Direct bond native interconnect and active base die

By directly bonding native interconnects and active base dies, the problems of low efficiency and high power of standard interfaces in microelectronics systems are solved, efficient inter-chip communication and adaptability are achieved, and cost and heat generation are reduced.

CN119028958APending Publication Date: 2024-11-26AIDEA SEMICONDUCTOR CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202411177513.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2016-10-07
Filing Date
2017-10-04
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

In existing microelectronics systems, the complexity and overhead of standard interfaces lead to low communication efficiency between chips, high power consumption, and difficulty in adapting to differences in different process nodes and operating voltages.

Method used

Using direct bonded native interconnects and active base dies, short data paths are formed through direct bonding of metal to metal, abandoning standard interfaces, and using customized logic and state components of active base dies for signal driving and voltage Adjust to adapt to multiple interconnection types and process nodes.

Benefits of technology

It significantly reduces size and power requirements, improves signal quality and timing, reduces heat generation and operating costs, adapts to the operating voltage of different chiplets, and realizes efficient data transmission and interconnection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119028958A_ABST
    Figure CN119028958A_ABST
Patent Text Reader

Abstract

The invention provides a directly bonded native interconnect and an active base die. In a microelectronic architecture, an active die or chiplet is connected to an active base die via its core-level conductors. These native interconnects provide a short data path, which abandons the overhead of a standard interface. Because the native interconnect is coupled to a bit, the system preserves a redistributed route. The base die may include custom logic, allowing the attached die to provide a storage function. The architecture can connect different interconnect types and chiplets from various process nodes operating at different voltages. The base die may have a state element for driving. Functional blocks on the base die receive native signals from different chiplets and communicate with all attached chiplets. The chiplet may share processing resources and memory resources of the base die. And the routing blockage is very small, so that the signal quality and the timing are improved. The system is capable of operating at a double data rate or a quadruple data rate. The architecture facilitates ASIC, ASSP, and FPGA ICs as well as neural networks, thereby reducing footprint and power requirements.
Need to check novelty before this filing date? Find Prior Art

Description

[0001] This application is a divisional application of an invention patent application with an international application date of October 4, 2017 (the earliest priority date is October 7, 2016), which entered the Chinese national phase on June 4, 2019, and has a Chinese national application number of 201780075216.6 and is named “Directly Bonded Native Interconnects and Active Base Dies”. Technical Field

[0002] The present invention relates generally to electronic circuits and more particularly to voltage sources. Background Art

[0003] In microelectronic systems, electronic circuits are made on wafers or semiconductor materials (such as silicon). The wafer with the electronic circuits may be bonded to one or more other wafers, bonded to a separate die, or itself cut into multiple dies, each containing a replica of the circuit. Each die with a functional integrated circuit is called a microchip or "chip". When specific functions from a functional library are assigned to separate chips, or when a large monolithic chip is simulated by a collection of smaller chips, these smaller chips or chips with specific or proprietary functions may be called "chiplets". As used herein, a chiplet generally refers to a complete subsystem IP core (intellectual property core) on a single die, a reusable logic unit. A library of chiplets can be used to provide conventional or complete IP block functions.

[0004] Conventionally, microchips and chiplets require standard interfaces to communicate and interact with each other and with the larger microelectronic layout that constitutes a microelectronic device. The use of such standard interfaces is expected and taken for granted in the industry. It is assumed in the industry that every logic block that requires input and output (I / O) will work through a standard interface that includes at least some I / O protocols. A standard interface can be formally defined as:

[0005] "A point of interconnection between two systems or between parts of a system (e.g., a processor and a peripheral) at which all physical, electrical, and logical parameters conform to predetermined values ​​and are otherwise common. Interfaces may be classified as standard based on manufacturer, industry, or international use. A processor's I / O channels may be classified as standard because they are common to all processors of that type, or common to more than one type of peripheral, but they may be manufacturer-specific. Some interfaces are de facto industry standards and can be used to connect devices from different vendors. Other interfaces are standardized through agreements within industry associations or international committees or consortia" (Dictionary of Computing 2004, first published by Oxford University Press in 2004).

[0006] Standard interfaces and I / O protocols provide well-characterized outputs with drivers large enough to power a variety of output loads and provide other benefits such as voltage leveling and buffered inputs with electrostatic discharge (ESD) protection. The tradeoff for these benefits is that the native signals generated by the specific logic or "core IP" of a given microchip must be adapted, modified, and often routed to have suitable compatibility with the standard interfaces. Standard interfaces, in turn, enable multiple independent chips to "talk" to each other in a standardized manner according to standardized protocols because these interfaces have standard pin geometry, designed serialization, standard voltages, standard timing, etc. to achieve common compatibility. But chiplets and the resulting 3D stacked IC structures tend to be larger, more complex, more expensive, generate more heat, and consume more power than they need to support the on-board standard interfaces and I / O protocols. Summary of the invention

[0007] The present invention provides a direct bonded native interconnect and active base die. A native interconnect is a metal-to-metal bond formed directly between a native conductor of a die and a conductor of a second die, thereby forgoing the need for the complexity and expense of a standard interface. The native conductor of a die is an electrical conductor that has electrical contact to the original or native signal of the die, operating at the core functional logic level of a particular die without requiring significant modification of the signal in order to connect to other dies.

[0008] In microelectronic architecture, the active die or chiplet is connected to the active base die via its core-level conductors. These native interconnects provide short data paths, which forgo the overhead of standard interfaces. Since the native interconnects are coupled to the bits, the system can save redistribution routing. The active base die can contain custom logic, allowing the attached die to provide storage functions.

[0009] The active base die can accommodate a variety of interconnect types and can accommodate chiplets from various process nodes and different operating voltages. The active base die can be driven by its own state elements or can be driven by state elements on attached chiplets at cross-die boundaries. The active base die receives native core-side signals from multiple different chiplets and is capable of bidirectional communication between the functional elements of the active base die and the attached chiplets. The active base die can significantly reduce the size and area footprint, and can reduce power requirements, especially for large hard chiplets. When necessary, the active base die can integrate repeater units for longer routes, and use data transmission schemes to improve signal quality, improve timing and provide native high-speed interfaces. The chiplets can share the processing resources and memory resources of the base die. Routing congestion is small because certain circuit elements on the chiplets can be oriented and / or aligned with circuit elements on the base die, thereby improving signal quality and timing. The system can optionally operate at double data rate (DDR) or quad data rate (QDR). The architecture benefits ASIC, ASSP and FPGA integrated circuits as well as large neural networks while reducing footprint and power requirements.

[0010] This Summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used as an aid in limiting the scope of the claimed subject matter. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Hereinafter, certain embodiments of the present disclosure will be described with reference to the accompanying drawings, wherein the same reference numerals represent the same elements. However, it should be understood that the accompanying drawings illustrate various specific implementations described herein and are not intended to limit the scope of the various technologies described herein.

[0012] Figure 1 is a diagram of an example of a conventional standard interface on each of the four sides of a conventional microchip.

[0013] Figure 2 is a diagram of a conventional monolithic integrated circuit layout with various functional blocks.

[0014] Figure 3 is a diagram showing an example wafer-to-wafer (W2W) fabrication of direct bonds between native conductors of a die on a first wafer and conductors of an active base die on a second wafer to form a native interface via a W2W bonding process.

[0015] Figure 4 Figures are provided that illustrate various example configurations of microelectronic devices incorporating native interconnects and active base dies.

[0016] Figure 5is a diagram of an example active base die included within an example microelectronic device.

[0017] Figure 6 is a diagram of an example core IP unit of an example chiplet.

[0018] Figure 7 is a diagram of an example active base die with voltage regulation domains.

[0019] Figure 8 is a diagram of an example active base die with a one-to-one voltage regulator.

[0020] Fig. 9 is a diagram of an example active base die including a clock for timing and synchronizing process steps and data transfers.

[0021] Fig.10 is a diagram of an example active base die with a negotiation engine or an out-of-order engine.

[0022] Fig.11 is a diagram of an example neural network implementation using an example active base die.

[0023] Fig.12 is a flow chart of an example method of making a microelectronic device including a native interconnect.

[0024] Fig.13 is a flow chart of an example method 1300 of providing a microchip architecture for semiconductor chiplets, wherein native core-side conductors of a plurality of chiplets are connected to an active base die.

[0025] Fig.14 is a flow chart of an example method 1400 of providing a microchip architecture for semiconductor chiplets, wherein voltages are adjusted to accommodate different chiplets.

[0026] Fig.15 is a flow chart of an example method for providing a microchip architecture for semiconductor chiplets using state elements in connected chiplets for signal driving in an active base die. DETAILED DESCRIPTION

[0027] Overview

[0028] The present disclosure describes example direct bonded native interconnects and active base dies. Example microelectronic devices have dies that are directly bonded to core-side conductors of one or more other dies, thereby providing a "native interconnect" that, in a specific implementation, can provide a unique interface between dies. Native interconnects can enable electronic circuits to span across different dies and across die boundaries between multiple different dies, but there is no standard interface and no input / output protocol across the cross-die boundaries that the direct bond connections to the native core-side conductors traverse.

[0029] As used herein, "standard interface" conforms to the dictionary definition given in the Background section above, and more briefly means "additional hardware, software, routing, logic, connections, or surface area added to the core logic entity or functionality of the die in order to meet industry or consortium specifications for interfacing, connecting, or communicating with other components or signals external to the die." As used herein, "direct bonding" refers to direct metal-to-metal contact, oxide bonding, or fusion bonding between two metals, such as a copper-to-copper (Cu-Cu) metal bond between two copper conductors in direct contact, which has at least partial lattice cohesion. Such direct bonding can be provided by hybrid bonding techniques, such as those described below. (Direct Bond Interconnect) technology and other metal bonding technologies (Xperi Group, Inc., San Jose, California, InvensasBonding Technologies). As used herein, "core" and "core side" refer to the location, signal and / or level that appears at the functional logic of a particular die, rather than the location, signal and / or level of the added standard interface defined by the alliance. Therefore, if the signal is operable at the core functional logic level of a particular die, the signal is original or "native" without certain modifications, such as additional serialization, added ESD protection, unless inherently provided by a specific circuit; with a non-serialized data path, it can pass through the die by simple latching, flipping or line coupling, without imposed input / output (I / O) protocols, etc. However, the native signal may undergo level shifting or voltage adjustment for the purpose of adaptation between dies of heterogeneous foundry sources, and is still a native signal used herein. As used herein, "active" (active base die) conforms to the usual meaning of active in the semiconductor field, rather than "passive". Active components include transistor logic and amplification elements, such as transistors. Passive components, on the other hand, introduce no net energy into the circuit and use no raw power, except that derived from other circuits connected to the passive circuit. Although the techniques shown herein generally refer to active dies, the techniques can also be applied to passive devices with the same or similar benefits.

[0030] A die's "native conductors" are electrical conductors that make electrical contact to the die's original or native signals, which, as described above, are signals that are operable at the core functional logic level of a particular die without requiring significant modification of the signal in order to connect to other dies.

[0031] Native interconnects for conducting such native signals from the core side of the die provide continuous circuits through two or more cross-die boundaries without amplifying or modifying the native signals, unless required to accommodate dies from different manufacturing processes. From a signal perspective, native signals from the IP core of one die are directly passed to other dies via direct-bonded native interconnects with no or negligible modification to the native signals, thereby abandoning standard interfaces and alliance-imposed input / output protocols.

[0032] Notably, the use of native interconnects made between different dies from heterogeneous foundry nodes or dies with incompatible manufacturing can enable such uninterrupted circuits to be advanced across or across die boundaries without interfaces and without input / output protocols. Thus, an example circuit can be advanced across a die boundary between a first die manufactured at a first foundry node that is directly bonded to a second die at a second foundry node with no other docking, or with only minimal latching or level shifting, for example, to balance voltages between dies. In one specific implementation, circuits disposed between multiple dies via directly bonded native interconnects can be advanced between an active base die and a proprietary chiplet die, or between dies (including an active base die) on each side of a wafer-to-wafer (W2W) process forming direct bonds, where at least some of the W2W direct bonds involve native conductors of the die on at least one side of the W2W bond.

[0033] In one implementation, microelectronic systems utilizing semiconductor chiplets can replicate various architectures, such as ASICs, ASSPs, and FPGAs, in a smaller, faster, and more power-efficient manner. As described above, a chiplet is a complete subsystem IP core (intellectual property core), such as a reusable logic unit on a single die.

[0034] Native interconnects may be fabricated during die-to-die or die-to-wafer direct bonding, which creates native interconnects between a first die (such as an active die or chiplet) and a second die (which may be an active base die). Native interconnects may also be made by direct bonding, for example, between an active base die on one wafer and other active die layers on other wafers during a wafer-to-wafer (W2W) process. One or more dies may be implemented in semiconductor material, although other materials, such as glass, oxide, or polymer may also be implemented as appropriate for a given design.

[0035] Figure 1An example comparison between a conventional microelectronic package 100 having a plurality of conventional chiplets 102 on a conventional interposer 104 and an example microelectronic package 106 presented on an active base die 108 is shown, as described below. The example microelectronic package 106 provides a much smaller physical package and significant improvements over the conventional package 100. The conventional chiplets 102 include, for example, a CPU core 110 surrounded by a conventional standard interface 112. The less improved package 106 includes the same CPU core 110, for example, attached directly to the active base die 108 in the absence of the conventional standard interface 112. The less improved package 106 is not only smaller, but is also more efficient, easier to manufacture, and has lower power requirements than its conventional counterpart package 100, and provides additional benefits thereto.

[0036] exist Figure 1 In the example, a conventional standard interface 112 may be located on each of the four sides of a conventional microchip or chiplet 102, such as a central processing unit (CPU) core 110. The standard interface 112 introduces a loss. Figure 1 It is apparent that the standard interface 112 significantly increases the area footprint of the example CPU core. If the CPU core 110 has dimensions of 3×5 mm, the CPU core 110 as a chiplet 110 with a standard interface 112 may have dimensions of 4×6 mm. Sometimes, the inclusion of the standard interface 112 effectively doubles the area footprint of a given chiplet. The standard interface 112 also draws significant additional power over the native logic of the CPU core 110 itself. For example, line drivers that are required in the standard interface 112 must be able to drive a large number of unknown output loads that may be connected, depending on unknown future uses. Because the standard interface 112 must be able to universally accommodate a large number of unknown output loads, conventional standard interfaces 112 typically have "overkill" of driver capacity and other capabilities that must be powered, but may be unnecessary for actual utilization of the chip.

[0037] The standard interface 112 also requires significant additional routing from the native interconnects of the core IP to the standard interface 112 so that the native signal first arrives at the standard interface 112. Therefore, the data path is longer and inherently less reliable, and congestion often occurs at the corner geometry of large chip layouts. In order to meet compatibility with the standard interface 112, the native signal is usually buffered, processed and doped by additional components such as inverters, repeaters, drivers, state machines, timers and regulators (for the standard interface 112, these components are added to the tube core). Due to the relatively large traditional pad size and line pitch of the standard interface, some conventional schemes add further complexity by multiplexing or serializing highly parallelized native signals via SerDes blocks or other interfaces, just to be able to leave the signal via a limited number of pins under the conventional large pitch constraints between tube cores. Therefore, in addition to increasing power requirements and requiring additional layout area, the standard interface 112 may be a troublesome bottleneck for I / O itself.

[0038] Figure 2 A conventional monolithic integrated circuit layout 200 is shown with various functional blocks 110, 202, 204, 206, ... n, relative to the conventional monolithic integrated circuit layout 200 shown in the present invention. Figure 2 1 and 2. The example microelectronic package 106 having the same functional blocks 110, 202, 204, 206 coupled to the active base die 108 as described in another part of FIG. Functional blocks or just "blocks" may consist of interfaces and specific implementations. Example blocks include multipliers, arithmetic logic units (ALUs), instruction decoders, digital signal processors (DSPs), etc.

[0039] Functional blocks 202 have been incorporated into active base die 108. In a two-dimensional (2D) floorplan of a conventional monolithic IC 200, it is apparent that some of blocks 204, 206 must have data paths 208 routed around or below an intermediate block in order to communicate with each other or with a third block 202. Conventional very large scale integration (VLSI) designs are typically significantly congested due to large hard IP blocks on the chip. For large processors, most on-chip signaling must bypass a large central cortex, resulting in high traffic density bypassing larger blocks. In many floorplans, the shortest route between two blocks may be a long distance around an intermediate block. These relatively long distances may also introduce the need for repeated instances of components (such as additional buffers, inverters, voltage regulators, repeaters, drivers, etc.), not to mention the additional routing itself, as circuit components become more distant from each other due to the layout of the floorplan.

[0040] The example microelectronic package 106 has functional blocks 110, 204, 206 coupled to the active base die 108 as chiplets via native interconnects 210 of the chiplets 110, 204, 206. As an intentional part of the design, the active base die 108 has incorporated functional blocks 202 into the active base die 108. The example active base die 108 can be designed to place the relevant functional blocks 202 near the native interconnects 210 of the chiplets 110 to which the functional blocks are to be connected. This results in direct routing between the components 110, 202 over very short data paths having a length comparable to the size of the native interconnects 210 of the chiplets themselves, which are in the order of microns.

[0041] Figure 3 is a diagram showing an example wafer-to-wafer (W2W) of direct bonding between native conductors of a die on a first wafer and conductors of an active base die on a second wafer to form a native interconnect via a W2W bonding process (such as referred to as hybrid bonding or DBI). The native conductors may be disposed on, at, or below a surface defined by an insulating material that may separate one or more native conductors from other conductive features including other native conductors. The insulating material may be polished to form an interface for bonding and electrical interconnection. When the insulating material of one die or wafer is brought into contact with another die or wafer (such as a die or wafer having a corresponding insulating and conductor interface), it may advantageously form a mechanical bond. The conductors may be melted together simultaneously or subsequently, such as by raising a temperature sufficient to cause the conductors to expand so that the opposing conductors are pressed together to form a continuous electrical connection.

[0042] Example microelectronic devices having the benefits of native interconnects and / or active base die 108, such as some of the devices shown below ( Figure 4 ), can be made of two or more semiconductor wafers 302, 304, 306, 108, which are aligned, bonded into a stack 308, and cut into 3D ICs containing native interconnects and / or active base dies 108. In one specific implementation, each wafer can be thinned before or after bonding to enhance signal transmission through layers and between layers. The bottom wafer 108 can have an active base die 108, while the upper wafers 302, 304, 306 can have other active dies to be directly bonded to the active base die 108 and to each other via directly bonded native interconnects. Cutting produces an example of an example microelectronic device 310. In some examples, the base die and / or wafer can be implemented in semiconductor, oxide, glass or other materials. For convenience and simplicity of discussion, this article generally uses a specific implementation of active devices formed in semiconductor materials.

[0043] Vertical connections between layers 302, 304, 306, 108 are imparted by a direct bonding process (such as DBI) to create native interconnects, but other conventional vertical connections can also be built into the wafer before bonding, or formed in the stack 310 after bonding. For example, semiconductor through vias (TSVs in this article) can optionally pass through the silicon or other semiconductor substrate between active layers and / or between active layers and external bonding pads. In general, TSVs, TOVs (through oxide vias) or TGVs (through glass vias) can be interconnected through the wafer material or other materials of the example active base tube core 108 to, for example, connect one side to another side.

[0044] In one implementation, the direct bonding process can be operated on heterogeneous wafers because heterogeneous integration does not stop the formation of native interconnects. Directly bonded native interconnects and the lack of standard interfaces using native interconnects also greatly contribute to signal propagation speed and power density prospects. Conventionally, up to one-third of the power used by a given die is due to its wiring, and native interconnects greatly reduce the length of conductors in the circuit, thereby greatly reducing the power requirements of a given die.

[0045] Native interconnects allow native signals to pass through the corresponding die while maintaining power consumption levels as if the native signals were kept on the chip. The shorter "wires" or conduction paths of native interconnects also reduce power consumption by creating less parasitic capacitance. Reducing overall power consumption also results in less heat generation, such as extended battery life, and overall lower operating costs.

[0046] Figure 4 Various example configurations of microelectronic devices incorporating active base die 108 are shown. Some example configurations show the result of die-to-die or die-to-wafer direct bonding to form native interconnects between a first die, such as chiplet 206, and a second die, such as active base die 108. Other configurations show native interconnects made by direct bonding between active base die 108 and IP core logic of other active dies via a wafer-to-wafer (W2W) process. Figure 4 The configurations shown in are examples of directly bonded native interconnects and active base dies. The examples shown are not intended to provide an exhaustive set of configurations. Many other configurations are possible. Two active dies connected by corresponding native conductors and / or one or more native interconnects do not have to be in a face-to-face configuration. The two active dies (such as an active base die and another active die, such as a chiplet) can be facing up or facing down. The example native interconnects do not have to be between face-to-face dies, but the active dies can also be, for example, face-to-back or back-to-back.

[0047] The example microelectronic device 402 includes a chiplet 404 that is directly bonded to the example active base die 108 in a die-to-die or die-to-wafer process.

[0048] The example microelectronic device 406 includes stacked chiplets 408 and unstacked chiplets 410 of various heights bonded directly to the example active base die 108 in a die-to-die or die-to-wafer process.

[0049] The example microelectronic device 412 includes a mixture of very small chiplets 414 (eg, micron-sized) and relatively large chiplets 416 bonded directly to the example active base die 108 in a die-to-die or die-to-wafer process.

[0050] The example microelectronic device 418 includes a very small chiplet 420 of dimensions 0.25 x 0.25 microns bonded directly to the example active base die 108 in, for example, a die-to-die or die-to-wafer process.

[0051] The example microelectronic device 422 includes a very small chiplet 424 of, for example, micron size bonded directly to an example active base die 108 of the same size or footprint as the example chiplet 424 .

[0052] The example microelectronic device 426 includes a large chiplet 428 bonded directly to an example active base die 108 of the same size or footprint as the chiplet 428. In general, the chiplet size is not required, but it is practical to make a given chiplet size a multiple or fraction of the size of the active base die 108.

[0053] The example microelectronic device 430 includes chiplets 432, 434, 436 bonded in a stacked fashion directly to an example active base die 108 of the same size or footprint as the chiplets 432, 434, 436. For example, this example configuration of the microelectronic device 430 using the active base die 108 to host one or more memory controllers can be used to make or emulate various types of high bandwidth memory modules (such as DDR4 SDRAM, DDR5 SDRAM), high bandwidth memory (HBM), hybrid memory cube (HMC), etc.

[0054] The example microelectronic device 438 includes example chiplets 440 , 442 bonded directly to opposite sides of an example active base die 108 having connecting conductors on both major sides.

[0055] The example microelectronic device 444 includes an example active base die 108 arranged in multiple planes with example chiplets directly bonded to multiple sides of the example active base die 108 .

[0056] The example microelectronic device 446 includes a plurality of example active base dies 108 , 108 ′, 108 ″ bonded to each other and to respective example chiplets 448 , 450 , 452 .

[0057] The example microelectronic device 454 includes an example active base die 108 embedded in a substrate 456. The example embedded active base die 108 has conductive contacts on opposing sides and is smaller than the chiplets 458, 460 bonded directly to the example active base die 108.

[0058] The example microelectronic device 462 includes an example active base die 108 embedded in an example chiplet 464. The example chiplet 464 with the embedded active base die 108 is directly bonded to another chiplet 466, both directly and also via the embedded active base die 108.

[0059] The example microelectronic device 468 includes the example vertical active base die 108 bonded directly to the sidewalls of chiplets in a chiplet stack 470 , which are bonded to the substrate 456 .

[0060] The example microelectronic device 472 includes an example active base die 108 that is directly bonded to a chiplet 474 and also accommodates a conventional standard interface 476 to connect the chiplet 478 .

[0061] The example microelectronic device 480 includes example chiplets 482 , 483 having native interconnects on two opposing sides of the chiplets 482 , 483 for direct bonding to a plurality of active base dies 108 , 108 ′.

[0062] The example microelectronic device 484 includes example chiplets 482, 483, 485 having native interconnects on two opposing sides of the chiplets 482, 483, 485 to directly bond to multiple active base dies 108, 108' and form chiplet stacks 483, 485 between the multiple active base dies 108, 108'.

[0063] The example microelectronic device 486 includes example chiplets 487 , 488 embedded in the example active base die 108 .

[0064] The exemplary microelectronic device 490 includes an exemplary active die 491 bonded directly to the active base die 108 in wafer-to-wafer (W2W) fabrication.

[0065] The exemplary microelectronic device 492 includes an exemplary active die 493 that is directly bonded to the active base die 108, individually and in a stack, in wafer-to-wafer (W2W) fabrication after the corresponding wafer is thinned to produce the thin microelectronic device 492. In addition to the size reduction provided by the thinned wafer, the thinned wafer, for example down to 3 μm, provides an easier and more efficient path for signal traversal after direct bonding.

[0066] The exemplary microelectronic device 494 includes an exemplary active die 495 directly bonded to the active base die 108, both singly and in stacks, in wafer-to-wafer (W2W) fabrication. The microelectronic device 494 also includes redistribution layer (RDL) features 496 and one or more through silicon vias (TSVs) 497.

[0067] The example microelectronic device 498 includes an example double-sided active base die 108 in wafer-to-wafer (W2W) manufacturing, which has active components and corresponding conductors on both sides of the active base die 108, and active dies 499 and x403 are built on both sides of the double-sided active base die 108.

[0068] Example microelectronic device x404 includes example active dies x406 and x408 bonded directly to one side of active base die 108 in wafer-to-wafer (W2W) fabrication, with chiplets x410 and x412 bonded directly to the opposite side of active base die 108 .

[0069] The example microelectronic device x414 includes back-to-back or stacked active base dies 108 and 108', with active components of the back-to-back active base dies 108 and 108' bonded and / or directly joined to each respective active base die 108 or 108'. The available sides of the back-to-back active base dies 108 and 108' can be directly bonded to the native interconnects of the respective chiplets x416 and x418 and the stack of chiplets x420 and x422, or can be directly bonded to other active dies via wafer-to-wafer (W2W) fabrication.

[0070] Figure 5An example active base die 108 is shown included within an example microelectronic device 502, such as an integrated circuit package 502. In one implementation, the die's native conductors 504, such as example chiplets 506, 508, 510, ... n, are connected directly to the active base die 108, rather than to conventional components, such as industry standard interfaces 112, conventional interconnect layers, or passive interposers that conventionally connect chiplets and dies to the package. The native conductors 504 may be native interconnects, contacts, lines, wires, or pads that are in electrical contact with the core side of the IP core, and are thus communicatively coupled to the native signals of a given chiplet 506. Some of the native conductors 504 of a chiplet 506 may be accessible by the manufacturer, i.e., the chiplets 506 may be manufactured specifically for a given active base die 108. Such connections between the native conductors 504 of the chiplet 506 and the active base die 108 can replace and eliminate the need for an industry standard interface 112 in the microelectronic device 502, thereby providing a plethora of benefits.

[0071] By utilizing chiplets 506 and their native interconnects (504) to connect directly to the active base die 108, an example system, such as a microprocessor system, can be partitioned among multiple configurable components. For example, certain functions, particularly more customized or confidential portions of the system, can be provided by circuits and blocks on the active base die 108. Certain other functions, such as more conventional or less customized portions of the system, can be provided by circuits and blocks on a secondary die, chiplets 506, 508, 510, ... n, particularly when the secondary die is significantly smaller than the active base die 108. The chiplets 506, 508, 510, ... n can be aligned and docked at one or more locations on the active base die 108 to closely interconnect with the relevant portions of the active base die 108.

[0072] As an example configuration, certain memory IP cores may be generally aligned with processor cores or execution engines to allow for minimum trace length and maximum speed. More common and standardized cores such as phase locked loops (PLLs), memories, etc. may be moved out of the active base die 108, thereby freeing up space on the active base die 108. Such partitioning may also allow the active base die 108 and various IP core dies to be produced at different semiconductor processing nodes and run at different voltages, all within the same example microelectronic device 502.

[0073] In one implementation, the active base die 108 may be formed at a first process node, such as 5 nm. The secondary dies 506, 508, 510, ... n may be formed at a more mature or traditional node, such as 250 nm. If both the active base die 108 and the secondary dies 506, 508, 510, ... n utilize fine pitch interconnect technology, such as described below (Direct Bond Interconnect) hybrid technology, then the underlying chips can be interconnected even though they have different process node parameters (Xperi Group, Inc., Ziptronix, San Jose, California). This inter-die interconnect capability greatly simplifies the required routing, especially compared to conventional integrated microprocessor dies. Utilizing multiple dies and chiplets 506 saves manufacturing costs because the active base die 108 and secondary dies 506, 508, 510...n can be produced at a much lower cost than the monolithic integrated die 200, and have a smaller size, better performance, and lower power requirements.

[0074] Example Active Base Die

[0075] In one implementation, the active base die 108 is a silicon or other semiconductor die, and can act like a substrate, physically supporting the smaller chiplets 506, 508, 510, ... n. In some implementations, the active base die 108 can be smaller than the attached chiplets. In some cases, the active base die 108 can be made of a substrate material such as a polymer, with an embedded semiconductor die, or the active base die 108 can be primarily silicon or semiconductor, with other materials present for various reasons. The active base die 108 contains active circuits and functional blocks 512, which provide a particular integrated circuit 502 with its functional identity. The customization of the particular microchip system at hand is in or on the active base die 108, while the chiplets 506 are typically standard, well-established, or ubiquitous units, often containing proprietary IP blocks.

[0076] The example active base die 108 can be distinguished initially from a conventional passive interposer, which has one or more layers of passive conductive lines, typically connecting, for example, a conventional standard interface 112 of various dies in a 2.5D assembly. The active base die 108 is able to connect directly to logic with minimal drive distance, whereas a conventional passive die would have too many crossovers and clutter. Although different from a passive interposer, in a specific implementation, the example active base die 108 can additionally combine all of the features of a passive interposer, as well as the features of the active base die 108 as described herein.

[0077] Further distinguishing the active base die 108 from a conventional passive interposer, the active base die 108 may include one or more state elements 514 that are typically found only on a single die for conventionally connecting blocks within a conventional chip, but the active base die 108 actively uses these same state elements to connect signals from one die or chiplet 506 to another die or chiplet. The active base die 108 may also recruit state elements on one or more chiplets 506, 508, 510, ... n for driving on the active base die 108.

[0078] The recruited state element 514 may be a single state element, or may be a plurality of state elements bundled together, such as inverters and repeaters, and components such as buffers, drivers, signal repeaters, state machines, voltage regulators, timing components, etc. However, in a specific implementation, these example components may be located only on the active base die 108, rather than on the chiplets 506, 508, 510 as in conventional techniques. Thus, the active base die 108 may have its own on-board state element 514 and other supporting components to coordinate and connect the different die and chiplets into a working microchip system, but depending on the specific implementation, it is also possible to utilize existing state elements, such as drivers, inverters, repeaters, etc., which are attached to the die and chiplets connected to the active base die 108.

[0079] In one implementation, the active base die 108 may have a design that also replaces the state machine with latches instead of flip-flops to improve performance and efficiency, and reduce power requirements, as described further below.

[0080] The active base die 108 uses chiplets 506, 508, 510, ... n and communicatively connects them together, rather than relying on a monolithic integrated circuit design. In addition, the length of the data path formed by the interconnects between the active base die 108 and the native conductors 504 of a given chiplet 506 can be short, for example, as short as 1 um or less. The active base die 108 is therefore able to receive native signals directly from different chiplets and can freely connect and adjust these native signals between different dies and chiplets, thereby directly routing signals over, under, or through large IP blocks, which typically constitute the main blockage in conventional large chips or processors.

[0081] The circuits and blocks 512 within the active base die 108 are arranged and customized to provide the particular microelectronic device 502 or system at hand and integrate the IP blocks of the chiplets 506 , 508 , 510 . . . n into the microelectronic device 502 .

[0082] The active base die 108 can be designed to make electrical contact with the native conductors 504 of the chiplets 506 when they are natively placed on each chiplet, instead of connecting each chiplet 506 to the conventional standard interface 112. Eliminating the conventional standard interface 112 eliminates various types of unnecessary overhead. Significant overhead is eliminated because the native signals of the chiplets 506, 508, 510...n can be passed directly and in an undoped state to the active base die 108 through the extremely short data paths of the native interconnects 504, typically consisting of just about each conductive contact point 516 between the corresponding native conductors 504, 504', 504' of the chiplets 506, 508, 510...n and the active base die 108. The short data paths and elimination of hardware that typically modifies the native signals to fit the standard interface 112 provide many benefits. Removing the standard interface 112 from the package 502 eliminates the entire hierarchy of data processing complexity, and providing a short data path to interface with the active base die 108 provides a domino effect of simplification.

[0083] Once the native signals of the chiplets 506 are passed to the active base die 108, they can be communicatively coupled to functional blocks 512 or other components formed in the active base die 108 at locations at or near the interconnects to the native conductors 504 of a given chiplet 506. Each active base die 108 can be customized to have effectively placed circuits and functional blocks for interfacing with the native conductors 504 of the attached chiplets 506, 508, 510, ... n. In turn, the native signals of each chiplet 506 are efficiently routed and modified within the active base die 108 as needed to other functional blocks 512 within the active base die 108, and significantly, to other dies or chiplets 508, 510, ... n that can contact the active base die 108 via their respective native conductors 504.

[0084] Thus, the active base die 108 can eliminate the interconnect placement, pad layout, and spacing requirements of the feature design of the industry standard interface 112. The example active base die 108 can save a significant amount of unnecessary redistribution routing because the chiplets 506 are directly connected to the active base die 108 regardless of where the native conductors 504 themselves are located on a given chiplet, resulting in a minimum drive distance.

[0085] The active base die 108 can accommodate multiple interconnect types on the same active base die 108, providing more flexibility than is available in conventional industry. While providing a customized architecture to enable bidirectional communication between the functional elements of the active base die 108 and the off-the-shelf chiplets 506, 508, 510, ... n, the active base die 108 also utilizes voltage regulation to accommodate voltage differences and address voltage leveling between different chiplets and components.

[0086] Using the example active base die 108 can significantly reduce the size and area of ​​the package 502 and reduce power requirements, especially when emulating large hard IP chips. If necessary, the example active base die 108 can integrate repeater units for longer paths. The example active base die 108 can also improve signal quality with data transmission schemes, improve timing, and provide native high-speed interfaces.

[0087] Example Chiplet Technology

[0088] Typically, a chiplet is a die that can be included in a 2.5D or 3D assembly, but is not on the base of the stack. The chiplet 506 can be manufactured in various silicon foundry (process) nodes, such as 250nm, 180nm...28nm, 22nm, 5nm, etc., and various flavors (HPP, HPC, HPC+, etc.), which may exhibit different operating voltages. Voltage differences may make the die mismatched, and having a conventional standard interface 112 is generally intended to remedy these operating voltage variations.

[0089] Silicon IP providers invest significant effort into characterizing and validating a certain IP for each foundry node and combination of features that the IP provider intends to offer in a chiplet 506. This characterization is performed over a space of different foundry process conditions, voltages, and temperatures.

[0090] Each additional IP variant is a significant financial burden and potential loss of opportunity. However, once the IP has been characterized and verified, the IP provider guarantees its performance unless the IP is modified. Once modifications are made, the characterization data is no longer valid and the IP provider no longer guarantees the performance of the IP and its chiplet implementations.

[0091] In various implementations, the chiplets 506, 508, 510, ... n can have their native core-side interconnects, but can be manufactured to not include conventional standard interfaces 112. In one implementation, each chiplet 506 can have minimal circuitry to attenuate the signal to a minimum threshold to prevent damage to the circuit. A given chiplet 506 can also have a regulator or state element recruited by the active base die 108 for the entire microchip system 502.

[0092] In one implementation, the example chiplet 510 has multiple independent functions and multiple ports that can communicate with multiple functional elements. The example chiplet 510 can have communication paths between its independent on-board functions. In one implementation, the chiplet 510 can be a memory device with two or more independently addressable memory blocks. The active base die 108 can interface with the native signals of such an example chiplet 510 and take advantage of these features.

[0093] Example Interconnects

[0094] Typically, for widespread commercial applications, conventional chiplets typically include proven silicon IP blocks. Typically these include at least one standard interface 112, and the die size and power grow to accommodate these standard interfaces 112, which are typically not optimized for the IP blocks. For larger systems like processor chips, standard interfaces 112 may need to be located on all sides of the processor at or beyond the periphery of the functional processor block. In addition, relatively lengthy routing from each edge of the processor core to the standard interface 112 may be required. If the processor is 3×5 mm in size and each standard interface 112 is 2 mm long, then routing of the 3 mm long edge typically needs to be reduced to a 2 mm long interface, and routing of the 5 mm long edge typically needs to be routed to one or two 2 mm long standard interfaces 112, all of which have an impact on routing length, congestion, and power requirements.

[0095] In one implementation, an example native interconnect using the active base die 108 is directly coupled to a native core-side interconnect 504 that is natively present on the chiplet 506. The native interconnect is intended to use the inherent native placement of the native conductors 504 as they are located on the chiplet 506 as placed by the manufacturer. By enlisting the native interconnects of the chiplets 506, 508, 510, ... n, rather than the conventional standard interface 112, the active base die 108 is intended to reproduce and improve various architectures, such as ASICs, ASSPs, and FPGAs.

[0096] The interconnection between the active base die 108 and the primary conductor 504 or other active die of the chiplets 506, 508, 510 can be achieved by a variety of different technologies. The signal spacing within a given die can be in the range of 0.1 to 5.0 microns. The primary conductor 504 can have a spacing of about 3um (micrometers), so the bonding technology must be able to target a small pad surface and align the conductors to be connected to each other sufficiently to meet the minimum overlap requirements for conductivity. Various techniques for fine pitch bonding can be used, such as copper diffusion bonding, in which two fine pitch copper conductors are pressed against each other when metal diffusion bonding occurs, usually under pressure and elevated temperatures. Where spacing permits, amalgams such as solder can be used. Copper nanoparticle technology and hybrid interconnect technology can also be used for interconnection. Wires can be used in some cases. Another example interconnect technology that can be used in some cases is described in U.S. patent application Ser. No. 15 / 257,427, filed on Sept. 6, 2016, and entitled “3D-Joining of Microelectronic Components with Conductively Self-Adjusting Anisotropic Matrix,” which is incorporated herein by reference in its entirety, wherein an anisotropic matrix of conductive nanotubes or wires automatically self-adjusts to form connections between conductors on two surfaces that are not perfectly aligned with each other, and to make no connection when there is no overlap between the conductors on the surfaces being connected.

[0097] In a specific implementation, application (Direct Bond Interconnect) Hybrid Bonding Technology. DBI bonding is currently available for fine pitch bonding in 3D and 2.5D integrated circuit components and can be applied to bond the native conductor 504 of chiplets 506, 508, 510...n to the active base die 108 (Xperi Group, Inc., Ziptronix, San Jose, California). See, for example, U.S. Patent No. 7,485,968, which is incorporated herein by reference in its entirety. DBI bonding technology has been demonstrated to have an interconnect pitch of 2um. DBI bonding technology has also been demonstrated to be as low as 1.6um pitch in a wafer-to-wafer approach, which does not achieve individual die pitch limitations through pick and place (P&P) operations (pick and place surface mount technology machines). With DBI technology, under-bump metallization (UBM), bottom fill, and micro-bumps are replaced by DBI metallization. Bonding at the die level is started at room temperature, followed by batch annealing at low temperatures. Direct bonding is also used in some cases (Xperi Group, Inc., Ziptronix, Inc., San Jose, CA).

[0098] Figure 6 An example core IP cell 600 of an example chiplet 506 is shown. Native core side interconnect pads 602 (not shown to scale) in the array 604 provide the native conductors 504 to bond to complementary bonding pads 606 on the active base die 108. DBI bonding or interconnection across an interface having a native conductor 504 on one side and complementary pads 606 or contacts to the active base bottom die 108 on the other side is scalable and limited only by the accuracy of chiplet placement during the pick and place (P&P) phase of the example operation. For example, if P&P can handle a placement accuracy of 1 um, and the pad overlap requirement is 50%, i.e., 50% of each pad 602 must overlap with the complementary pad 606 in both the x and y axes, then for these or other native conductors 504, for 2×2 um pads 602, the minimum pad spacing should be greater than 3 um. If both the x and y axes are shifted (misaligned) by the maximum allowed of 50% per axis, then 25% or one quarter of the pad 602 is allowed to overlap with the complementary pad 606 .

[0099] This fine pitch bonding of the interconnect 602 with DBI bonding and other techniques enables interconnects between the pads 606 or contacts of the active base die 108 and the native conductors 504 (core side interconnect pads 602) of the chiplet 506 with minimal or no changes to the native pitch, placement, and geometric pad configuration of the silicon proven IP and the core IP unit 600 of the chiplet. Most core side interconnects are currently at 3um pitch and DBI bonding can be performed in the array 604. In one specific implementation, a larger pitch can be used in the small array 604, such as four rows of pads 602 or native conductors 504 at 12um pitch. This means that the conductive routing to the array 604 will be at least an order of magnitude shorter than the routing required to connect to the traditional standard interface 112. The native interconnects 602 have a fine enough pitch so that they can exist in sufficient numbers to eliminate conventional serialization of the output to accommodate the limited pin count of the standard interface 112. This also eliminates the burden of delay and having to provide power for conventional serialization because no buffers or entire artificial interface construction are required.

[0100] Voltage Adaptation in Active Base Die

[0101] The active base die 108 can provide voltage adaptability in order to couple with different chiplets 506, 508, 510, ... n, which can have different operating voltages from each other. For example, a half-node 28nm chiplet can operate in a voltage range of 0.9 to 1.1 volts, while a 5nm chiplet can operate at 0.6 to 0.85 volts, with no voltage range overlap. To accommodate these voltage differences, the active base die 108 can also provide improved voltage control over conventional voltage leveling measures by enabling a greater number of independent power domains, each of which can be independently managed in the active base die 108. For example, this can allow a CPU core to run at an elevated voltage and frequency to meet a large computational load, while other cores also execute lower priority code at a lower voltage and frequency to save power. Adding one or more voltage conversion stages can also improve power efficiency. The active base die 108 can provide such adaptive voltage leveling in a variety of ways. Figure 7 An example microelectronic package 700 is shown with an active base die 108 and voltage regulators 702, 704. In one implementation, the active base die 108 has a compact voltage regulator dedicated to each set of chiplets 506, 508, 510, ... n ("chiplet"), thereby creating a corresponding voltage domain 710 for the chipset. That is, different chipsets each share a dedicated voltage regulator 702 or 704 integrated in the active base die 108. The voltage regulator 702 provides a potential of 1.2 volts to the chipset, which includes the chiplets 506, 508, 510 in the domain 710. The voltage regulator 704 provides a potential of 1.0 volts to the chipset, which includes the chiplets 706, 708 in the domain 712. In one implementation, these voltage regulators 702, 704 can be passive.

[0102] Figure 8 An example microelectronic package 800 is shown having an active base die 108 and multiple voltage regulators 802, 804, 806 and 808, 810. In this implementation, a single voltage regulator is placed in the active base die 108 near the I / O interface of each chiplet 506, 508, 510 and 706, 708. This one-per-die approach ensures that each chiplet 506, 508, 510, 706, 708 has its required voltage level, and the approach can improve power integrity. Because the voltage regulators 802, 804, 806, 808, 810 are closer to their respective dies, there are fewer parasitic effects, and therefore less IR drop and droop.

[0103] In another implementation, the active base die 108 has voltage control capability to overdrive or underdrive the chiplets 506, 508, 510, 706, 708. Overdriving or underdriving enables proper voltage overlap for voltage leveling or enables better operation between dies with different operating voltages.

[0104] Thus, the example active base die 108 can accommodate chiplets 506 , 508 , 510 at various different operating voltages for various semiconductor manufacturing technologies, either by providing a one-to-one regulator for the various chiplets, or by having different voltage domains on the active base die 108 for groups of chiplets.

[0105] Timing and Priority

[0106] Fig. 9 An example microelectronic package 900 with an active base die 108 is shown, which includes a clock 902 for timing and synchronizing processing steps and data transmission. The example active base die 108 can provide a global or regional clock signal in the active base die 108, which can also be used for timing and synchronization interaction with the chiplets 506. The clock signaling is enhanced to synchronize data transmission, which utilizes the short data paths of the native interconnect to interconnect with the chiplets 506, 508, 510...n and minimal routing blocking, thereby improving performance. Depending on the specific implementation, the active base die 108 can have a clock 902 internally or externally thereto, and in various specific implementations, various communication channel schemes can be included, such as an on-board communication network 904 or a bus 906 or bus.

[0107] Fig.10 An example microelectronic package 1000 is shown with an active base die 108 and an example negotiation engine 1002 or out-of-order engine. The example active base die 108 and negotiation engine 1002 can improve performance by determining which functional blocks in the active base die 108 have current priority for unidirectional or bidirectional communications with chiplets 506, 508, 510...n. The active base die 108 can also determine the priority among the execution engines and functional blocks for a given instruction, both within the active base die 108 and relative to the chiplets 506, 508, 510...n. In communications and priorities, the active base die 108 has the advantage of large IP blocks being located in the chiplets 506, thereby alleviating many routing congestion. This can move data traffic away from the backbone of the layout, thereby improving timing. Soft logic can also be improved over a larger area, which eliminates mitigation circuits that are typically used to retime and resend signals.

[0108] During RTL design, logic synthesis applied to the design of the active base die 108 can place repeater cells where needed for longer data routing. Where applicable, flip state machines can be replaced with latches to further improve efficiency. Repeaters and signal repeaters for longer channel lengths can be inserted using synthesis tools such as timing closure tools as needed during design. Synthesis tools can also simulate the microchip system 502, perform retiming and level shifting, and can insert inverting nodes into the design to close timing paths.

[0109] The active base die 108 typically has fewer repeaters than a comparable conventional layout because blocking is reduced by moving large IP blocks to the chiplets 506. Also, there are shorter path delays due to the direct and very short interconnects between the native interconnects 504 of the chiplets 506 and the active base die 108. Alternatively, the chiplet timing can be turned off to the state drivers and electronic design automation applied at the hierarchy level.

[0110] In one implementation, the active base die 108 achieves improved performance by using a double data rate (DDR) data transfer protocol, transferring data on both the rising and falling edges of an onboard clock signal. In another implementation, the active base die 108 may use a quad data rate (QDR), performing four data transfers per clock cycle.

[0111] The active base die 108 may also utilize other devices to accelerate performance, such as a negotiation engine 1002 or an out-of-order engine to stage data and instructions between execution engines.

[0112] Neural Network Implementation

[0113] Fig.11 An example neural network implementation of a microelectronic device 1100 is shown using an example active base die 108. Typically, an architecture for configuring a neural network may include many large conventional general purpose processors, where the cores of conventional hardware are recruited through programming to establish a neural network paradigm.

[0114] However, to build 3D volumes of neurons or convolutional neural networks for image analysis, machine vision, computer vision, or various forms of artificial intelligence, the recruitment and layout of conventional large processors becomes cumbersome and ultimately fails in the task, or provides an inefficient solution because the large processors are not really optimized for the nuances and larger neuron layouts of evolving neural network architectures.

[0115] Fig.11The active base die 108 in the FIG. 104 provides an ideal medium for growing larger and more complex neural network topologies. First, the active base die 108 can be scaled to large sizes and can contain advantageously repeated instances of the supporting elements required for a given neural network architecture. Next, the large monolithic conventional processor of a conventional network design can be replaced by a larger field of one or more repeatable and very small processing elements, each of which is represented in a chiplet 1102 coupled to the active base die 108 for very efficient and unburdened processing of native signals from each of these processing elements 1102. The active base die 108 can also include a global synchronization clock 902 to coordinate the timing across the field of multiple chiplets 1102 providing processing elements. The clock 902 can make the active base die 108 scalable for very large neural network configurations.

[0116] The physical structure of the active base die 108 with a field of attached processing element chiplets 1102 can represent the neurons and synapses of neural networks and biological brain system models, rather than typically imposing a neural network paradigm on a general purpose CPU chip, which is not up to the task of representing evolving neuronal architectures and may ultimately not have the number of transistors required to represent biological neural networks or perform higher artificial intelligence.

[0117] Process Sharing

[0118] The example active base die 108 provides a unique opportunity for shared processing between dies or chiplets 506, 508, 510...n. The active base die 108 can be equipped with time borrowing capabilities to save power, reduce latency, and reduce footprint. In one specific implementation, the active base die 108 can implement an architecture in which a given functional element of the active base die 108 can communicate with multiple chiplets 506, 508, 510...n and can negotiate the priority of a specific communication among multiple other functional elements. It is worth noting that the active base die 108 can share processes and resources in the active base die 108 between chiplets of various technologies, such as chiplets manufactured at different foundry process nodes.

[0119] The active base die 108 can enable chiplets of various technologies to share one or more common memories, whereas conventionally each processor has its own dedicated coupled memory. The active base die 108 can allow external memory to be used as embedded memory with process sharing. In such a configuration, memory access does not need to be performed through a memory interface every time, such as a DBI bond of a native interconnect 504 to an attached chiplet 506, 508, 510...n, but memory access can be configured directly through the active base die. In addition, since certain processes can be configured as redundant and used to improve the yield of the stack by having one block on a given die share repair functionality with another block that may have a fault within the redundant block, the repair capability is enhanced. This capability is enhanced at least in part due to the number of interconnects available through the DBI process, the proximity of adjacent blocks on both sides and across the interface, and the elimination of most of the routing required in conventional arrangements.

[0120] Example Method

[0121] Fig.12 An example method 1200 of making a microelectronic device having native interconnects is shown. The operations of the example method 1200 are shown as separate blocks.

[0122] At block 1202 , native core-side conductors of a first die are directly bonded to conductors of a second die to form native interconnects between the first die and the second die.

[0123] At block 1204 , circuitry of the first die extends across a die boundary between the first die and the second die via a native interconnect, the circuitry spanning the native interconnect.

[0124] At block 1206 , a native signal of the IP core of the first die is passed between the core of the first die and at least one functional block of the second die by crossing a circuit through the native interconnect.

[0125] The native interconnect provided by the example method 1200 can provide a unique interface between a first die and a second die, while the native interconnect abandons standard interface geometry and input / output protocols. The first die can be made of a first manufacturing process node, and the second die is made of a different second manufacturing process node. When a native signal is passed through the native interconnect, the interface protocol and input / output protocol between the first die and the second die are abandoned across the circuit passing through the native interconnect.

[0126] The example method 1200 may also include directly bonding native core-side conductors of the plurality of dies across the plurality of die boundaries of the plurality of dies to form a plurality of native interconnects, and spanning the circuits through the plurality of native interconnects across the plurality of die boundaries. The plurality of native interconnects provide interfaces between the plurality of dies, and the interfaces abandon interface protocols and input / output protocols between the plurality of dies.

[0127] Example method 1200 can pass native signals between a functional block of a first die and one or more functional blocks of one or more other dies of a plurality of dies through one or more native interconnects while abandoning interface protocols and input / output protocols between the plurality of dies. Native signals can be passed unmodified between a core of a first die and at least one functional block of a second die through a circuit spanning through the native interconnects.

[0128] The native signal may be level shifted between the core of the first die and at least one functional block of the second die by circuitry crossing through the native interconnect, the level shifting accommodating a difference in operating voltage between the first die and the second die.

[0129] The example method 1200 can be implemented in a wafer-to-wafer (W2W) bonding process, for example, where a first die is on a first wafer and a second die is on a second wafer, and where the W2W bonding process includes directly bonding a native core-side conductor of the first die to a conductor of the second die to form a native interconnect between the first die and the second die, the native interconnect extending one or more circuits across a die boundary between the first die and the second die, the one or more circuits spanning through one or more native interconnects, the native interconnect providing an interface between the respective dies that abandons an interface protocol and an input / output protocol between the respective dies. The first wafer and the second wafer are made of heterogeneous foundry nodes, or the first die and the second die are made of incompatible manufacturing processes. In a specific implementation, the example method 1200 can directly bond native core-side conductors between some components of the first wafer and the second wafer to form native interconnects for transmitting native signals, but create other interfaces or standard interfaces on other components of the wafer for transmitting amplified signals in a microelectronic device produced by the W2W process.

[0130] The first die or the second die can be an active base die. The first die can also be a chiplet that includes an IP logic core and the second die includes an active base die. In some cases, the size of the chiplet can range from 0.25×0.25 microns, for example, to the same size as the active base die. The example method 1200 can stack the active base die and multiple chiplets in a stack or 3D stacked IC structure having multiple layers, wherein each layer in the stack or 3D stacked IC structure is directly bonded to form native interconnects between dies at different layers.

[0131] Fig.13 An example method 1300 of providing a microchip architecture for a semiconductor chiplet is shown. In the flow chart, the operations of the method 1300 are shown as separate blocks.

[0132] Native core-side conductors of a plurality of chiplets are connected to an active base die at block 1302. Native interconnects coupled to the active base die avoid the need for industry standard interfaces for conventional chiplets.

[0133] At block 1304 , native signals from each of the plurality of chiplets are received at one or more functional blocks in the active base die.

[0134] At block 1306 , bidirectional communications are conveyed between at least one functional block in the active base die and the plurality of chiplets across at least one cross-die boundary.

[0135] Fig.14 Another example method 1400 of providing a microchip architecture for semiconductor chiplets, including voltage regulation to accommodate different chiplets, is shown. In the flow chart, the operations of the method 1400 are shown as separate blocks.

[0136] At block 1402, a chiplet is selected for connection to an active base die.

[0137] At box 1404, the native core-side conductors in the multiple chiplets are variously connected to the active base tube die using native interconnects selected from the group consisting of: direct bond interconnect (DBI) metallization layers, copper to copper diffusion bonding, connections to conductive nanotubes, metal to metal contacts, hybrid interconnects.

[0138] At block 1406 , voltages are adjusted to accommodate chiplets from different semiconductor process nodes and / or chiplets having different operating voltages to the active base die via the chiplets' respective native interconnects.

[0139] Fig.15Another example method 1500 is shown that provides a microchip architecture of semiconductor chiplets using state elements in connected chiplets for signal driving in an active base die. In the flow chart, the operations of the method 1500 are shown as separate blocks.

[0140] At block 1502, native core-side conductors of a plurality of chiplets are connected to an active base die.

[0141] At block 1504, the active base die uses state elements of one or more chiplets in the plurality of chiplets for driving signals across a cross-die boundary between the active base die and the one or more chiplets. The cross-die boundary may be only 1 um thick, or even smaller.

[0142] In the above description and accompanying drawings, specific terms and figure marks have been set forth to provide a thorough understanding of the disclosed embodiments. In some cases, the terms and symbols may imply specific details that are not required for practicing those embodiments. For example, any specific size, quantity, material type, manufacturing steps, etc. may be different from those described above in alternative embodiments. The term "coupling" used herein means direct connection and connection through one or more intermediate circuits or structures. The terms "example", "embodiment" and "specific implementation" are used to express examples, rather than preferences or requirements. In addition, the terms "may" and "can" are used interchangeably to represent optional (allowed) topics. The absence of any one term should not be interpreted as meaning that a given feature or technology is required.

[0143] Various modifications and changes may be made to the embodiments presented herein without departing from the broader spirit and scope of the present disclosure. For example, the features or aspects of any embodiment may be used in combination with any other embodiment in the embodiments, or in place of corresponding features or aspects thereof. Therefore, the specification and drawings should be regarded as illustrative rather than restrictive.

[0144] While a limited number of embodiments have been disclosed with respect to the present disclosure, those skilled in the art having benefit of this disclosure will appreciate that numerous modifications and variations are possible given the description. The appended claims are intended to cover such modifications and variations as fall within the true spirit and scope of the present disclosure.

Claims

1. A system, include: an active base die; as well as a first chiplet attached to the active base die by a first hybrid bonding, wherein: The first chiplet includes a first functional block; The first hybrid bond forms a first data connection that communicatively couples the first functional block to the active base die; and The first data connection passes raw data signals bidirectionally between the first chiplet and the active base die.

2. The system according to claim 1, wherein the first functional block comprises a multiplier, an arithmetic logic unit (ALU), an instruction decoder, a digital signal processor (DSP), a subsystem intellectual property (IP) core, or a combination thereof. 3 . The system of claim 1 , wherein the first functional block comprises an intellectual property (IP) core, the IP core comprising a memory controller or a reusable logic unit.

4. The system of claim 1, wherein the first data connection bidirectionally transfers the raw data signal between the first chiplet and the active base die without using a standard interface including an input / output (I / O) protocol. 5 . The system of claim 1 , wherein the first hybrid bond has a pitch between 0.1 μm and 5 μm.

6. The system of claim 1, wherein the first hybrid bond forms a continuous circuit between the first chiplet and the active base die.

7. The system of claim 6, wherein the first chiplet and the active base die are configured for two-way communication via the continuous circuit.

8. The system of claim 1 further comprising a second chiplet attached to the active base die by a second hybrid bonding.

9. The system of claim 8, wherein the active base die includes one or more voltage regulators or voltage regulation domains for regulating voltage between the first chiplet and the second chiplet.

10. The system of claim 1, wherein the first chiplet comprises a first conductor, the first conductor is hybrid bonded to the active base die, and the first conductor comprises a repeater, a buffer, a driver, a redriver, a state machine, a voltage regulator, a timing component, or a combination thereof.

11. The system of claim 8, wherein the first chiplet and the second chiplet are attached to opposite sides of the active base die.

12. The system of claim 1, further comprising a second chiplet attached to the first chiplet, wherein the second chiplet is communicatively coupled to the active base die through an interconnect disposed through the first chiplet.

13. The system of claim 8, further comprising a third chiplet disposed on the second chiplet.

14. The system according to claim 1, further comprising: include: a second chiplet attached to the first chiplet by a second hybrid bonding, wherein: The second chiplet includes a second functional block; and The second hybrid bond forms a second data connection that communicatively couples the second functional block to the first chiplet.

15. The system of claim 14, wherein at least one of the first functional block or the second functional block comprises a memory controller.

16. The system of claim 1, further comprising a plurality of micron-sized second chiplets attached to the active base die by a second hybrid bonding, in: The second chiplets each include a second functional block; and The second hybrid bond forms a second data connection that communicatively couples the second functional block to the active base die.

17. A system, include: an active base die; as well as a first chiplet attached to the active base die by a first hybrid bonding, wherein: The first chiplet includes a first functional block; The first hybrid bond forms a first data connection that bidirectionally communicatively couples the first functional block to the active base die; and At least some of the first data connections convey raw data signals between the first chiplet and the active base die.

18. A system, include: an active base die; as well as a first chiplet attached to the active base die by a first hybrid bonding, wherein: The first chiplet includes a first functional block; The first hybrid bond forms a first data connection that communicatively couples the first functional block to the active base die; At least one of the first data connections communicates a raw data signal from the first chiplet to the active base die; as well as At least one of the first data connections transfers a raw data signal from the active base die to the first chiplet.

Citation Information

Patent Citations

  • 3d-joining of microelectronic components with conductively self-adjusting anisotropic matrix

    US20170069595A1

  • 3D IC method and device

    US7485968B2