primary pattern.
74. The computing system of claim 73, wherein the first component and the second component comprise at least two different components selected from a graphics processingAttorney Docket No. MTSH-152WO01unit (GPU), a central processing unit (CPU), a memory module, an input / output (I / O) chip, a link chiplet, a network interface card (NIC), a tensor processing unit (TPU), a data processing unit (DPU), a switch, or a system-on-a-chip (SOC).
75. The computing system of claim 73, wherein the exposed universal bridge leads are symmetric about a plane parallel to the plurality of surfaces.
76. The computing system of claim 75, wherein the exposed universal bridge leads are glide symmetric about a plane parallel to the plurality of surfaces.
77. The computing system of claim 73, wherein the computing system is a component of an immersion cooling system, and the component is immersed in an immersion cooling liquid.
78. The computing system of claim 73, wherein the exposed universal bridge leads are arranged in the primary pattern.
79. The computing system of claim 73, wherein the exposed universal bridge leads are arranged in a second pattern that mirrors the primary pattern.
80. The computing system of claim 73, wherein the universal bridge comprises an universal bridge substrate and a plurality of traces.
81. The computing system of claim 80, wherein the universal bridge substrate comprises at least one of silicon, silicon nitride, silicon dioxide, or printed circuit board (PCB).
82. The computing system of claim 80, wherein the plurality of traces comprise gold, copper, or aluminum.
83. The computing system of claim 80, wherein the universal bridge comprises a first integrated circuit (IC) configured to control a routing of communications through the plurality of traces.
84. The computing system of claim 83, wherein the first IC comprises a field-programmable gate array (FPGA).
85. The computing system of claim 73, wherein:Attorney Docket No. MTSH-152WO01the first component is not configured to perform the second function; andthe second component is not configured to perform the first function.
86. The computing system of claim 73, wherein:the first component is configured to perform the second function; andthe second component is configured to perform the first function.
87. The computing system of claim 73, wherein the universal bridge is disposed above the substrate.
88. The computing system of claim 73, wherein the universal bridge is disposed in the substrate.
89. The computing system of claim 88, wherein the universal bridge is disposed in at least one of a redistribution layer (RDL) or a local silicon interconnect (LSI) layer.
