GPU asynchronous direct memory access application

Abstract

The invention discloses a GPU asynchronous direct memory access application. One embodiment provides a graphics processor, comprising: a base die comprising a plurality of chiplet slots; and a plurality of chiplets, the plurality of chiplets being coupled with the plurality of chiplet slots. A chiplet of the plurality of chiplets includes: a graphics core cluster, the graphics core cluster including a plurality of graphics cores; a distributed shared local memory, the distributed shared local memory including a shared local memory within each of the plurality of graphics cores; and a direct memory access engine within each of the plurality of graphics cores, the direct memory access engine configured to asynchronously copy data from a memory device to the distributed shared local memory.

Description

Cross-references

[0001] This application claims priority to U.S. Provisional Application No. 63 / 566,260, filed on March 16, 2024, which is hereby incorporated by reference herein. Technical Field

[0002] The present disclosure relates generally to data processing, and more particularly to data processing via general-purpose graphics processing units. Background Art

[0003] In traditional hardware, if four thread groups (each dispatched to a separate graphics core) all load the same tile from memory into shared local memory, four separate loads from the L2 cache are executed. These loads may be duplicates and place unnecessary bandwidth pressure on the L2 cache and the L2 cache interconnect. BRIEF DESCRIPTION OF THE DRAWINGS

[0004] The embodiments described herein are illustrated by way of example and not limitation in the figures of the accompanying drawings in which like reference numerals indicate similar elements, and in which:

[0005] Figure 1 is a block diagram illustrating a computer system configured to implement one or more aspects of the embodiments described herein;

[0006] Figure 2A-2E shows parallel processor components, including a graphics multiprocessor;

[0007] Figure 3 A graphics processing unit is shown that includes a collection of dedicated graphics processing resources arranged as a multi-core group.

[0008] Figures 4A-4E An exemplary architecture is shown in which multiple GPUs are communicatively coupled to multiple multi-core processors;

[0009] Figure 5 shows the graphics processing pipeline;

[0010] Figure 6 Showing the machine learning software stack;

[0011] Figure 7 A general purpose graphics processing unit is shown;

[0012] Figure 8 shows a multi-GPU computing system;

[0013] Figure 9A-9B illustrates the layers of an exemplary deep neural network;

[0014] Figures 10A-10B An exemplary language model is shown;

[0015] Figure 11 Demonstrates the training and deployment of deep neural networks;

[0016] Figure 12A is a block diagram illustrating distributed learning;

[0017] Figure 12B is a block diagram illustrating a programmable network interface and a data processing unit;

[0018] Figure 13 An exemplary inference system-on-chip (SOC) suitable for performing inference using a trained model is shown;

[0019] Figure 14 It is a block diagram of the processing system;

[0020] Figures 15A-15C A computing system and a graphics processor are shown.

[0021] Figure 16 is a block diagram of a graphics processor, which may be a discrete or integrated graphics processing unit;

[0022] Figures 17A-17B A block diagram illustrating an additional graphics processor and computing accelerator architecture;

[0023] Figures 18A-18C shows thread execution logic including an array of processing elements employed in a graphics processor core;

[0024] Figure 19 illustrates a tile of a multi-chip processor according to one embodiment;

[0025] Figure 20 is a block diagram illustrating a graphics processor instruction format;

[0026] Figure 21 is a block diagram of the attached graphics processor architecture;

[0027] Figures 22A-22B Shows the graphics processor command format and command sequence;

[0028] Figure 23 An exemplary graphical software architecture for a data processing system is shown;

[0029] Figure 24 is a block diagram illustrating an IP core development system;

[0030] Figure 25A A cross-sectional side view of an integrated circuit package assembly is shown including multiple units of hardware logic chiplets connected to a substrate (e.g., base die);

[0031] Figure 25BA package assembly including interchangeable chiplets is shown;

[0032] Figure 26 is a block diagram illustrating a system-on-chip integrated circuit;

[0033] Figure 27 shows a graphics processing system including a graphics core cluster according to an embodiment;

[0034] Figure 28 A system is shown that enables data sharing across graphics cores within a graphics core cluster;

[0035] Figure 29 A system including a graphics core cluster with asynchronous DMA and MMA circuit modules is shown;

[0036] Figure 30 Showing a graphics core component capable of asynchronous DMA operations according to an embodiment;

[0037] Figure 31 is a block diagram illustrating circuit modules associated with an asynchronous DMA unit according to embodiments described herein;

[0038] Figure 32 A method for performing asynchronous DMA operations on a graphics processor is shown;

[0039] Figure 33 A method of performing cooperative asynchronous DMA operations on surfaces in memory is shown.

[0040] Figure 34 A method of synchronizing asynchronous DMA operations via addressable barrier objects is shown; and

[0041] Figure 35 is a block diagram of a computing device including a graphics processor according to one embodiment. DETAILED DESCRIPTION

[0042] Current parallel graphics data processing includes systems and methods developed to perform specific operations on graphics data, such as linear interpolation, tessellation, rasterization, texture mapping, depth testing, and the like. Traditionally, graphics processors have used fixed-function compute units to process graphics data. However, more recently, portions of graphics processors have been made programmable, enabling such processors to support a wider variety of operations for processing vertex and fragment data.

[0043] To further improve performance, graphics processors typically implement processing techniques such as pipelining, which attempts to process as much graphics data as possible in parallel across different parts of the graphics pipeline. Parallel graphics processors with a single instruction multiple thread (SIMT) architecture are designed to maximize the amount of parallel processing in the graphics pipeline. In a SIMT architecture, groups of parallel threads attempt to execute program instructions together synchronously as often as possible to improve processing efficiency.

[0044] A graphics processing unit (GPU) is communicatively coupled to a host / processor core to accelerate, for example, graphics operations, machine learning operations, pattern analysis operations, and / or various general-purpose GPU (GPGPU) functions. The GPU can be communicatively coupled to the host processor / core via a bus or another interconnect (e.g., a high-speed interconnect such as PCIe or NVLink). Alternatively, the GPU can be integrated on the same package or chip as the core and communicatively coupled to the core via an internal processor bus / interconnect (i.e., inside the package or chip). Regardless of the manner in which the GPU is connected, the processor core can assign work to the GPU in the form of a sequence of commands / instructions contained in a work descriptor. The GPU then uses dedicated circuit modules / logic to efficiently process these commands / instructions.

[0045] Although the techniques described herein are primarily discussed in the context of GPUs, the techniques may also be implemented in other types of processors, including but not limited to general-purpose processors and accelerator devices such as artificial intelligence accelerators, vision processors, and neural processing units.

[0046] In the following description, numerous specific details are set forth to provide a more thorough understanding. However, it will be apparent to those skilled in the art that the embodiments described herein can be practiced without one or more of these specific details. In other instances, well-known features are not described in order to avoid obscuring the details of the present embodiments. System Overview

[0047] Figure 11 is a block diagram illustrating a computing system 100 configured to implement one or more aspects of the embodiments described herein. Computing system 100 includes a processing subsystem 101 having one or more processors 102 and system memory 104 communicating via an interconnect path, which may include a memory hub 105. Memory hub 105 may be a separate component within a chipset assembly or may be integrated within one or more processors 102. Memory hub 105 is coupled to an I / O subsystem 111 via a communication link 106. I / O subsystem 111 includes an I / O hub 107 that enables computing system 100 to receive input from one or more input devices 108. Additionally, I / O hub 107 enables a display controller (which may be included in one or more processors 102) to provide output to one or more display devices 110A. In one embodiment, the one or more display devices 110A coupled to I / O hub 107 may include local, internal, or embedded display devices.

[0048] The processing subsystem 101 includes, for example, one or more parallel processors 112 coupled to the memory hub 105 via a communication link 113 (such as a bus or fabric). The communication link 113 can be one of any number of standards-based communication link technologies or protocols, such as, but not limited to, PCI Express, or can be a vendor-specific communication interface or communication fabric. The one or more parallel processors 112 can form a parallel or vector processing system in a computational cluster that can include a large number of processing cores and / or processing clusters, such as, a many-integrated-core (MIC) processor. For example, the one or more parallel processors 112 form a graphics processing subsystem that can output pixels to one of the one or more display devices 110A coupled via the I / O hub 107. The one or more parallel processors 112 can also include a display controller and a display interface (not shown) to enable a direct connection to the one or more display devices 110B.

[0049] Within the I / O subsystem 111, a system storage unit 114 may be connected to the I / O hub 107, thereby providing a storage mechanism for the computing system 100. An I / O switch 116 may be used to provide an interface mechanism to enable connection between the I / O hub 107 and other components, such as a network adapter 118 and / or a wireless network adapter 119 that may be integrated into the platform, as well as various other devices that may be added via one or more plug-in devices 120. The plug-in device(s) 120 may also include, for example, one or more external graphics processor devices, graphics cards, and / or computing accelerators. The network adapter 118 may be an Ethernet adapter or another wired network adapter. The wireless network adapter 119 may include one or more of the following: Wi-Fi, Bluetooth, near field communication (NFC), or other network devices including one or more wireless radio devices.

[0050] The computing system 100 may include other components not explicitly shown, including USB or other port connections, optical storage drives, video capture devices, etc., which may also be connected to the I / O hub 107. Figure 1 The communication paths for interconnecting the various components in the system may be implemented using any suitable protocol, such as a PCI (Peripheral Component Interconnect) based protocol (e.g., PCI Express), or any other bus or point-to-point communication interface and / or protocol(s) such as NVLink high-speed interconnect, Compute Express Link™, or the like. TM, CXL™) (e.g., CXL.mem), Infinity Fabric (IF), Ethernet (IEEE 802.3), remote direct memory access (RDMA), InfiniBand, Internet Wide Area RDMA Protocol (iWARP), Transmission Control Protocol (TCP), User Datagram Protocol (UDP), quick UDP Internet Connections (QUIC), RDMA over Converged Ethernet (RoCE), Intel Quick Path Interconnect (QPI), Intel Ultra Path Interconnect (UPI), Intel On-Chip System Fabric (IOSF), Omnipath, HyperTransport, Advanced Microcontroller Bus Architecture (AMD Bus Architecture (AMBA) interconnect, OpenCAPI, Gen-Z, Cache Coherent Interconnect for Accelerators (CCIX), 3GPP Long Term Evolution (LTE) (4G), 3GPP 5G and its variants, or wired or wireless interconnect protocols known in the art. In some examples, data can be copied or stored to the virtualized storage node using a protocol such as non-volatile memory express (NVMe) (NVME-oF) or NVMe over fabrics.

[0051] One or more parallel processors 112 may include circuit modules optimized for graphics and video processing (including, for example, video output circuit modules) and constitute a graphics processing unit (GPU). Alternatively or additionally, as described in more detail herein, one or more parallel processors 112 may include circuit modules optimized for general-purpose processing while retaining the underlying computing architecture. The components of the computing system 100 may be integrated with one or more other system elements on a single integrated circuit. For example, one or more parallel processors 112, the memory hub 105, (one or more) processors 102, and the I / O hub 107 may be integrated into a system on a chip (SoC) integrated circuit. Alternatively, the components of the computing system 100 may be integrated into a single package to form a system-in-package (SIP) configuration. In one embodiment, at least a portion of the components of the computing system 100 may be integrated into a multi-chip module (MCM), which may be interconnected with other multi-chip modules into a modular computing system. In some configurations, in addition to the processor(s) 102 and the parallel processor(s) 112, the computing system 100 also includes one or more accelerator devices 130 coupled to the memory hub 105. The accelerator device(s) 130 are configured to perform domain-specific acceleration of workloads to process tasks that are computationally intensive or require high throughput. The accelerator device(s) 130 can reduce the burden placed on the processor(s) 102 and / or the parallel processor(s) 112 of the computing system 100. The accelerator device(s) 130 can include, but are not limited to, intelligent network interface cards, data processing units, cryptographic accelerators, storage accelerators, artificial intelligence (AI) accelerators, neural processing units (NPUs), storage accelerators, and / or video transcoding accelerators.

[0052] It will be appreciated that the computing system 100 shown herein is illustrative and that variations and modifications are possible. The connection topology, including the number and arrangement of bridges, the number of processor(s) 102, and the number of parallel processor(s) 112, may be modified as desired. For example, the system memory 104 may be connected to the processor(s) 102 directly rather than through a bridge, while other devices communicate with the system memory 104 via the memory hub 105 and the processor(s) 102. In other alternative topologies, the parallel processor(s) 112 are connected to the I / O hub 107 or directly to one of the processor(s) 102 rather than to the memory hub 105. In other embodiments, the I / O hub 107 and the memory hub 105 may be integrated into a single chip. It is also possible for two or more sets of processors 102 to be attached via multiple sockets, which may be coupled to two or more instances of the parallel processor(s) 112.

[0053] Some of the specific components shown herein are optional and may not be included in all implementations of computing system 100. For example, any number of plug-in cards or peripherals may be supported, or some components may be eliminated. Additionally, some architectures may be specific to the implementation of the computer system 100. Figure 1 Different terminology is used for components that are similar to those shown in FIG. For example, memory hub 105 may be referred to as a north bridge in some architectures, while I / O hub 107 may be referred to as a south bridge.

[0054] Figure 2A A parallel processor 200 is shown. The parallel processor 200 may be a GPU, a GPGPU, or the like as described herein. The various components of the parallel processor 200 may be implemented using one or more integrated circuit devices, such as a programmable processor, an application specific integrated circuit (ASIC), or a field programmable gate array (FPGA). The parallel processor 200 may be Figure 1 One or more of the parallel processor(s) 112 shown in .

[0055] The parallel processor 200 includes a parallel processing unit 202. The parallel processing unit includes an I / O unit 204 that enables communication with other devices, including other instances of the parallel processing unit 202. The I / O unit 204 can be directly connected to the other devices. For example, the I / O unit 204 is connected to the other devices using a hub or switch interface (such as, memory hub 105). The connection between the memory hub 105 and the I / O unit 204 forms a communication link 113. Within the parallel processing unit 202, the I / O unit 204 is connected to a host interface 206 and a memory crossbar switch 216, wherein the host interface 206 receives commands related to performing processing operations and the memory crossbar switch 216 receives commands related to performing memory operations. In one embodiment, the I / O unit 204 is configured to enable secure I / O operations via trusted execution environment (TEE)-I / O support. TEE-IO enables trusted I / O virtualization, where a trust relationship can be directly established between a secure virtual environment, such as a trusted virtual machine, and a parallel processor 200 or a secure partition of a parallel processor.

[0056] When host interface 206 receives command buffers via I / O unit 204, it can direct work operations for executing those commands to front-end 208. In one embodiment, front-end 208 is coupled to scheduler 210, which is configured to dispatch commands or other work items to processing cluster array 212. Scheduler 210 ensures that processing cluster array 212 is properly configured and in a valid state before tasks are dispatched to processing clusters within the processing cluster array 212. Scheduler 210 can be implemented via firmware logic executing on a microcontroller. A microcontroller-implemented scheduler 210 can be configured to perform complex scheduling and work dispatch operations at both coarse and fine granularity, thereby enabling fast preemption and context switching of threads executing on processing cluster array 212. Preferably, host software can confirm workloads for scheduling on processing cluster array 212 via one of multiple graphics processing doorbells. In other examples, polling for new workloads or interrupts can be used to identify or indicate the availability of work to be executed. The workload may then be automatically distributed across the processing cluster array 212 by the scheduler 210 logic within the scheduler microcontroller.

[0057] The processing cluster array 212 may include up to "N" processing clusters (e.g., cluster 214A, cluster 214B, through cluster 214N). Each cluster 214A-214N in the processing cluster array 212 may execute a large number of concurrent threads. The scheduler 210 may assign work to the clusters 214A-214N in the processing cluster array 212 using various scheduling and / or work distribution algorithms that may vary depending on the workload generated for each type of program or computation. Scheduling may be handled dynamically by the scheduler 210 or may be assisted in part by compiler logic during the compilation of program logic configured for execution by the processing cluster array 212. Optionally, different clusters 214A-214N in the processing cluster array 212 may be assigned to process different types of programs or to perform different types of computations.

[0058] Processing cluster array 212 may be configured to perform various types of parallel processing operations. For example, processing cluster array 212 may be configured to perform general-purpose parallel computing operations. For example, processing cluster array 212 may include logic for performing processing tasks including filtering of video and / or audio data, performing modeling operations including physics operations, and performing data transformations.

[0059] Processing cluster array 212 is configured to perform parallel graphics processing operations. In such embodiments where parallel processor 200 is configured to perform graphics processing operations, processing cluster array 212 may include additional logic for supporting the execution of such graphics processing operations, including but not limited to texture sampling logic for performing texture operations, as well as tessellation logic and other vertex processing logic. In addition, processing cluster array 212 may be configured to execute graphics processing related shader programs, such as, but not limited to, vertex shaders, tessellation shaders, geometry shaders, and pixel shaders. Parallel processing unit 202 may transfer data from system memory via I / O unit 204 for processing. During processing, the transferred data may be stored in on-chip memory (e.g., parallel processor memory 222) during processing and subsequently written back to system memory.

[0060] In embodiments where parallel processing units 202 are used to perform graphics processing, scheduler 210 may be configured to divide the processing workload into tasks of approximately equal size to better enable distribution of graphics processing operations to multiple clusters 214A-214N in processing cluster array 212. In some of these embodiments, portions of processing cluster array 212 may be configured to perform different types of processing. For example, a first portion may be configured to perform vertex shading and topology generation, a second portion may be configured to perform tessellation and geometry shading, and a third portion may be configured to perform pixel shading or other screen-space operations to produce a rendered image for display. Intermediate data generated by one or more of clusters 214A-214N may be stored in a buffer to allow the intermediate data to be transferred between clusters 214A-214N for further processing.

[0061] During operation, processing cluster array 212 may receive processing tasks to be executed via scheduler 210, which receives commands defining the processing tasks from front-end 208. For graphics processing operations, a processing task may include data to be processed, such as surface (patch) data, primitive data, vertex data, and / or pixel data, as well as state parameters and an index of commands that define how the data is to be processed (e.g., what program is to be executed). Scheduler 210 may be configured to fetch an index corresponding to a task, or may receive an index from front-end 208. Front-end 208 may be configured to ensure that processing cluster array 212 is configured in a valid state before a workload specified by an incoming command buffer (e.g., a batch buffer, a push buffer, etc.) is initiated.

[0062] Each of the one or more instances of parallel processing unit 202 can be coupled to parallel processor memory 222. Parallel processor memory 222 can be accessed via memory crossbar switch 216, which can receive memory requests from processing cluster array 212 and I / O unit 204. Memory crossbar switch 216 can access parallel processor memory 222 via memory interface 218. Memory interface 218 can include a plurality of partition units (e.g., partition unit 220A, partition unit 220B, through partition unit 220N), each of which can be coupled to a portion (e.g., memory cells) of parallel processor memory 222. The number of partition units 220A-220N can be configured to be equal to the number of memory cells, such that each partition unit 220A-220N has a corresponding memory cell 224A-224N. In other embodiments, the number of partition units 220A-220N may not be equal to the number of memory cells.

[0063] Memory units 224A-224N may include various types of memory devices, including dynamic random-access memory (DRAM) or graphics random access memory, such as synchronous graphics random access memory (SGRAM), including graphics double data rate (GDDR) memory. Optionally, memory units 224A-224N may also include 3D stacked memory, including, but not limited to, high bandwidth memory (HBM). Those skilled in the art will appreciate that the specific implementation of memory units 224A-224N may vary and may be selected from a variety of conventional designs. Render targets, such as frame buffers or texture maps, may be stored across memory units 224A-224N, allowing partition units 220A-220N to write portions of each render target in parallel to efficiently use the available bandwidth of parallel processor memory 222. In some embodiments, local instances of parallel processor memory 222 may be eliminated in favor of a unified memory design that utilizes system memory in conjunction with local cache memory.

[0064] Optionally, any of the clusters 214A-214N in the processing cluster array 212 has the capability to process data to be written to any of the memory units 224A-224N within the parallel processor memory 222. The memory crossbar 216 can be configured to transmit the output of each cluster 214A-214N to any partition unit 220A-220N or to another cluster 214A-214N, which can perform additional processing operations on the output. Each cluster 214A-214N can communicate with a memory interface 218 via the memory crossbar 216 to read from or write to various external memory devices. In one embodiment having a memory crossbar 216, the memory crossbar 216 has connections to a memory interface 218 for communicating with the I / O unit 204 and to a local instance of parallel processor memory 222, thereby enabling processing units within different processing clusters 214A-214N to communicate with system memory or other memory that is not local to the parallel processing unit 202. In general, the memory crossbar 216 may be capable of separating traffic flows between the clusters 214A-214N and the partition units 220A-220N using virtual channels, for example.

[0065] Although a single instance of parallel processing unit 202 is shown within parallel processor 200, any number of instances of parallel processing unit 202 may be included. For example, multiple instances of parallel processing unit 202 may be provided on a single plug-in card, or multiple plug-in cards may be interconnected. For example, parallel processor 200 may be included in Figure 1 The plug-in device in the (one or more) plug-in device 120 can be a graphics card (such as a discrete graphics card including one or more GPUs, one or more memory devices, and device-to-device or network or fabric interfaces). Different instances of parallel processing unit 202 can be configured to interoperate even if different instances have different numbers of processing cores, different amounts of local parallel processor memory, and / or other configuration differences. Optionally, some instances of parallel processing unit 202 may include higher precision floating point units relative to other instances. Systems including one or more instances of parallel processing unit 202 or parallel processor 200 can be implemented in various configurations and form factors, including but not limited to desktop computers, laptop computers, or handheld personal computers, servers, workstations, game consoles, and / or embedded systems. The orchestrator can use one or more of the following to form a composite node for workload execution: decomposed processor resources, cache resources, memory resources, storage resources, and networking resources.

[0066] In one embodiment, the parallel processing unit 202 can be partitioned into multiple instances. Those multiple instances can be configured to execute workloads associated with different clients in an isolated manner, thereby providing a predetermined quality of service for each client. For example, each cluster 214A-214N can be partitioned and isolated from other clusters, allowing the processing cluster array 212 to be divided into multiple computing partitions or instances. In such a configuration, workloads executed on isolated partitions are protected from errors or errors associated with different workloads executed on different partitions. Partition units 220A-220N can be configured to enable dedicated and / or isolated paths to the memory of the cluster 214A-214N associated with the corresponding computing partition. This data path isolation allows computing resources within a partition to communicate with one or more assigned memory units 224A-224N without being interfered with by the activities of other partitions. In one embodiment, data path isolation can be enhanced by encrypting data in the memory of various partitions with encryption keys unique to the associated partition, so that data is protected on a per-partition basis both at rest and during transition.

[0067] Figure 2B is a block diagram of the partition unit 220. The partition unit 220 may be Figure 2A20N。 As shown, the partition unit 220 includes an L2 cache 221, a frame buffer interface 225, and a ROP 226 (raster operation unit). The L2 cache 221 is a read / write cache configured to perform load and store operations received from the memory crossbar switch 216 and the ROP 226. Read misses and urgent writeback requests are output by the L2 cache 221 to the frame buffer interface 225 for processing. Updates can also be sent to the frame buffer via the frame buffer interface 225 for processing. In one embodiment, the frame buffer interface 225 is connected to the frame buffer interface 225. Figure 2A Partition unit 220 may also interface with one of the memory units in the parallel processor memory 222 via a memory controller (not shown).

[0068] In graphics applications, ROP 226 is a processing unit that performs raster operations such as stenciling, z-testing, blending, and the like. ROP 226 then outputs processed graphics data, which is stored in graphics memory. In some embodiments, ROP 226 includes or is coupled to a codec (CODEC) 227 that includes compression logic for compressing depth or color data written to memory or L2 cache 221 and decompressing depth or color data read from memory or L2 cache 221. The compression logic may be lossless compression logic that utilizes one or more of a variety of compression algorithms. The type of compression performed by CODEC 227 may vary based on the statistical characteristics of the data being compressed. For example, in one embodiment, delta color compression is performed on the depth and color data on a per-slice basis. In one embodiment, CODEC 227 includes compression and decompression logic that can compress and decompress the computational data associated with machine learning operations. CODEC 227 can, for example, compress sparse matrix data for sparse machine learning operations. CODEC 227 can also compress sparse matrix data encoded in sparse matrix format (e.g., coordinate list encoding (COO), compressed sparse row (CSR), compressed sparse column (CSC), etc.) to generate compressed and encoded sparse matrix data. Compressed and encoded sparse matrix data can be decompressed and / or decoded before being processed by a processing element, or a processing element can be configured to consume compressed, encoded, or compressed and encoded data for processing. In one embodiment, CODEC 227 can be configured as a general-purpose data compression engine for use in GPU database acceleration and large-capacity data analysis.

[0069] ROP 226 may be included in each processing cluster (e.g., Figure 2A 214N) rather than being included within partition unit 220. In such embodiments, read and write requests for pixel data rather than pixel fragment data are routed through memory crossbar 216. The processed graphics data may be displayed on a display device such as a Figure 1 10B), is routed for further processing by the processor(s) 102, or is routed for processing by the processor(s) 102. Figure 2A The processing is further processed by one of the processing entities within the parallel processor 200.

[0070] Figure 2C is a block diagram of a processing cluster 214 within a parallel processing unit. For example, processing cluster 214 represents Figure 2A An instance of a processing cluster in one of the processing clusters 214A-214N. The processing cluster 214 can be configured to execute many threads in parallel, where the term "thread" refers to an instance of a specific program executed on a specific set of input data. Alternatively, a single-instruction, multiple-data (SIMD) instruction issuance technology can be used to support the parallel execution of a large number of threads without providing multiple independent instruction units. Alternatively, a single-instruction, multiple-thread (SIMT) technology can be used to use a common instruction unit configured to issue instructions to a set of processing engines within each processing cluster in the processing cluster to support the parallel execution of a large number of generally synchronized threads. Unlike the SIMD execution mechanism in which all processing engines typically execute the same instruction, SIMT execution allows different threads to more easily follow divergent execution paths through a given thread program. Those skilled in the art will understand that the SIMD processing mechanism represents a functional subset of the SIMT processing mechanism.

[0071] The operation of the processing cluster 214 may be controlled via a pipeline manager 232 that distributes processing tasks to SIMT parallel processors. The pipeline manager 232 receives data from Figure 2A The graphics multiprocessor 234 receives instructions from the scheduler 210 and manages the execution of those instructions via the graphics multiprocessor 234 and / or the texture unit 236. The graphics multiprocessor 234 is an exemplary instance of a SIMT parallel processor. However, various types of SIMT parallel processors with different architectures may be included in the processing cluster 214. One or more instances of the graphics multiprocessor 234 may be included in the processing cluster 214. The graphics multiprocessor 234 may also be referred to as a streaming multiprocessor (SM) and is capable of executing a large number of execution threads simultaneously.

[0072] The graphics multiprocessor 234 can process data, and the data crossbar 240 can be used to distribute the processed data to one of multiple possible destinations, including instances of the graphics multiprocessor 234 within the processing cluster 214. The pipeline manager 232 can facilitate the distribution of processed data by specifying the destination for the processed data to be distributed via the data crossbar 240. Each graphics multiprocessor 234 within the processing cluster 214 can include the same set of function execution logic (e.g., arithmetic logic unit, load-store unit, etc.). The function execution logic can be configured in a pipelined manner, in which new instructions can be issued before previous instructions are completed. The function execution logic supports a variety of operations, including integer and floating-point arithmetic, comparison operations, Boolean operations, shifts, and the calculation of various algebraic functions. The same functional unit hardware can be used to perform different operations, and any combination of functional units can be present.

[0073] Instructions transmitted to the processing cluster 214 constitute threads. A collection of threads executed across a collection of parallel processing engines is a thread group. Thread groups execute the same program on different input data. Each thread within a thread group can be assigned to a different processing engine within the graphics multiprocessor 234. A thread group may include fewer threads than the number of processing engines within the graphics multiprocessor 234. When a thread group includes fewer threads than the number of processing engines, one or more of the processing engines may be idle during the cycles in which the thread group is being processed. A thread group may also include more threads than the number of processing engines within the graphics multiprocessor 234. When a thread group includes more threads than the number of processing engines within the graphics multiprocessor 234, processing may be performed in consecutive clock cycles. Optionally, multiple thread groups may be executed concurrently on the graphics multiprocessor 234.

[0074] The graphics multiprocessor 234 may include internal cache memory to perform load and store operations. Alternatively, the graphics multiprocessor 234 may forgo the internal cache and use cache memory within the processing cluster 214 (e.g., level 1 (L1) cache 248). Each graphics multiprocessor 234 also has a partition unit (e.g., Figure 2A20N) are shared among all instances of processing cluster 214 and can be used to transfer data between threads. Graphics multiprocessor 234 can also access off-chip global memory, which can include one or more of local parallel processor memory and / or system memory. Any memory external to parallel processing unit 202 can be used as global memory. In embodiments where processing cluster 214 includes multiple instances of graphics multiprocessor 234, common instructions and data can be shared, which can be stored in L1 cache 248.

[0075] Each processing cluster 214 may include an MMU 245 (memory management unit) configured to map virtual addresses to physical addresses. In other embodiments, one or more instances of the MMU 245 may reside in Figure 2A The MMU 245 includes a set of page table entries (PTEs) for mapping virtual addresses to physical addresses of the slice, and optionally includes a cache line index. The MMU 245 may include an address translation lookaside buffer (TLB) or cache that may reside within the graphics multiprocessor 234 or L1 cache 248 of the processing cluster 214. The physical address is processed to distribute surface data access locality, thereby allowing efficient request interleaving between partition units. The cache line index can be used to determine whether a request for a cache line is a hit or a miss.

[0076] In graphics and compute applications, the processing clusters 214 may be configured such that each graphics multiprocessor 234 is coupled to a texture unit 236 for performing texture mapping operations, such as determining texture sample locations, reading texture data, and filtering texture data. Texture data is read from an internal texture L1 cache (not shown), or in some embodiments, from an L1 cache within the graphics multiprocessor 234 and retrieved from an L2 cache, local parallel processor memory, or system memory as needed. Each graphics multiprocessor 234 outputs processed tasks to a data crossbar 240 to provide the processed tasks to another processing cluster 214 for further processing, or stores the processed tasks in an L2 cache, local parallel processor memory, or system memory via a memory crossbar 216. A pre-raster operations unit (pre-ROP) 242 is configured to receive data from the graphics multiprocessor 234 and direct the data to ROP units that may interface with partition units (e.g., Figure 2AThe preROP 242 unit may perform optimizations for color blending, organize pixel color data, and perform address translation.

[0077] It will be appreciated that the core architecture described herein is illustrative and that variations and modifications are possible. Any number of processing units (e.g., graphics multiprocessor 234, texture unit 236, preROP 242, etc.) may be included within processing cluster 214. Parallel processing units as described herein may include any number of instances of processing cluster 214. Optionally, each processing cluster 214 may be configured to operate independently of other instances of processing cluster 214 using, for example, separate and distinct processing units, L1 caches, L2 caches, to facilitate data and fault isolation.

[0078] Figure 2D An example of a graphics multiprocessor 234 is shown, wherein the graphics multiprocessor 234 is coupled to the pipeline manager 232 of the processing cluster 214. The graphics multiprocessor 234 has an execution pipeline that includes, but is not limited to, an instruction cache 252, an instruction unit 254, an address mapping unit 256, a register file 258, one or more general-purpose graphics processing unit cores (GPGPU cores 262), and one or more load / store units 266. The GPGPU cores 262 and the load / store units 266 are coupled to a cache memory 272 and a shared memory 270 via a memory and cache interconnect 268. The graphics multiprocessor 234 may also include a ray tracing core 263, which includes hardware logic for accelerating ray tracing operations, and a tensor core 264, which includes hardware logic for accelerating tensor (e.g., matrix) operations. The instruction cache 252 may receive a stream of instructions to be executed from the pipeline manager 232. Instructions are cached in the instruction cache 252 and dispatched for execution by the instruction unit 254. The instruction unit 254 can dispatch instructions as groups of threads (e.g., warps), where each thread in the thread group is assigned to a different execution unit within the GPGPU core 262. Instructions can access any of the local address space, the shared address space, or the global address space by specifying an address within the unified address space. The address mapping unit 256 can be used to translate the address in the unified address space into different memory addresses that can be accessed by the load / store unit 266.

[0079] The register file 258 provides a collection of registers for the functional units of the graphics multiprocessor 234. The register file 258 provides temporary storage for operands for the data paths of the functional units (e.g., GPGPU core 262, load / store unit 266) connected to the graphics multiprocessor 234. The register file 258 may be divided among each of the functional units so that each functional unit is allocated a dedicated portion of the register file 258. For example, the register file 258 may be divided among different groups of units executed by the graphics multiprocessor 234.

[0080] The GPGPU cores 262 may each include a floating point unit (FPU) and / or an integer arithmetic logic unit (ALU) for executing instructions of the graphics multiprocessor 234. In some implementations, the GPGPU cores 262 may include hardware logic that may otherwise reside within the tensor core 264 and / or the ray tracing core 263. The GPGPU cores 262 may be architecturally similar or architecturally different. For example, and in one embodiment, a first portion of the GPGPU core 262 includes a single-precision FPU and integer ALU, while a second portion of the GPGPU core includes a double-precision FPU. Optionally, the FPU may implement the IEEE 754-2008 standard for floating-point arithmetic or be capable of variable-precision floating-point arithmetic. In one embodiment, a single-precision FPU or a separate set of FPUs may be configured to perform operations on 16-bit floating-point operands, such as operands in half-precision format or bfloat16 format (e.g., Brain floating point), which is a 16-bit floating-point format with one sign bit, eight exponent bits, and eight significant bits, seven of which are explicitly stored. The FPUs within one or more GPGPU cores 262 may also support one or more 8-bit floating-point formats. Supported 8-bit floating-point formats include the E4M3 format with a 4-bit exponent and a 3-bit mantissa and the E5M2 format with a 5-bit exponent and a 2-bit mantissa. The graphics multiprocessor 234 may additionally include one or more fixed-function or special-function units for performing specific functions, such as copying rectangles or pixel blending operations. One or more of the GPGPU cores may also include fixed-function or special-function logic.

[0081] GPGPU core 262 may include SIMD logic capable of executing a single instruction on multiple sets of data. Optionally, GPGPU core 262 can physically execute SIMD4, SIMD8 and SIMD16 instructions, and logically execute SIMD1, SIMD2 and SIMD32 instructions. SIMD instructions for GPGPU core can be generated by a shader compiler at compile time, or automatically generated when executing a program written and compiled for a single program multiple data (SPMD) or SIMT architecture. Multiple threads of a program configured for a SIMT execution model can be executed via a single SIMD instruction. For example, and in one embodiment, eight SIMT threads performing the same or similar operation can be executed in parallel as SIMD8 instructions. In one embodiment, a cell group of 32 SIMT threads can be executed as a single SIMD32 instruction. Cell group divergence can be handled via multiple SIMD instructions.

[0082] The memory and cache interconnect 268 is an interconnect network that connects each of the functional units in the graphics multiprocessor 234 to the register file 258 and to the shared memory 270. For example, the memory and cache interconnect 268 is a crossbar interconnect that allows the load / store unit 266 to perform load and store operations between the shared memory 270 and the register file 258. The register file 258 can operate at the same frequency as the GPGPU core 262, so data transfers between the GPGPU core 262 and the register file 258 are very low-latency. Shared memory 270 can be used to enable communication between threads executing on the functional units within the graphics multiprocessor 234. Shared memory 270 can also be used as a managed cached program. Cache memory 272 can be used as an automatically managed data cache, for example, to cache texture data communicated between the functional units and the texture unit 236. Shared memory 270 and cache memory 272 can be coupled to the data crossbar 240 to enable communication with other components of the processing cluster, thereby facilitating the collaborative execution of cluster workgroups by multiple graphics multiprocessors within the processing cluster. Threads executing on GPGPU core 262 can programmatically store data in shared memory in addition to automatically cached data stored in cache memory 272. In one embodiment, shared memory 270 and cache memory 272 can be combined into a single configurable memory unit that can be selectively configured as either cache memory 272 or shared memory 270.

[0083] Figure 2E Shown with respect to Figure 2DThe graphics multiprocessor 235 is an alternative configuration of the graphics multiprocessor 234. The disclosure of any feature described herein in conjunction with the graphics multiprocessor 235 also discloses the corresponding Figure 2D The graphics multiprocessor 234 may be combined with, but is not limited to, the graphics multiprocessor 234. Figure 2E The graphics multiprocessor 235 includes relative Figure 2D The graphics multiprocessor 235 may include multiple additional instances of execution resources 286A-286D of the graphics multiprocessor 234. For example, the graphics multiprocessor 235 may include multiple instruction units 254A-254D, register files 258A-258D, and texture units 280A-280D. The graphics multiprocessor 235 also includes multiple sets of graphics or compute execution units (e.g., GPGPU cores 262A-262D, ray tracing cores 263A-263D, tensor cores 264A-264D) and multiple sets of load / store units 266A-266D. The execution resources 286A-286D may work in conjunction with the texture unit(s) 280A-280D for texture operations while sharing the instruction cache 252, shared memory 270, and cache memories 272A-272B. In one embodiment, execution resources 286A-286D additionally include multi-function units (MUFUs) and / or special function units (SFUs) (e.g., MFUs 267A-267D) for performing specialized mathematical operations such as transcendental operations including exponential, logarithmic, and trigonometric functions.

[0084] The components can communicate via interconnect fabric 290. Interconnect fabric 290 may include one or more crossbar switches to enable communication between the components of graphics multiprocessor 235. GPGPU cores 262A-262D, ray tracing cores 263A-263B, and tensor cores 264A-264D can each communicate with shared memory 270 via interconnect fabric 290. Interconnect fabric 290 can arbitrate communications within graphics multiprocessor 235 to ensure fair bandwidth allocation between components. In one embodiment, interconnect fabric 290 can be a separate, high-speed network fabric layer upon which each component of graphics multiprocessor 235 is stacked. Components of graphics multiprocessor 235 can also communicate with remote components via interconnect fabric 290.

[0085] In one embodiment, the graphics multiprocessor 235 includes a tensor transfer engine 292, which is a copy engine configurable to accelerate the movement of tensor data into and out of the graphics multiprocessor 235. The tensor transfer engine 292 can accelerate tensor memory operations by asynchronously performing address generation and data movement operations for N-dimensional blocks of tensor data, which offloads operations that would otherwise be manually performed by program code executed by the graphics multiprocessor 235. The tensor transfer engine 292 can be configured to copy data between, for example, the shared memory 270 and / or cache memories 272A-272B of the graphics multiprocessor 235 and memory external to the graphics multiprocessor 235, such as graphics processor global memory (e.g., parallel processor memory 222). In one embodiment, data transfers performed by the tensor transfer engine 292 can be configured to selectively bypass various levels of intermediate data storage between source and destination memory. For example, transfers between global memory and shared memory 270 can bypass register files 258A-258D. In one embodiment, threads can be synchronized on asynchronous tensor transfers via a non-blocking barrier synchronization mechanism.

[0086] In various embodiments, the graphics multiprocessor 235 can be customized for specific use cases by including or excluding certain components, thereby allowing various implementations of the graphics multiprocessor 235 to be customized for target power, performance, and area characteristics. For example, a compute-oriented variant of the graphics multiprocessor 235 that will not perform graphics operations can exclude ray tracing cores 263A-263D. A fully graphics-oriented variant can exclude the tensor transfer engine 292, while a graphics-oriented variant that is additionally configured to accelerate neural network inference can include at least one version of the tensor transfer engine 292.

[0087] Those skilled in the art will understand that Figure 1 and Figure 2A-2E The architecture described in is illustrative and not limiting with respect to the scope of the present embodiments. Thus, the techniques described herein may be implemented on any appropriately configured processing unit without departing from the scope of the embodiments described herein, including but not limited to: one or more mobile application processors; one or more desktop or server central processing units (CPUs), including multi-core CPUs; one or more parallel processor units, such as, Figure 2A and a parallel processing unit 202 and one or more graphics processors or special purpose processing units.

[0088] The parallel processor or GPGPU described herein is communicatively coupled to a host / processor core to accelerate graphics operations, machine learning operations, pattern analysis operations, and various general-purpose GPU (GPGPU) functions. The GPU can be communicatively coupled to the host processor / core via a bus or another interconnect (e.g., a high-speed interconnect such as PCIe, NVLink, or other known, standardized, or proprietary protocols). In other embodiments, the GPU can be integrated on the same package or chip as the core and communicatively coupled to the core via an internal processor bus / interconnect (i.e., inside the package or chip). Regardless of the manner in which the GPU is connected, the processor core can assign work to the GPU in the form of a sequence of commands / instructions contained in a work descriptor. The GPU then uses dedicated circuit modules / logic to efficiently process these commands / instructions.

[0089] Figure 3 A graphics processing unit (GPU) 380 is shown, which includes a collection of dedicated graphics processing resources arranged into multiple core groups 365A-365N. Multiple core groups 365A-365N correspond to Figure 2D Graphics multiprocessor 234 or Figure 2E While details are provided for a single example of multi-core groups 365A-365N (e.g., multi-core group 365A), it will be appreciated that other multi-core groups 365B-365N may be equipped with the same or similar sets of graphics processing resources. Details described with respect to multi-core groups 365A-365N may also apply to graphics multiprocessors 234 or 235, as described herein.

[0090] As shown, multi-core group 365A may include a graphics core 370, a tensor core 371, and a ray tracing core 372. Graphics core 370 is similar to GPGPU cores 262A-262D and is configurable to execute instructions to perform graphics and / or general computing operations. Scheduler / dispatcher 368 schedules and dispatches graphics threads for execution on various cores within multi-core group 365A. Register files 369 are included, which store operand values ​​used by the cores when executing graphics or general computing operations on the threads being executed. These register files 369 may include, for example, registers configurable to store integer values ​​or floating-point values, including vector registers for storing packed integer and / or floating-point data elements and slice registers for storing tensor / matrix values. Slice registers may be implemented as multidimensional registers comprising a combined set of vector registers.

[0091] One or more combined first level (L1) caches and shared memory units 373 store graphics data locally within each multi-core group 365A, such as texture data, vertex data, pixel data, ray data, bounding volume data, etc. One or more texture units 374 can also be used to perform texture operations, such as texture mapping and sampling. A second level (L2) cache 375 shared by all multi-core groups 365A-365N or a subset of multi-core groups 365A-365N stores graphics data and / or instructions for multiple concurrent graphics threads. As shown, the L2 cache 375 can be shared across multiple multi-core groups 365A-365N. One or more memory controllers 367 couple the GPU 380 to memory 366, which can be system memory (e.g., DRAM) and / or dedicated graphics memory (e.g., GDDR6 memory).

[0092] An input / output circuit module (I / O circuit module 363) couples GPU 380 to one or more I / O devices 362, such as digital signal processors (DSPs), network controllers, or user input devices. On-chip interconnects may be used to couple I / O devices 362 to GPU 380 and memory 366. At least one I / O memory management unit (IOMMU 364) of I / O circuit module 363 couples I / O devices 362 directly to memory 366. Optionally, IOMMU 364 manages multiple sets of page tables used to map virtual addresses to physical addresses in memory 366. I / O devices 362, CPU(s) 361, and GPU 380 can then share the same virtual address space.

[0093] In one implementation of IOMMU 364, IOMMU 364 supports virtualization. In this case, it can manage a first set of page tables for mapping guest / graphics virtual addresses to guest / graphics physical addresses and a second set of page tables for mapping guest / graphics physical addresses to system / host physical addresses (e.g., within memory 366). The base address of each of the first set of page tables and the second set of page tables can be stored in a control register and swapped out upon context switch (e.g., so that the new context is provided with access to the relevant set of page tables). Although not described in Figure 3 , but each of the cores within the multi-core group 365A-365N may include a translation lookaside buffer (TLB) for caching guest virtual to guest physical translations, guest physical to host physical translations, and guest virtual to host physical translations.

[0094] (One or more) CPU 361, GPU 380 and I / O device 362 can be integrated on a single semiconductor chip and / or chip package. Memory 366 can be integrated on the same chip or can be coupled to one or more memory controllers 367 via an off-chip interface. In one implementation, memory 366 includes GDDR6 memory that shares the same virtual address space as other physical system-level memory, but the basic principles described herein are not limited to this particular implementation.

[0095] The tensor core 371 may include multiple execution units specifically designed to perform matrix operations, which are fundamental computational operations for performing deep learning operations. For example, synchronized matrix multiplication operations may be used for neural network training and inference. The tensor core 371 may perform matrix processing using a variety of operand precisions, including single-precision floating point (e.g., 32-bit), half-precision floating point (e.g., 16-bit), integer word (16-bit), byte (8-bit), and half-byte (4-bit). For example, a neural network implementation extracts features from each rendered scene, potentially combining details from multiple frames to construct a high-quality final image.

[0096] In deep learning implementations, parallel matrix multiplication work can be scheduled for execution on Tensor Core 371. Training neural networks, in particular, requires a large number of matrix dot product operations. To handle the inner product formulation of NxNxN matrix multiplication, Tensor Core 371 may include at least N dot product processing elements. Before the matrix multiplication begins, a complete matrix is ​​loaded into the slice registers, and for each of N cycles, at least one column of the second matrix is ​​loaded. For each cycle, there are N dot products processed.

[0097] Depending on the specific implementation, matrix elements can be stored with different precisions, including 16-bit words, 8-bit bytes (e.g., INT8), and 4-bit nibbles (e.g., INT4). Different precision modes can be specified for the tensor core 371 to ensure that the most efficient precision is used for different workloads (e.g., such as inference workloads, which can tolerate quantization to bytes and nibbles). Supported formats also include 64-bit floating point (FP64) and non-IEEE floating point formats, such as bfloat16 format. One embodiment includes support for a reduced precision tensor floating point (TF32) mode that performs calculations using the range of FP32 (8 bits) and the precision of FP16 (10 bits). Reduced precision TF32 operations can be performed on FP32 inputs and produce FP32 outputs with higher performance relative to FP32 and increased precision relative to FP16. In one embodiment, one or more of 8-bit floating point (FP8), 6-bit floating point (FP6), and 4-bit floating point (FP4) formats are supported, including a floating point format denoted as the microscale (MX) format.

[0098] In one embodiment, the tensor core 371 supports a sparse operation mode for matrices in which the vast majority of values ​​are zero. The tensor core 371 includes support for sparse input matrices encoded in a sparse matrix representation (e.g., coordinate list encoding (COO), compressed sparse row (CSR), compressed sparse column (CSC), etc.). The tensor core 371 also includes support for compressed sparse matrix representations in cases where the sparse matrix representation can be further compressed. Compressed matrix data, encoded matrix data, and / or compressed and encoded matrix data, along with associated compression and / or encoding metadata, can be read by the tensor core 371, and non-zero values ​​can be extracted. For example, for a given input matrix A, non-zero values ​​can be loaded from a compressed and / or encoded representation of at least a portion of matrix A. Based on the position of the non-zero values ​​in matrix A (which can be determined from the index or coordinate metadata associated with the non-zero values), the corresponding values ​​in the input matrix B can be loaded. Depending on the operation to be performed (e.g., multiplication), loading values ​​from input matrix B may be bypassed if the corresponding value is a zero value. In one embodiment, the pairing of values ​​for certain operations (such as multiplication) may be pre-scanned by the scheduler logic, and only operations between non-zero inputs may be scheduled. Depending on the dimensions of matrices A and B and the operation to be performed, the output matrix C may be dense or sparse. In the case where the output matrix C is sparse and depending on the configuration of the tensor core 371, the output matrix C may be output in a compressed format, sparsely coded, or compressed sparsely coded.

[0099] Ray tracing core 372 may accelerate ray tracing operations for both real-time and non-real-time ray tracing implementations. Specifically, ray tracing core 372 may include a ray traversal / intersection circuit module for performing ray traversals and identifying intersections between rays and primitives enclosed within a bounding volume hierarchy (BVH) using a bounding volume hierarchy. Ray tracing core 372 may also include a circuit module for performing depth testing and culling (e.g., using a Z-buffer or similar arrangement). In one implementation, ray tracing core 372 performs traversal and intersection operations in conjunction with the image denoising techniques described herein, at least portions of which may be executed on tensor core 371. For example, tensor core 371 may implement a deep learning neural network to perform denoising on frames generated by ray tracing core 372. However, the CPU(s) 361 , graphics core 370 , and / or ray tracing core 372 may also implement all or portions of the denoising and / or deep learning algorithms.

[0100] Furthermore, as described above, a distributed approach to noise reduction can be employed, wherein GPU 380 is located in a computing device coupled to other computing devices via a network or high-speed interconnect. According to this distributed approach, the interconnected computing devices can share neural network learning / training data to improve the speed at which the entire system learns to perform noise reduction for different types of image frames and / or different graphics applications.

[0101] The ray tracing core 372 can handle all BVH traversals and / or ray-primitive intersections, thereby freeing the graphics core 370 from being overloaded with thousands of instructions for each ray. For example, each ray tracing core 372 includes a first set of specialized circuit modules for performing bounding box tests (e.g., for traversal operations) and / or a second set of specialized circuit modules for performing ray-triangle intersection tests (e.g., intersecting rays that have already been traversed). Thus, for example, the multi-core group 365A can simply start ray probing, and the ray tracing core 372 independently performs ray traversals and intersections and returns hit data (e.g., hit, no hit, multiple hits, etc.) to the thread context. The graphics core 370 and tensor core 371 are then freed to perform other graphics or compute work while the ray tracing core 372 performs traversal and intersection operations. Optionally, each ray tracing core 372 can include a traversal unit for performing BVH test operations and / or an intersection unit for performing ray-primitive intersection tests. The intersection unit generates "hit," "no hit," or "multiple hits" responses, which it provides to the appropriate threads. During the traversal and intersection operations, execution resources of other cores (e.g., graphics core 370 and tensor core 371) are freed up to perform other forms of graphics work. In an alternative embodiment described below, a hybrid rasterization / ray tracing approach is used in which rendering operations are distributed between graphics core 370 and ray tracing core 372.

[0102] The ray tracing core 372 may include hardware support for a ray tracing instruction set, such as Microsoft's DirectX Ray Tracing (DXR), which includes the DispatchRays command; and ray generation shaders, nearest hit shaders, any hit shaders, and miss shaders, which enable the assignment of a unique set of shaders and textures to each object. Another ray tracing platform that may be supported by the ray tracing core 372, graphics core 370, and tensor core 371 is the Vulkan API (e.g., Vulkan version 1.1.85, or later). However, it should be noted that the basic principles described herein are not limited to any particular ray tracing ISA. In general, the ray tracing core 372, tensor core 371, and graphics core 370 may support a ray tracing instruction set that includes instructions / functions for one or more of the following: ray generation, nearest hit, any hit, ray-primitive intersection, primitive-by-primitive and hierarchy bounding box construction, miss, access, and exceptions. More specifically, preferred embodiments include ray tracing instructions for performing one or more of the following functions:

[0103] Light Generation- Ray generation instructions can be executed for each pixel, sample, or other user-defined work assignment.

[0104] Recent Hits - Can execute nearest hit instructions to locate the closest intersection of a ray with a primitive within the scene.

[0105] Any hit - Any hit instruction identifies multiple intersections between rays and primitives within the scene, potentially identifying a new closest intersection point.

[0106] intersect - The Intersect instruction performs a ray-primitive intersection test and outputs the result.

[0107] Per-primitive bounding box construction - This instruction builds a bounding box around a given primitive or group of primitives (e.g., when building a new BVH or other acceleration data structure).

[0108] miss - Indicates that the ray missed the scene or all geometry within the specified area of ​​the scene.

[0109] access ——Indicates the subvolume that the ray will traverse.

[0110] abnormal - Includes various types of exception handlers (e.g., called for various error conditions).

[0111] In one embodiment, the ray tracing core 372 may be adapted to accelerate general computational operations using computational techniques similar to ray intersection testing. A computational framework may be provided that enables shader programs to be compiled into low-level instructions and / or primitives that perform general computational operations via the ray tracing core. Exemplary computational problems that may benefit from computational operations performed on the ray tracing core 372 include computations involving the propagation of beams, waves, rays, or particles within a coordinate space. Interactions associated with that propagation may be computed relative to geometry or meshes within the coordinate space. For example, computations associated with the propagation of electromagnetic signals through an environment may be accelerated using instructions or primitives that are executed via the ray tracing core. Refraction and reflection of signals through objects in the environment may be computed as direct ray tracing simulations.

[0112] The ray tracing core 372 can also be used to perform calculations that are not directly similar to ray tracing. For example, the ray tracing core 372 can be used to accelerate mesh projection, mesh refinement, and volume sampling calculations. General coordinate space calculations, such as nearest neighbor calculations, can also be performed. For example, a set of points near a given point can be found by defining a bounding box around the point in coordinate space. The BVH and ray detection logic within the ray tracing core 372 can then be used to determine the set of intersections of points within the bounding box. The intersections constitute the origin and the nearest neighbors of that origin. The calculations performed using the ray tracing core 370 can be performed in parallel with the calculations performed on the graphics core 372 and the tensor core 371. The shader compiler can be configured to compile compute shaders or other general graphics processing programs into low-level primitives that can be parallelized across the graphics core 370, the tensor core 371, and the ray tracing core 372. Technologies for GPU to host processor interconnection

[0113] Figure 4A In FIG. 4 , a plurality of GPUs 410-413 (e.g., such as Figure 2A ) is communicatively coupled to a plurality of multi-core processors 405-406 via high-speed links 440A-440D (e.g., buses, point-to-point interconnects, etc.). Depending on the implementation, the high-speed links 440A-440D may support 4 GB / s, 30 GB / s, 80 GB / s, or higher communication throughput. Various interconnect protocols may be used, including but not limited to PCIe 4.0, PCIe 5.0, PCIe 6.0, and various NVLink and NVLink-C2C (chip-to-chip) interconnect protocols (e.g., NVLink v5). However, the underlying principles described herein are not limited to any particular communication protocol or throughput.

[0114] Two or more of the GPUs 410-413 may be interconnected via high-speed links 442A-442B, which may be implemented using the same or different protocols / links as those used for high-speed links 440A-440D. Similarly, two or more of the multi-core processors 405-406 may be connected via high-speed link 443, which may be a symmetric multi-processor (SMP) bus operating at 20 GB / s, 30 GB / s, 120 GB / s, or lower or higher speeds. Alternatively, Figure 4A All communications between the various system components shown in FIG. 5 can be implemented using the same protocol / links (eg, through a common interconnect structure). However, as mentioned, the basic principles described herein are not limited to any particular type of interconnect technology.

[0115] Each of multi-core processors 405 and 406 may be communicatively coupled to processor memories 401-402 via memory interconnects 430A-430B, respectively, and each GPU 410-413 may be communicatively coupled to GPU memories 420-423 via GPU memory interconnects 450A-450D, respectively. Memory interconnects 430A-430B and 450A-450D may utilize the same or different memory access technologies. By way of example and not limitation, processor memories 401-402 and GPU memories 420-423 may be volatile memories such as dynamic random access memory (DRAM) (including stacked DRAM), graphics DDR SDRAM (GDDR) (e.g., GDDR5, GDDR6), or high bandwidth memory (HBM), and / or may be non-volatile memories such as 3D XPoint / Optane or Nano-Ram. For example, a portion of the memory may be volatile memory and another portion may be non-volatile memory (e.g., using a two-level memory (2LM) hierarchy). The memory subsystem as described herein is compatible with several memory technologies, such as the double data rate version published by the Joint Electronic Device Engineering Council (JEDEC).

[0116] As described below, although each processor 405-406 and GPU 410-413 may be physically coupled to a specific processor memory 401-402 and GPU memory 420-423, respectively, a unified memory architecture may be implemented in which the same virtual system address space (also referred to as an "effective address" space) is distributed across all of the various physical memories. For example, processor memories 401-402 may each include 64GB of system memory address space, and GPU memories 420-423 may each include 32GB of system memory address space (yielding a total of 256GB of addressable memory in this example).

[0117] Figure 4BAdditional optional details of the interconnection between processor 407 and graphics accelerator 446 are shown. Graphics accelerator 446 may include one or more GPU chips integrated on a line card coupled to processor 407 via high-speed link 440. Alternatively, graphics accelerator 446 may be integrated on the same package or chip as processor 407. Processor 407 includes multiple cores 460A-460D, each of which has a translation lookaside buffer 461A-461D and one or more caches 462A-462D. The cores may include various other components for executing instructions and processing data, which are not shown to avoid obscuring the basic principles of the components described herein (e.g., instruction fetch unit, branch prediction unit, decoder, execution unit, reorder buffer, etc.). Caches 462A-462D may include a first level (L1) cache and a second level (L2) cache. In addition, one or more shared caches 456 may be included in the cache hierarchy and shared by the collection of cores 460A-460D. For example, one embodiment of processor 407 includes 24 cores, each of which has its own L1 cache, twelve shared L2 caches, and twelve shared L3 caches. In this embodiment, one of the L2 cache and the L3 cache is shared by two adjacent cores. Processor 407 is connected to system memory 441, which may include processor memories 401-402.

[0118] Coherence is maintained for data and instructions stored in each cache 462A-462D, the shared cache(s) 456, and the system memory 441 via inter-core communication over a coherence bus 464. For example, each cache may have associated therewith cache coherence logic / circuitry modules that communicate over the coherence bus 464 in response to a detected read or write to a particular cache line. In one implementation, a cache snooping protocol is implemented over the coherence bus 464 to snoop cache accesses. Cache snooping / coherence techniques are well understood by those skilled in the art and will not be described in detail herein to avoid obscuring the underlying principles described herein. Proxy circuitry 425 may be provided that communicatively couples the graphics accelerator 446 to the coherence bus 464, thereby allowing the graphics accelerator 446 to participate in the cache coherence protocol as a peer of the core. Specifically, an interface 435 provides connectivity to the proxy circuitry 425 over a high-speed link 440 (e.g., a PCIe bus, NVLink, etc.), and an interface 437 connects the graphics accelerator 446 to the high-speed link 440.

[0119] In one implementation, interface 437 is coupled to an accelerator integrated circuit 436 that provides cache management, memory access, context management, and interrupt management services on behalf of graphics processing engines 431, 432, ..., N of graphics accelerator 446. Graphics processing engines 431, 432, ..., N may each comprise a separate graphics processing unit (GPU). Alternatively, graphics processing engines 431, 432, ..., N may comprise different types of graphics processing engines within a GPU, such as a graphics execution unit, a media processing engine (e.g., a video encoder / decoder), a sampler, and a block image transfer (BLIT) engine. In other words, the graphics accelerator may comprise graphics processing engines 431-432, ..., N of a single GPU, or graphics processing engines 431-432, ..., N may be associated with multiple GPUs integrated on a common package, line card, or chip. Graphics processing engines 431-432, ..., N may be configured using any graphics processor or computing accelerator architecture described herein. The work to be performed by graphics processing engines 431 , 432 may be specified via work descriptors that provide an indication of the work to be done by graphics accelerator 446 .

[0120] The accelerator integrated circuit 436 may include a memory management unit (MMU 439) for performing various memory management functions, such as virtual-to-physical memory translation (also known as effective-to-real memory translation) and a memory access protocol for accessing system memory 441. The MMU 439 may also include a translation lookaside buffer (TLB) (not shown) for caching virtual / effective-to-physical / real address translations. In one implementation, cache 438 stores commands and data for efficient access by graphics processing engines 431, 432, ..., N. Data stored in cache 438 and graphics memories 433-434, ..., M may be kept consistent with core caches 462A-462D, shared cache(s) 456, and system memory 441. As mentioned, this may be accomplished via proxy circuitry 425, which participates in cache coherence mechanisms on behalf of cache 438 and graphics memories 433-434, ..., M (e.g., sending updates to cache 438 related to modifications / accesses of cache lines on processor caches 462A-462D, shared cache(s) 456, and receiving updates from cache 438).

[0121] Registers 445 store context data for threads executed by graphics processing engines 431-432, ..., N, and context management circuitry 448 manages these thread contexts. For example, context management circuitry 448 may perform save and restore operations to save and restore the context of each thread during a context switch (e.g., where a first thread is saved and a second thread is restored so that the second thread can be executed by the graphics processing engine). For example, upon a context switch, context management circuitry 448 may store current register values ​​to a designated area in memory (e.g., identified by a context pointer). It may then restore the register values ​​upon returning to that context. Interrupt management circuitry 447 may, for example, receive interrupts from a system device and process the interrupts received from the system device.

[0122] In one implementation, the virtual / effective address from the graphics processing engine is translated into an actual / physical address in the system memory 441 by the MMU 439. Optionally, the accelerator integrated circuit 436 supports multiple (e.g., 4, 8, 16) graphics accelerators 446 and / or other accelerator devices. The graphics accelerator 446 can be dedicated to a single application executed on the processor 407, or can be shared between multiple applications. Optionally, a virtualized graphics execution environment is provided in which the resources of the graphics processing engines 431-432, ..., N are shared with multiple applications, virtual machines (VMs), or containers. Resources can be subdivided into "slices" that are allocated to different VMs and / or applications based on the processing requirements and priorities associated with the VMs and / or applications or based on a predetermined partition profile for the graphics accelerator 446. VM and container can be used interchangeably herein.

[0123] A virtual machine (VM) can be software that runs an operating system and one or more applications. A VM can be defined by specifications, configuration files, virtual disk files, non-volatile random access memory (NVRAM) settings files, and log files, and is backed up by the physical resources of the host computing platform. A VM may include an operating system (OS) or application environment installed on software that mimics dedicated hardware. End users have the same experience on a virtual machine as they would on dedicated hardware. Specialized software called a hypervisor completely emulates the CPU, memory, hard disk, network, and other hardware resources of a PC client or server, enabling virtual machines to share resources. A hypervisor can emulate multiple virtual hardware platforms that are isolated from each other, allowing virtual machines to run on the same underlying physical host. Server, VMware ESXi and other operating systems.

[0124] A container is a package of applications, configurations, and dependencies so that the application runs reliably from one computing environment to another. Containers can share the operating system installed on a server platform and run as isolated processes. A container is a package of software that contains everything the software needs to run, such as system tools, libraries, and settings. Containers are not installed like traditional software programs, which allows them to be isolated from other software and from the operating system itself. The isolated nature of containers provides several benefits. First, the software in the container will run the same way in different environments. For example, a container containing PHP and MySQL can be installed in Computers and The machines will run in exactly the same way on both. Secondly, containers provide increased security because the software will not affect the host operating system. While installed applications may change system settings and modify resources (such as the Windows Registry), containers can only modify settings within the container.

[0125] Thus, the accelerator integrated circuit 436 acts as a bridge to the system for the graphics accelerator 446 and provides address translation and system memory caching services. In one embodiment, to facilitate bridging functionality, the accelerator integrated circuit 436 may also include shared I / O 497 (e.g., PCIe, USB, or other elements) and hardware to enable system control of voltage, clock control, performance, thermal, and security. The shared I / O 497 may utilize separate physical connections or may span the high-speed link 440. In addition, the accelerator integrated circuit 436 may provide virtualization facilities for the host processor to manage virtualization of the graphics processing engine, interrupts, and memory management.

[0126] Because the hardware resources of the graphics processing engines 431-432, ..., N are explicitly mapped to the actual address space seen by the processor 407, any host processor can directly address these resources using effective address values. An optional function of the accelerator integrated circuit 436 is to physically separate the graphics processing engines 431-432, ..., N so that they appear as independent units to the system. In one embodiment, the accelerator integrated circuit 436 includes a security circuit module 444 that enables configurable cryptographic isolation of data associated with each resource slice. Different slices can be associated with different security domains, so that data associated with various security domains is encrypted using different cryptographic keys. In one embodiment, the security domains of the graphics accelerator 446 can be integrated into a trusted execution environment supported by the processor 407. In one embodiment, secure I / O capabilities can be implemented that allow each security domain to be presented as a separate trusted I / O device, which can enable support for trusted DMA and MMIO operations.

[0127] One or more graphics memories 433-434, ..., M may be coupled to each of the graphics processing engines 431-432, ..., N, respectively. The graphics memories 433-434, ..., M store instructions and data processed by each of the graphics processing engines 431-432, ..., N. The graphics memories 433-434, ..., M may be volatile memories such as DRAM (including stacked DRAM), GDDR memories (e.g., GDDR5, GDDR6), or HBM, and / or may be non-volatile memories such as 3D XPoint / Optane, Samsung Z-NAND, or Nano-Ram.

[0128] To reduce the amount of data traffic on high-speed link 440, biasing techniques may be used to ensure that the data stored in graphics memories 433-434, ..., M is the data that will be used most frequently by graphics processing engines 431-432, ..., N and preferably not used (at least not frequently) by cores 460A-460D. Similarly, the biasing mechanism attempts to keep data needed by the cores (and preferably not graphics processing engines 431-432, ..., N) within caches 462A-462D, shared cache(s) 456, and system memory 441.

[0129] In an alternative variation, an accelerator integrated circuit 436 is integrated within the processor 407, and the graphics processing engines 431-432, ..., N communicate with the accelerator integrated circuit 436 via a high-speed link 440, via an interface 437 and an interface 435 (which again may utilize any form of bus, structure, or interface protocol). In this variation, the accelerator integrated circuit 436 performs the same operations as those described above.

[0130] The described embodiments may support different programming models, including a dedicated process programming model (without graphics accelerator virtualization) and a shared programming model (with virtualization). The latter may include programming models controlled by accelerator integrated circuit 436 and programming models controlled by graphics accelerator 446. In the dedicated process model embodiment, graphics processing engines 431, 432, ..., N may be dedicated to a single application or process under a single operating system. A single application may funnel requests from other applications to graphics processing engines 431, 432, ..., N, thereby providing virtualization within a VM / partition. In the dedicated process programming model, graphics processing engines 431, 432, ..., N may be shared by multiple VMs / application partitions. The shared model requires the hypervisor to virtualize graphics processing engines 431-432, ..., N to allow access by each operating system. For a single-partition system without a hypervisor, graphics processing engines 431-432, ..., N are owned by the operating system. In both cases, the operating system can virtualize graphics processing engines 431-432, ..., N to provide access to each process or application. For a shared programming model, graphics accelerator 446 or individual graphics processing engines 431-432, ..., N use process handles to select process elements. Process elements can be stored in system memory 441 and can be addressable using the effective address to real address translation techniques described herein. The process handle can be an implementation-specific value provided to the host process when registering its context with the graphics processing engine 431-432, ..., N. This can be performed by the host process by calling system software to add the process element to a process element linked list. The lower 16 bits of the process handle can be the offset of the process element within the process element linked list.

[0131] Figure 4C An accelerator integrated slice 490 is shown. As used herein, a "slice" comprises a specified portion of the processing resources of an accelerator integrated circuit 436. The application's address space 482 within system memory 441 stores process elements 483. Process elements 483 may be stored in response to a GPU call 481 from an application 480 executing on processor 407. Process elements 483 contain the process state of application 480. The work descriptor (WD 484) contained in process element 483 may be a single job requested by the application, or may contain a pointer to a job queue. In the latter case, WD 484 is a pointer to a job request queue in the application's address space 482.

[0132] The graphics accelerator 446 and / or the individual graphics processing engines 431-432, ..., N can be shared by all processes in the system or a subset of the processes in the system. For example, the techniques described herein may include infrastructure for establishing process state and sending WD 484 to the graphics accelerator 446 to start a job in a virtualized environment.

[0133] In one implementation, the dedicated process programming model is implementation-specific. In this model, a single process owns graphics accelerator 446 or a separate graphics processing engine (e.g., graphics processing engine 431). Because graphics accelerator 446 is owned by a single process, when graphics accelerator 446 is assigned, the hypervisor initializes accelerator integrated circuit 436 for the owning partition, and the operating system initializes accelerator integrated circuit 436 for the owning process.

[0134] In operation, a fetch unit 491 in the accelerator integrated slice 490 fetches a pending WD 484. WD 484 includes an indication of work to be performed by one or more graphics processing engines of the graphics accelerator 446. As shown, data from WD 484 may be stored in registers 445 and used by the MMU 439, interrupt management circuitry 447, and / or context management circuitry 448. For example, the MMU 439 may include a segment / page walk circuit module for accessing segment tables / page tables 486 within the OS virtual address space 485. The interrupt management circuitry 447 may process interrupt events 492 received from the graphics accelerator 446. When executing graphics operations, effective addresses 493 generated by the graphics processing engines 431-432, ..., N are translated into real addresses by the MMU 439.

[0135] Registers 445 may be replicated for each graphics processing engine 431-432, ..., N and / or graphics accelerator 446 and may be initialized by a hypervisor or operating system. Each of these replicated registers may be included in an accelerator integrated slice 490. In one embodiment, each graphics processing engine 431-432, ..., N may be presented to a hypervisor 496 as a different graphics processor device. Quality of service (QoS) settings may be configured for clients of a particular graphics processing engine 431-432, ..., N. Encryption and physical data isolation between clients of each engine may be achieved via isolated memory access paths and automatic data encryption (for each client). Exemplary registers that may be initialized by a hypervisor are shown in Table 1. Table 1 - Registers initialized by the hypervisor 1 Slice Control Register 2 Real Address (RA) dispatched process area pointer 3 Permission Mask Override Register 4 Interrupt vector table entry offset 5 Interrupt vector table entry limit 6 Status Register 7 Logical partition ID 8 Real Address (RA) Manager Accelerator Utilizes Record Pointers 9 Storage Description Register

[0136] Example registers that may be initialized by the operating system are shown in Table 2. Table 2 - Registers initialized by the operating system

[0137] Each WD 484 may be specific to a particular graphics accelerator 446 and / or graphics processing engine 431-432, ..., N. It contains all the information required by the graphics processing engine to do its work, or it may be a pointer to a memory location where a command queue is located where an application has set up work to be done.

[0138] Figure 4D Shows additional optional details of the shared model. It includes a list of process elements 499 is stored in the hypervisor real address space 498. Hypervisor real address space 498 is accessible via a hypervisor 496 , which virtualizes the graphics processing engine to the operating system 495 .

[0139] The shared programming model allows all processes or a subset of processes from all partitions in the system or a subset of partitions in the system to use the graphics accelerator 446. There are two programming models in which the graphics accelerator 446 is shared by multiple processes and partitions: time-sharing and graphics-directed sharing.

[0140] In this model, hypervisor 496 owns graphics accelerator 446 and makes its functionality available to all operating systems 495. In order for graphics accelerator 446 to support virtualization by hypervisor 496, graphics accelerator 446 may adhere to the following requirements: 1) Application job requests must be autonomous (i.e., no state needs to be maintained between jobs), or graphics accelerator 446 must provide a context save and restore mechanism. 2) Graphics accelerator 446 guarantees that application job requests are completed within a specified amount of time, including any translation errors, or graphics accelerator 446 provides the ability to preempt the processing of jobs. 3) Graphics accelerator 446 must ensure fairness between processes when operating in a directed sharing programming model.

[0141] For the directed sharing model, the application 480 may be required to make an operating system 495 system call using a graphics accelerator 446 type, a work descriptor (WD), an authority mask register (AMR) value, and a context save / restore area pointer (CSRP). The graphics accelerator 446 type describes the target acceleration function for the system call. The graphics accelerator 446 type can be a system-specific value. The WD is formatted specifically for the graphics accelerator 446 and can take the form of a graphics accelerator 446 command, an effective address pointer to a user-defined structure, an effective address pointer to a command queue, or any other data structure describing the work to be performed by the graphics accelerator 446. In one embodiment, the AMR value is the AMR state to be used for the current process. The value passed to the operating system is similar to the application setting the AMR. If the accelerator integrated circuit 436 and graphics accelerator 446 implementation do not support the User Authority Mask Override Register (UAMOR), the operating system may apply the current UAMOR value to the AMR value before passing the AMR in the hypervisor call. The hypervisor 496 may optionally apply the current Authority Mask Override Register (AMOR) value before placing the AMR into the process element 483. The CSRP may be one of the registers 445 that contains the effective address of an area in the application's address space 482 for the graphics accelerator 446 to use to save and restore context state. This pointer is optional if no state is required to be saved between jobs or when a job is preempted. The context save / restore area may be pinned system memory.

[0142] Upon receiving the system call, operating system 495 may verify that application 480 is registered and has been given permission to use graphics accelerator 446. Operating system 495 then calls hypervisor 496 with the information shown in Table 3. Table 3 - OS call parameters to the hypervisor

[0143] Upon receiving the hypervisor call, the hypervisor 496 verifies that the operating system 495 has registered and been given permission to use the graphics accelerator 446. The hypervisor 496 then places the process element 483 in the process element linked list for the corresponding type of graphics accelerator 446. The process element may include the information shown in Table 4. Table 4 - Process element information 1 Work Descriptor (WD) 2 The authority mask register (AMR) value (potentially masked). 3 Effective Address (EA) Context Save / Restore Region Pointer (CSRP) 4 Process ID (PID) and optional thread ID (TID) 5 Virtual Address (VA) Accelerator Utilization Record Pointer (AURP) 6 Virtual address of the storage segment table pointer (SSTP) 7 Logical Interrupt Service Number (LISN) 8 Interrupt vector table, derived from the hypervisor call parameters. 9 Status register (SR) value 10 Logical partition ID (LPID) 11 Real Address (RA) Manager Accelerator Utilizes Record Pointers 12 Storage Descriptor Register (SDR)

[0144] The hypervisor may initialize the registers 445 of the accelerator integrated slice 490 .

[0145] like Figure 4E As shown in , in an optional implementation, a unified memory addressable via a common virtual memory address space is employed, which is used to access physical processor memories 401-402 and GPU memories 420-423. In this implementation, operations executed on GPUs 410-413 utilize the same virtual / effective memory address space to access processor memories 401-402 and vice versa, thereby simplifying programmability. A first portion of the virtual / effective address space can be allocated to processor memory 401, a second portion can be allocated to second processor memory 402, a third portion can be allocated to GPU memory 420, and so on. The entire virtual / effective memory space (sometimes referred to as the effective address space) can thus be distributed across each of processor memories 401-402 and GPU memories 420-423, allowing any processor or GPU to access any physical memory using a virtual address mapped to that physical memory.

[0146] Bias / coherence management circuit modules 494A-494E within one or more of the MMUs 439A-439E may be provided to ensure cache coherence between the caches of the host processor (e.g., the multi-core processor 405) and the caches of the GPUs 410-413 and implement biasing techniques that indicate the physical memory in which certain types of data should be stored. Figure 4E Multiple instances of bias / coherence management circuit modules 494A- 494E are shown in , but bias / coherence circuit modules may be implemented within an MMU of one or more host processors and / or within the accelerator integrated circuit 436 .

[0147] The GPU-attached memory 420-423 can be mapped as part of system memory and accessed using shared virtual memory (SVM) technology, but without the typical performance drawbacks associated with full system cache coherence. The ability of the GPU-attached memory 420-423 to be accessed as system memory without the heavy cache coherence overhead provides a beneficial operating environment for GPU migration. This arrangement allows host processor software to set operands and access computation results without the overhead of traditional I / O DMA data copying. Such traditional copying involves driver calls, interrupts, and memory mapped I / O (MMIO) accesses, all of which are inefficient relative to simple memory accesses. At the same time, the ability to access the GPU-attached memory 420-423 without cache coherence overhead can be critical to the execution time of the migrated computation. For example, in situations with large amounts of streaming write-to-memory traffic, cache coherence overhead can significantly reduce the effective write bandwidth seen by the GPUs 410-413. The efficiency of operand setup, the efficiency of result access, and the efficiency of GPU computation all play a role in determining the effectiveness of GPU migration.

[0148] The selection between GPU bias and host processor bias can be driven by a bias tracker data structure. For example, a bias table can be used, which can be a page-granular structure (i.e., controlled at the granularity of a memory page) that includes one or two bits per GPU-attached memory page. The bias table can be implemented in a stolen memory range of one or more GPU-attached memories 420-423, with or without a bias cache in GPUs 410-413 (e.g., for caching frequently / recently used entries of the bias table). Alternatively, the entire bias table can be maintained within the GPU.

[0149] In one implementation, the bias table entry associated with each access to the GPU-attached memory 420-423 is accessed before the actual access to the GPU memory, resulting in the following operations. First, local requests from the GPUs 410-413 for pages that find their presence in the GPU bias are forwarded directly to the corresponding GPU memory 420-423. Local requests from the GPU for pages that find their presence in the host bias are forwarded to the processor (e.g., as discussed above, over a high-speed link). Optionally, requests from the host processor for pages that find their presence in the host processor bias complete the request like normal memory reads. Alternatively, requests involving pages that are GPU biased may be forwarded to the GPUs 410-413. If the GPU is not currently using the page, the GPU may then transition the page to the host processor bias. The bias state of a page may be changed by a software-based mechanism, a hardware-assisted software-based mechanism, or, for a limited set of cases, by a purely hardware-based mechanism. One mechanism for changing the bias state employs an API call (e.g., OpenCL), which in turn calls the GPU's device driver, which in turn sends a message (or queues a command descriptor) to the GPU directing the GPU to change the bias state and perform a cache flush operation in the host for some transitions. A cache flush operation is required for transitions from host processor bias to GPU bias, but not for the reverse transition.

[0150] Cache coherency can be maintained by temporarily rendering GPU-biased pages that are not cacheable by the host processor. To access these pages, the host processor may request access from GPU 410, which may or may not grant immediate access, depending on the implementation. Therefore, to reduce communication between the host processor and GPU 410, it is beneficial to ensure that GPU-biased pages are those required by the GPU but not by the host processor, and vice versa. Graphics processing pipeline

[0151] Figure 5 A graphics processing pipeline 500 is shown. A graphics multiprocessor such as Figure 2D Graphics multiprocessor 234 or Figure 2E The graphics multiprocessor 235 of FIG. 5 may implement the graphics processing pipeline 500. The graphics multiprocessor may be included in a parallel processing subsystem as described herein, such as a parallel processing subsystem. Figure 2A The parallel processor 200 can be used with Figure 1 The parallel processor(s) 112 are related and may be used in place of one of those parallel processors. Various parallel processing systems may be implemented via parallel processing units (e.g., Figure 2AThe graphics processing pipeline 500 may be implemented using one or more instances of the parallel processing unit 202 of FIG. For example, a shader unit (e.g., Figure 2C The graphics multiprocessor 234 of FIG5 may be configured to perform the functions of one or more of the following: a vertex processing unit 504, a tessellation control processing unit 508, a tessellation evaluation processing unit 512, a geometry processing unit 516, and a fragment / pixel processing unit 524. The functions of the data assembler 502, the primitive assemblers 506, 514, 518, the tessellation unit 510, the rasterizer 522, and the raster operation unit 526 may also be performed by a processing cluster (e.g., Figure 2A Other processing engines within the processing cluster 214 and corresponding partition units (e.g., Figure 2A The graphics processing pipeline 500 may also be implemented using dedicated processing units for one or more functions. It is also possible that one or more portions of the graphics processing pipeline 500 are executed by parallel processing logic within a general-purpose processor (e.g., a CPU). Optionally, one or more portions of the graphics processing pipeline 500 may access on-chip memory (e.g., such as a CPU) via a memory interface 528. Figure 2A The parallel processor memory 222 in the memory interface 528 can be Figure 2A The graphics processing pipeline 500 may also be connected to the memory interface 218 via the graphics processing pipeline 500. Figure 3 This is achieved using the multi-core group 365A in .

[0152] The data assembler 502 is a processing unit that collects vertex data for surfaces and primitives. The data assembler 502 then outputs the vertex data, including vertex attributes, to the vertex processing unit 504. The vertex processing unit 504 is a programmable execution unit that executes vertex shader programs, thereby lighting and transforming the vertex data as specified by the vertex shader programs. The vertex processing unit 504 reads data stored in cache, local, or system memory for use in processing vertex data and can be programmed to transform vertex data from an object-based coordinate representation to a world space coordinate space or a normalized device coordinate space.

[0153] A first instance of primitive assembler 506 receives vertex attributes from vertex processing unit 504. Primitive assembler 506 reads the stored vertex attributes as needed and builds graphics primitives for processing by tessellation control processing unit 508. Graphics primitives include triangles, line segments, points, patches, etc. as supported by various graphics processing application programming interfaces (APIs).

[0154] The tessellation control processing unit 508 treats the input vertices as control points of a geometry patch. The control points are transformed from an input representation of the patch (e.g., a basis for the patch) into a representation suitable for use in surface estimation by the tessellation evaluation processing unit 512. The tessellation control processing unit 508 may also calculate tessellation factors for the edges of the geometry patch. The tessellation factors are applied to individual edges and quantify the view-dependent level of detail associated with the edge. The tessellation unit 510 is configured to receive the tessellation factors for the edges of the patch and to tessellate the patch into a plurality of geometric primitives (e.g., line, triangle, or quadrilateral primitives), which are passed to the tessellation evaluation processing unit 512. The tessellation evaluation processing unit 512 operates on the parameterized coordinates of the tessellated patch to generate a surface representation and vertex attributes for each vertex associated with the geometric primitive.

[0155] A second instance of the primitive assembler 514 receives vertex attributes from the tessellation evaluation processing unit 512 as needed, reads the stored vertex attributes, and constructs graphics primitives for processing by the geometry processing unit 516. The geometry processing unit 516 is a programmable execution unit that executes a geometry shader program to transform the graphics primitives received from the primitive assembler 514 as specified by the geometry shader program. The geometry processing unit 516 can be programmed to tessellate the graphics primitive into one or more new graphics primitives and calculate parameters used to rasterize the new graphics primitives.

[0156] The geometry processing unit 516 may be capable of adding or removing elements from the geometry stream. The geometry processing unit 516 outputs parameters and vertices that specify new graphics primitives to the primitive assembler 518. The primitive assembler 518 receives the parameters and vertices from the geometry processing unit 516 and constructs graphics primitives for processing by the viewport scaling, culling, and clipping unit 520. The geometry processing unit 516 reads data stored in parallel processor memory or system memory for use in processing geometry data. The viewport scaling, culling, and clipping unit 520 performs clipping, culling, and viewport scaling, and outputs the processed graphics primitives to the rasterizer 522.

[0157] The rasterizer 522 can perform depth culling and other depth-based optimizations. The rasterizer 522 also performs scan conversion on new graphics primitives to generate fragments and outputs those fragments and associated coverage data to the fragment / pixel processing unit 524. The fragment / pixel processing unit 524 is a programmable execution unit configured to execute fragment shader programs or pixel shader programs. The fragment / pixel processing unit 524 transforms fragments or pixels received from the rasterizer 522 as specified by the fragment or pixel shader program. For example, the fragment / pixel processing unit 524 can be programmed to perform operations including, but not limited to, texture mapping, shading, blending, texture correction, and perspective correction to produce shaded fragments or pixels that are output to the raster operation unit 526. The fragment / pixel processing unit 524 can read data stored in parallel processor memory or system memory for use in processing the fragment data. The fragment or pixel shader program can be configured to shade at sample, pixel, slice, or other granularity, depending on the sampling rate configured for the processing unit.

[0158] Raster operation unit 526 is a processing unit that performs raster operations and outputs pixel data to be stored in graphics memory (e.g., Figure 2A The parallel processor memory 222 and / or Figure 1 The raster operations unit 526 may be configured to compress z-data or color data written to memory and decompress z-data or color data read from memory. Machine Learning Overview

[0159] The architecture described above can be applied to perform training and reasoning operations using machine learning models. Machine learning has been successful in solving many types of tasks. The calculations generated when training and using machine learning algorithms (e.g., neural networks) are naturally suitable for efficient parallel implementation. Accordingly, parallel processors such as general-purpose graphics processing units (GPGPUs) have played an important role in the actual implementation of deep neural networks. Parallel graphics processors with single instruction multiple threads (SIMT) architectures are designed to maximize the amount of parallel processing in the graphics pipeline. In the SIMT architecture, groups of parallel threads attempt to execute program instructions together synchronously as frequently as possible to increase processing efficiency. The efficiency provided by the parallel machine learning algorithm implementation allows the use of high-capacity networks and enables training of those networks to larger data sets.

[0160] Machine learning algorithms are algorithms that can learn from a dataset. For example, machine learning algorithms can be designed to model high-level abstractions within a dataset. For example, image recognition algorithms can be used to determine which of several categories a given input belongs to; regression algorithms can output a numerical value given an input; and pattern recognition algorithms can be used to generate converted text or perform text-to-speech and / or speech recognition.

[0161] An exemplary type of machine learning algorithm is a neural network. There are many types of neural networks; a simple type of neural network is a feedforward network. A feedforward network can be implemented as an acyclic graph in which nodes are arranged in layers. Typically, a feedforward network topology includes an input layer and an output layer separated by at least one hidden layer. The hidden layer transforms the input received by the input layer into a representation useful for generating the output in the output layer. The nodes of the network are fully connected via edges to nodes in adjacent layers, but there are no edges between nodes within each layer. Data received at the nodes of the input layer of the feedforward network is propagated (i.e., "fed forward") to the nodes of the output layer via an activation function, which calculates the state of the nodes in each successive layer of the network based on coefficients ("weights") associated with each of the edges connecting these layers. The output from the neural network algorithm can take various forms, depending on the specific model being represented by the algorithm being executed.

[0162] Before a machine learning algorithm can be used to model a particular problem, the algorithm is trained using a training data set. Training a neural network involves: selecting a network topology; using a set of training data that represents the problem being modeled by the network; and adjusting weights until the network model performs with minimal error for all instances of the training data set. For example, during a supervised learning training process for a neural network, the output generated by the network in response to an input representing an instance in the training data set is compared to the "correct" labeled output for that instance, an error signal representing the difference between the output and the labeled output is calculated, and as the error signal is propagated backward through the layers of the network, the weights associated with the connections are adjusted to minimize that error. The network is considered "trained" when the error for each of the outputs generated from the instances of the training data set is minimized.

[0163] The accuracy of a machine learning algorithm can be significantly affected by the quality of the dataset used to train the algorithm. The training process can be computationally intensive and can require a significant amount of time on a conventional general-purpose processor. Accordingly, many types of machine learning algorithms are trained using parallel processing hardware. This is particularly useful for optimizing the training of neural networks, as the calculations performed when adjusting the coefficients in a neural network are naturally suited to parallel implementation. In particular, many machine learning algorithms and software applications have been adapted to take advantage of the parallel processing hardware within general-purpose graphics processing devices.

[0164] Figure 6 is a generalized diagram of a machine learning software stack 600. A machine learning application 602 is any logic that can be configured to train a neural network using a training dataset or to implement machine intelligence using a trained deep neural network. The machine learning application 602 may include training and inference functionality for the neural network and / or specialized software that can be used to train the neural network prior to deployment. The machine learning application 602 may implement any type of machine intelligence, including but not limited to image recognition, mapping and localization, autonomous navigation, speech synthesis, medical imaging, or language translation. Example machine learning applications 602 include but are not limited to voice-based virtual assistants, image or facial recognition algorithms, autonomous navigation, and software tools used to train the machine learning models used by the machine learning application 602.

[0165] Hardware acceleration for machine learning applications 602 can be achieved via a machine learning framework 604. The machine learning framework 604 can provide a library of machine learning primitives. Machine learning primitives are basic operations typically performed by machine learning algorithms. Without the machine learning framework 604, developers of machine learning algorithms would be required to create and optimize the main computational logic associated with the machine learning algorithm, and then re-optimize the computational logic when new parallel processors are developed. Instead, machine learning applications can be configured to use the primitives provided by the machine learning framework 604 to perform the necessary computations. Exemplary primitives include tensor convolutions, activation functions, and pooling, which are computational operations performed when training convolutional neural networks (CNNs). The machine learning framework 604 can also provide primitives to implement basic linear algebra subroutines performed by many machine learning algorithms, such as matrix and vector operations. Examples of machine learning frameworks 604 include, but are not limited to, TensorFlow, TensorRT, PyTorch, MXNet, Caffe, and other advanced machine learning frameworks.

[0166] The machine learning framework 604 can process input data received from the machine learning application 602 and generate appropriate input to the computation framework 606. The computation framework 606 can abstract the underlying instructions provided to the GPGPU driver 608, enabling the machine learning framework 604 to utilize hardware acceleration via the GPGPU hardware 610 without requiring the machine learning framework 604 to be intimately familiar with the architecture of the GPGPU hardware 610. Furthermore, the computation framework 606 can enable hardware acceleration for the machine learning framework 604 across various types and generations of GPGPU hardware 610. In one example, the computation framework 606 can include the CUDA computation framework and associated machine learning libraries, such as the CUDA Deep Neural Network (cuDNN) library. The machine learning software stack 600 can also include a communication library or framework to facilitate multi-GPU and multi-node computation. GPGPU machine learning acceleration

[0167] Figure 7 A general purpose graphics processing unit (GPGPU 700) is shown, which may be Figure 2A Parallel processor 200 or Figure 1 The general processing unit can be configured to provide hardware acceleration support for primitives provided by machine learning frameworks to accelerate the processing of computational workloads associated with training deep neural networks. In addition, the GPGPU 700 can be directly connected to other instances of the GPGPU to create a multi-GPU cluster, thereby improving the training speed of deep neural networks in particular. Primitives are also supported to accelerate inference operations for deployed neural networks.

[0168] GPGPU 700 includes a host interface 702 for enabling connection to a host processor. Host interface 702 may be a PCI Express interface. However, the host interface may also be a vendor-specific communication interface or communication structure. GPGPU 700 receives commands from the host processor and uses a global scheduler 704 to distribute execution threads associated with those commands to a collection of processing clusters 706A-706H. The processing clusters 706A-706H share a cache memory 708. The cache memory 708 may act as a higher level cache for the cache memory within the processing clusters 706A-706H. The processing clusters 706A-706H shown may be used with Figure 2A Corresponding to the processing clusters 214A-214N in.

[0169] GPGPU 700 includes memory 714A-714B coupled to processing clusters 706A-706H via a set of memory controllers 712A-712B. Memories 714A-714B may include various types of memory devices, including dynamic random access memory (DRAM) or graphics random access memory, such as synchronous graphics random access memory (SGRAM), including graphics double data rate (GDDR) memory. Memories 714A-714B may also include 3D stacked memory, including but not limited to high bandwidth memory (HBM).

[0170] Each of processing clusters 706A-706H may include a collection of graphics multiprocessors, such as Figure 2D Graphics multiprocessor 234, Figure 2E The graphics multiprocessor 235 may include a Figure 3 The graphics multiprocessors of the compute clusters include multiple types of integer and floating-point logic units capable of performing computational operations with a range of precision, including those suitable for machine learning computations. For example, at least a subset of the floating-point units in each of the processing clusters 706A-706H can be configured to perform 16-bit or 32-bit floating-point operations, while a different subset of the floating-point units can be configured to perform 64-bit floating-point operations.

[0171] Multiple instances of GPGPU 700 can be configured to operate as a computing cluster. The communication mechanisms used by the computing cluster for synchronization and data exchange vary across embodiments. For example, multiple instances of GPGPU 700 communicate via a host interface 702. In one embodiment, GPGPU 700 includes an I / O hub 709 that couples GPGPU 700 to a GPU link 710, which enables direct connections to other instances of the GPGPU. GPU link 710 can be coupled to a dedicated GPU-to-GPU bridge that enables communication and synchronization between multiple instances of GPGPU 700. Optionally, GPU link 710 is coupled to a high-speed interconnect to transmit and receive data to and from other GPGPUs or parallel processors. Multiple instances of GPGPU 700 can be located in separate data processing systems and can communicate via a network device accessible via host interface 702. GPU link 710 may be configured to enable connection to a host processor in addition to or as an alternative to host interface 702 .

[0172] While the illustrated configuration of GPGPU 700 can be configured to train neural networks, alternative configurations of GPGPU 700 can be configured for deployment within a high-performance or low-power inference platform. In the inference configuration, GPGPU 700 includes fewer of the processing clusters 706A-706H relative to the training configuration. Additionally, the memory technology associated with memory 714A-714B can differ between the inference configuration and the training configuration. In one embodiment, the inference configuration of GPGPU 700 can support inference-specific instructions. For example, the inference configuration can provide support for one or more 8-bit integer or floating point dot product instructions, which are typically used during inference operations for deployed neural networks.

[0173] Figure 8 A multi-GPU computing system 800 is shown. The multi-GPU computing system 800 may include a processor 802 coupled to a plurality of GPGPUs 806A-806D via a host interface switch 804. The host interface switch 804 may be a PCI Express switching device that couples the processor 802 to a PCI Express bus, over which the processor 802 can communicate with the set of GPGPUs 806A-806D. Each of the plurality of GPGPUs 806A-806D may be a Figure 7 700. GPGPUs 806A-806D may be interconnected via a collection of high-speed point-to-point GPU-to-GPU links (P2PGPU links 816). The high-speed GPU-to-GPU links may be connected to each of GPGPUs 806A-806D via a dedicated GPU link, such as a dedicated GPU link. Figure 7 P2P GPU link 816 enables direct communication between each of GPGPUs 806A-806D without requiring communication over the host interface bus to which processor 802 is connected. With GPU-to-GPU traffic directed to the P2P GPU link, the host interface bus remains available for system memory access or for communicating with other instances of multi-GPU computing system 800, for example, via one or more network devices. Figure 8 GPGPUs 806A-806D are connected to processor 802 via host interface switch 804, but processor 802 may alternatively include direct support for P2P GPU link 816 and connect directly to GPGPUs 806A-806D. In one embodiment, P2P GPU link 816 enables multi-GPU computing system 800 to operate as a single logical GPU. Machine Learning Neural Network Implementation

[0174] The computing architecture described herein can be configured to perform a type of parallel processing that is particularly well-suited for training and deploying neural networks for machine learning. A neural network can be generalized as a network of functions having graph relationships. As is well known in the art, there are various types of neural network implementations used in machine learning. One exemplary type of neural network is a feedforward network as previously described. A second exemplary type of neural network is a convolutional neural network (CNN), while a third exemplary type of neural network is a recurrent neural network (RNN).

[0175] CNNs are specialized feedforward neural networks for processing data with a known, grid-like topology (such as image data). Accordingly, CNNs are commonly used in computational vision and image recognition applications, but they can also be used for other types of pattern recognition, such as speech and language processing. The nodes in the input layer of a CNN are organized into a collection of "filters" (feature detectors inspired by the receptive fields found in the retina), and the output of each filter set is propagated to nodes in successive layers of the network. The calculations used in a CNN involve applying a convolution mathematical operation to each filter to produce the output of that filter. Convolution is a specialized mathematical operation that is performed by two functions to produce a third function that is a modified version of one of the two original functions. In convolutional network terminology, the first function to be convolved can be called the input, and the second function can be called the convolution kernel. The output can be called a feature map. The input to the convolution layer can be a multidimensional data array that defines the various color components of the input image. The convolution kernel can be a multidimensional parameter array, where the parameters are adapted through the training process used for the neural network.

[0176] An RNN is a series of feedforward neural networks that include feedback connections between layers. RNNs are able to model sequential data by sharing parameter data across different parts of the neural network. The architecture of an RNN includes loops that represent the impact of the current value of a variable on its own value at future times, as at least part of the output data from the RNN is used as feedback for processing subsequent inputs in the sequence. This feature makes RNNs particularly useful for language processing due to the mutable nature of language data, which can be composed.

[0177] The figures described below present exemplary feedforward networks, CNN networks, and RNN networks, and describe the general process for training and deploying each of those types of networks, respectively. It will be understood that these descriptions are exemplary and non-limiting with respect to any specific embodiment described herein, and that the concepts shown are generally applicable to deep neural networks and machine learning techniques in general.

[0178] The deep neural networks used in deep learning typically include a front-end network for performing feature recognition, which is coupled to a back-end network that represents a mathematical model that can perform operations (e.g., object classification, speech recognition, etc.) based on the feature representations provided to the mathematical model. Deep learning enables machine learning to be performed without the need for manual feature engineering for the model. Instead, deep neural networks can learn features based on statistical structures or correlations within the input data. The learned features can be provided to a mathematical model that can map the detected features to an output. The mathematical model used by the network is typically dedicated to a specific task to be performed, and different models will be used to perform different tasks.

[0179] Once the neural network is structured, a learning model can be applied to the network to train the network to perform a specific task. The learning model describes how to adjust the weights within the model to reduce the output error of the network. Back propagation of errors is a common method for training neural networks. The input vector is presented to the network for processing. The output of the network is compared with the desired output using a loss function, and an error value is calculated for each neuron in the output layer. Subsequently, the error value is propagated backward until each neuron has an associated error value that roughly represents the contribution of the neuron to the original output. The network can then learn from those errors using an algorithm (such as a stochastic gradient descent algorithm) to update the weights of the neural network.

[0180] Figure 9A-9B An exemplary convolutional neural network is shown. Figure 9A Shows the various layers within CNN. Figure 9A As shown in , an exemplary CNN for modeling image processing may receive an input 902 describing the red, green, and blue (RGB) components of an input image. The input 902 may be processed by a plurality of convolutional layers (e.g., a first convolutional layer 904, a second convolutional layer 906). The outputs from the plurality of convolutional layers may optionally be processed by a fully connected layer 908. As previously described for feedforward networks, neurons in a fully connected layer have full connections to all activations in the previous layer. The outputs from the fully connected layer 908 may be used to generate output results from the network. Activations within the fully connected layer 908 may be calculated using matrix multiplication rather than convolution. Not all CNN implementations utilize the fully connected layer 908. For example, in some implementations, the second convolutional layer 906 may generate the output of the CNN.

[0181] The convolutional layers are sparsely connected, which is different from the traditional neural network configuration found in the fully connected layer 908. Traditional neural network layers are fully connected so that every output unit interacts with every input unit. However, as shown, the convolutional layers are sparsely connected because the output of the convolution of the receptive field (rather than the corresponding state value of each node in the receptive field) is input to the nodes of the subsequent layer. The kernel associated with the convolutional layer performs a convolution operation, the output of which is sent to the next layer. The dimensionality reduction performed within the convolutional layer is one aspect that enables CNNs to scale to process large images.

[0182] Figure 9B 1 shows exemplary computational stages within a convolutional layer of a CNN. Input 912 to a convolutional layer of the CNN can be processed in three stages of a convolutional layer 914. These three stages can include a convolution stage 916, a detector stage 918, and a pooling stage 920. The convolutional layer 914 can then output data to a subsequent convolutional layer. The final convolutional layer of the network can generate output feature map data or provide input to a fully connected layer, for example, to generate a classification value for input to the CNN.

[0183] Several convolutions are performed in parallel in the convolution stage 916 to produce a set of linear activations. The convolution stage 916 may include an affine transformation, which is any transformation that can be defined as a linear transformation plus a translation. Affine transformations include rotations, translations, scaling, and combinations of these transformations. The convolution stage calculates the output of a function (e.g., a neuron) connected to specific regions in the input, which can be determined as local regions associated with the neuron. The neuron calculates the dot product between the weight of the neuron and the weight of the region in the local input to which the neuron is connected. The output from the convolution stage 916 defines the set of linear activations processed by the successive stages of the convolution layer 914.

[0184] The linear activations may be processed by the detector stage 918. In the detector stage 918, each linear activation is processed by a nonlinear activation function. Nonlinear activation functions increase the nonlinear nature of the overall network without affecting the receptive field of the convolutional layers. Several types of nonlinear activation functions may be used. One particular type is the rectified linear unit (ReLU), which uses an activation function defined as f(x) = max(0, x) such that the threshold of the activation is zero.

[0185] The pooling stage 920 uses a pooling function that replaces the output of the second convolutional layer 906 with summary statistics of nearby outputs. The pooling function can be used to introduce translation invariance into the neural network so that small translations to the input do not change the pooled output. The invariance of local translation can be useful in scenarios where the exact location of the feature in the input data is more important. Various types of pooling functions can be used during the pooling stage 920, including maximum pooling, average pooling, and l2 norm pooling. In addition, some CNN implementations do not include a pooling stage. Instead, such implementations are alternative and additional convolution stages with an increased span relative to the previous convolution stage.

[0186] The output from the convolutional layer 914 may then be processed by the next layer 922. The next layer 922 may be an additional convolutional layer or one of the fully connected layers 908. For example, Figure 9A The first convolutional layer 904 can output to the second convolutional layer 906, and the second convolutional layer can output to the first layer in the fully connected layer 908.

[0187] A variant of CNN is the convolutional deep belief network, which has a similar structure to CNN and is trained in a similar manner to the deep belief network. A deep belief network (DBN) is a generative neural network consisting of multiple layers of stochastic (random) variables. DBNs can be trained layer by layer using greedy unsupervised learning. The learned weights of the DBN can then be used to provide a pre-trained neural network by determining the optimal initial set of weights for the neural network.

[0188] Figures 10A-10B An exemplary language model is shown. Figure 10AA recurrent neural network (RNN 1000) is shown. In a recurrent neural network (RNN), the previous state of the network influences the output of the current state of the network. RNNs can be built in various ways and using various functions. The use of RNNs generally revolves around using mathematical models to predict the future based on previous sequences of inputs. For example, RNNs can be used to perform statistical language modeling to predict upcoming words given a previous sequence of words. RNN 1000 can be described as having an input layer 1002 that receives an input vector, a hidden layer 1004 for implementing a recurrent function, a feedback mechanism 1005 for enabling a 'memory' of previous states, and an output layer 1006 for outputting a result. RNN 1000 operates based on time steps. The state of the RNN at a given time step is influenced by the feedback mechanism 1005 based on the previous time step. For a given time step, the state of the hidden layer 1004 is defined by the previous state and the input at the current time step. The initial input (x1) at the first time step can be processed by the hidden layer 1004. The second input (x2) can be processed by the hidden layer 1004 using the state information determined during the processing of the initial input (x1). The given state can be calculated as s t =f(Ux t +Ws t-1 ), where U and W are parameter matrices. The function f is typically nonlinear, such as a variant of the hyperbolic tangent function (Tanh) or the rectifier function f(x) = max(0, x). However, the specific mathematical function used in the hidden layer 1004 may vary depending on the specific implementation details of the RNN 1000. It is also possible to achieve acceleration for changes on the RNN network. An example RNN variant is the long short-term memory (LSTM) RNN. LSTM RNNs are able to learn long-term dependencies, which may be necessary for processing long language sequences.

[0189] Figure 10B The baseline components of the transformer model 1010 are shown. The transformer model 1010 solves problems found in RNN models with long input sequences and enables increased parallelization. The transformer model 1010 and its variants are used to build language models (LLMs) that perform various tasks such as machine translation, automatic summarization, and dialogue management. The transformer model 1010 can also be configured to perform image generation tasks such as text-to-image generation.

[0190] The transformer model 1010 includes multiple instances of an encoder 1016 and a decoder 1026. The encoders are stacked end-to-end, with the output of the final encoder being routed as input to the multi-head attention layer of each decoder 1026. The input into the bottom-most instance of the encoder 1016 is processed by the input embedding 1012, which converts the input token into a vector that can be processed by the encoder 1016. The output dictionary is vectorized by the output embedding 1022 before entering the bottom-most instance of the decoder 1026. Positional encodings 1014 and 1024 are added to the input and output vectors at the bottom of the encoder 1016 and decoder 1026 stacks. The positional encodings 1014 and 1024 inject information about the relative or absolute position of the token in the sequence of tokens to be processed, because the transformer model 1010 does not naturally encode the order of tokens.

[0191] The encoder 1016 of the transformer model 1010 analyzes the input text and creates multiple hidden states that preserve the context and meaning of the text data. The encoder 1016 layer forms part of the core of the transformer architecture, although variants of the transformer model 1010 with only the decoder 1026 are also possible. The encoder 1016 includes two sublayers, a multi-head attention (MHA) sublayer and a feedforward network (FFN) sublayer. The MHA sublayer performs multiple concurrent self-attention operations to calculate attention scores, which enables the transformer model 1010 to weigh the importance and relative relationships of different tokens in the input sequence in a context-aware manner. The FFN sublayer is a positional fully connected feedforward neural network. The output of each sublayer is processed by an addition and normalization (A&N) operation defined as LayerNorm(x+Sublayer(x)), where the output of the sublayer is added to the input of the sublayer and normalized. In some implementations, the normalization operation can be performed before the sublayer rather than after.

[0192] Decoder 1026 includes three sublayers: a masked MHA sublayer, an MHA sublayer, and a FFN sublayer. The masked MHA sublayer is similar to the MHA layer, except that masking is applied to prevent query positions from focusing on keys for future positions. The MHA sublayer of decoder 1026 is similar to the MHA sublayer of the encoder, with additional input from encoder 1016. The FFN of decoder 1026 is the same as the FFN of encoder 1016. The linear and softmax blocks take the output of the last instance of decoder 1026 of the decoder stack and generate a probability distribution representing the output probabilities.

[0193] GPU acceleration can also be used for variants of the transformer model 1010 that replace some or all of the FFN sublayers with sparse mixture of experts (MoE) layers. Each MoE layer includes multiple experts, where each expert is a neural network. The MoE layer can itself be an FFN, or it can also be an MoE, allowing for hierarchical MoE layers.

[0194] Training of the transformer model 1010 can be optimized through the use of adaptive precision logic, which adjusts the precision of calculations applied during training. The adaptive precision logic can attempt to use the smallest possible data types during training without significantly reducing training accuracy. For example, to minimize data loss due to the use of 16-bit, 8-bit, and 4-bit floating point formats, dynamic scaling and casting can be applied during training based on statistical analysis of the tensor data generated during training. The tensor data generated during training can be statistically analyzed using various analysis techniques to determine a set of scaling factors to be applied to data blocks (such as inputs to each layer of the transformer model 1010). For example, absolute minimum and / or maximum values ​​can be used to determine scaling factors that can prevent underflow or overflow of low-precision floating point data types.

[0195] Figure 11 The training and deployment of a deep neural network is shown. Once a given network has been structured for a task, the neural network is trained using a training dataset 1102. A training framework 1104 has been developed to enable hardware acceleration of the training process. For example, Figure 6 The machine learning framework 604 can be configured as a training framework 1104. The training framework 1104 can access an untrained neural network 1106 and enable the untrained neural network to be trained using the parallel processing resources described herein to generate a trained neural network 1108. To begin the training process, initial weights can be selected randomly or by pre-training using a deep belief network. Subsequently, a training cycle is performed in a supervised or unsupervised manner.

[0196] Supervised learning is a learning method in which training is performed as a mediated operation, such as when a training dataset 1102 includes inputs paired with their expected outputs, or when the training dataset includes inputs with known outputs and the outputs of the neural network are manually graded. The network processes the inputs and compares the resulting output to the expected output or set of expected outputs. Errors are then propagated back through the system. The training framework 1104 can be adjusted to adjust the weights controlling the untrained neural network 1106. The training framework 1104 can provide tools to monitor how well the untrained neural network 1106 is converging on a model suitable for generating the correct answer based on the known input data. The training process occurs iteratively as the network's weights are adjusted to refine the outputs generated by the neural network. The training process can continue until the neural network reaches a statistically expected level of accuracy associated with the trained neural network 1108. The trained neural network 1108 can then be deployed to implement any number of machine learning operations to generate inference results 1114 based on the input of new data 1112.

[0197] Unsupervised learning is a learning method in which a network attempts to train itself using unlabeled data. Therefore, for unsupervised learning, the training data set 1102 will include input data without any associated output data. The untrained neural network 1106 can learn groupings within the unlabeled inputs and can determine how individual inputs relate to the entire data set. Unsupervised training can be used to generate self-organizing maps, which are a type of trained neural network 1108 that can perform operations useful in reducing the dimensionality of data. Unsupervised training can also be used to perform anomaly detection, which allows identification of data points in the input data set that deviate from the normal pattern of the data.

[0198] Variations of supervised and unsupervised training can also be employed. Semi-supervised learning is a technique in which a mixture of labeled and unlabeled data having the same distribution is included in the training data set 1102. Incremental learning is a variation of supervised learning in which input data is continuously used to further train the model. Incremental learning enables a trained neural network 1108 to adapt to new data 1112 without forgetting the knowledge embedded in the network during initial training. Whether supervised or unsupervised, the training process for particularly deep neural networks can be too computationally intensive for a single computing node. A distributed network of computing nodes can be used instead of a single computing node to accelerate the training process.

[0199] Figure 12Ais a block diagram illustrating distributed learning; distributed learning is a training model that uses multiple distributed computing nodes to perform supervised or unsupervised training of neural networks. Each distributed computing node may include one or more host processors and one or more general processing nodes, such as Figure 7 As shown, distributed learning can be performed with model parallelism 1202, data parallelism 1204, or a combination of model and data parallelism 1206.

[0200] In model parallelism 1202, different compute nodes in a distributed system can perform training computations for different parts of a single network. For example, each layer of a neural network can be trained by a different processing node of the distributed system. Benefits of model parallelism include the ability to scale to extremely large models. Splitting the computations associated with different layers of a neural network enables training of very large neural networks, where the weights of all layers would not fit into the memory of a single compute node. In some instances, model parallelism can be particularly useful when performing unsupervised training of large neural networks. Some implementations of model parallelism 1202 may also be referred to as tensor parallelism.

[0201] In data parallelism 1204, different nodes of the distributed network have complete instances of the model, and each node receives a different portion of the data. The results from the different nodes are then combined. Although different approaches to data parallelism are possible, all data parallel training approaches require techniques for combining the results and synchronizing the model parameters between each node. Exemplary approaches for combining data include parameter averaging and update-based data parallelism. Parameter averaging trains each node on a subset of the training data and sets global parameters (e.g., weights, biases) to the average of the parameters from each node. Parameter averaging uses a central parameter server that maintains parameter data. Update-based data parallelism is similar to parameter averaging, except that updates to the model are transmitted instead of parameters from the node to the parameter server. In addition, update-based data parallelism can be performed in a decentralized manner, where updates are compressed and transmitted between nodes.

[0202] Combined model and data parallelism 1206 can be implemented, for example, in a distributed system where each compute node includes multiple GPUs. Combined model and data parallelism 1206 can also be referred to as hybrid parallelism. Each node can have a complete instance of the model, with separate GPUs within each node being used to train different parts of the model. Distributed training has increased overhead relative to training on a single machine. However, the parallel processors and GPGPUs described herein can each implement various techniques to reduce the overhead of distributed training, including techniques for enabling high-bandwidth GPU-to-GPU data transfer and accelerated remote data synchronization. Pipeline parallelism is a variation of combined model and data parallelism 1206, in which different nodes contain less than the entire model, but more than a single layer of the model. In pipeline parallelism, different groups of layers or sub-models are distributed across different processing nodes. Another variation is expert parallelism, which routes requests for specific experts within the model to different GPUs. For example, expert parallelism can be used with MoE transformer models.

[0203] Figure 12B 1 is a block diagram illustrating a programmable network interface 1210 and a data processing unit. The programmable network interface 1210 is a programmable network engine that can be used to accelerate network-based computing tasks within a distributed environment. The programmable network interface 1210 can be coupled to a host system via a host interface 1270. The programmable network interface 1210 can be used to accelerate network or storage operations for a CPU or GPU of the host system. The host system can be, for example, a node of a distributed learning system for performing distributed training, such as Figure 12A The host system can also be a data center node within a data center.

[0204] In one embodiment, access to a remote storage device containing model data may be accelerated by the programmable network interface 1210. For example, the programmable network interface 1210 may be configured to present the remote storage device as a local storage device of the host system. The programmable network interface 1210 may also accelerate remote direct memory access (RDMA) operations performed between a GPU of the host system and a GPU of the remote system. In one embodiment, the programmable network interface 1210 may be capable of implementing storage functionality such as, but not limited to, NVME-oF. The programmable network interface 1210 may also accelerate encryption, data integrity, compression, and other operations for the remote storage device on behalf of the host system, thereby allowing the remote storage device to approach the latency of a storage device directly attached to the host system.

[0205] Programmable network interface 1210 can also perform resource allocation and management on behalf of the host system. Storage security operations can be migrated to programmable network interface 1210 and performed in conjunction with the allocation and management of remote storage resources. Network-based operations for managing access to remote storage devices that would otherwise be performed by the host system's processor can instead be performed by programmable network interface 1210.

[0206] In one embodiment, network and / or data security operations can be migrated from the host system to the programmable network interface 1210. Data center security policies for data center nodes can be handled by the programmable network interface 1210 rather than by the host system's processor. For example, the programmable network interface 1210 can detect and mitigate attempted network-based attacks (e.g., DDoS) on the host system, thereby preventing the attack from compromising the availability of the host system.

[0207] Programmable network interface 1210 may include a system on a chip (SoC 1220) that executes an operating system via multiple processor cores 1222. Processor core 1222 may include a general-purpose processor (e.g., CPU) core. In one embodiment, processor core 1222 may also include one or more GPU cores. SoC 1220 may execute instructions stored in memory device 1240. Storage device 1250 may store local operating system data. Storage device 1250 and memory device 1240 may also be used to cache remote data for a host system. Network ports 1260A-1260B enable connection to a network or fabric and facilitate network access for SoC 1220 and for the host system via host interface 1270. Programmable network interface 1210 may also include an I / O interface 1275, such as a USB interface. I / O interface 1275 may be used to couple external devices to programmable network interface 1210 or as a debug interface. Programmable network interface 1210 also includes a management interface 1230 that enables software on a host device to manage and configure programmable network interface 1210 and / or SoC 1220. In one embodiment, programmable network interface 1210 may also include one or more accelerators or GPUs 1245 to accept migration of parallel computing tasks from SoC 1220, a host system, or a remote system coupled via network ports 1260A-1260B. Example Machine Learning Applications

[0208] Machine learning can be applied to solve various technical problems, including but not limited to computer vision, autonomous driving and navigation, speech recognition and language processing. Computer vision has traditionally been one of the most active research areas for machine learning applications. The application range of computer vision is from reproducing human visual capabilities (such as recognizing faces) to creating new categories of visual capabilities. For example, computer vision applications can be configured to identify sound waves from vibrations induced in objects visible in a video. Parallel processor-accelerated machine learning enables computer vision applications to be trained using significantly larger training data sets than previously feasible, and enables reasoning systems to be deployed using low-power parallel processors.

[0209] Parallel processor-accelerated machine learning has autonomous driving applications, including lane and road sign recognition, obstacle avoidance, navigation, and driving control. Accelerated machine learning techniques can be used to train driving models based on datasets that define appropriate responses to specific training inputs. The parallel processors described herein may enable rapid training of increasingly complex neural networks for autonomous driving solutions and enable the deployment of low-power inference processors in mobile platforms suitable for integration into autonomous vehicles.

[0210] Deep neural networks accelerated by parallel processors have enabled machine learning approaches for automatic speech recognition (ASR). ASR involves creating a function that computes the most likely speech sequence given a sequence of input sounds. Accelerated machine learning using deep neural networks has enabled an alternative to the hidden Markov models (HMMs) and Gaussian mixture models (GMMs) previously used for ASR.

[0211] Parallel processor-accelerated machine learning can also be used to accelerate natural language processing. The automated learning process can utilize statistical inference algorithms to generate models that are robust to erroneous or unfamiliar inputs. Exemplary natural language processor applications include automatic machine translation between human languages.

[0212] The parallel processing platform for machine learning can be divided into a training platform and a deployment platform. The training platform is generally highly parallel and includes optimizations for accelerating multi-GPU single-node training and multi-node multi-GPU training. Exemplary parallel processors suitable for training include Figure 7 GPGPU700 and Figure 8 In contrast, deployed machine learning platforms typically include lower-powered parallel processors suitable for use in products such as cameras, autonomous robots, and autonomous vehicles.

[0213] Furthermore, machine learning techniques can be applied to accelerate or enhance graphics processing activities. For example, a machine learning model can be trained to recognize the output generated by a GPU-accelerated application and generate a magnified version of that output. Such techniques can be applied to accelerate the generation of high-resolution images for gaming applications. Various other graphics pipeline activities can benefit from the use of machine learning. For example, a machine learning model can be trained to perform tessellation operations on geometric data to increase the complexity of the geometric model, thereby allowing more detailed geometry to be automatically generated from relatively low-detail geometry.

[0214] Figure 13 An exemplary inference system-on-chip (SOC) 1300 suitable for performing inference using a trained model is shown. The SOC 1300 may integrate processing components, including a media processor 1302, a vision processor 1304, a GPGPU 1306, and a multi-core processor 1308. The GPGPU 1306 may be a GPGPU as described herein, such as GPGPU 700, and the multi-core processor 1308 may be a multi-core processor as described herein, such as multi-core processors 405-406. The SOC 1300 may additionally include on-chip memory 1305, which may enable a shared on-chip data pool accessible by each of the processing components. The processing components may be optimized for low-power operation to enable deployment of various machine learning platforms, including autonomous vehicles and autonomous robots. For example, an implementation of the SOC 1300 may be used as part of a master control system for an autonomous vehicle. Where the SOC 1300 is configured for use in an autonomous vehicle, the SOC is designed and configured to comply with relevant functional safety standards of the deployment jurisdiction.

[0215] During operation, the media processor 1302 and the vision processor 1304 can work in tandem to accelerate computer vision operations. The media processor 1302 can be capable of low-latency decoding of multiple high-resolution (e.g., 4K, 8K) video streams. The decoded video streams can be written to a buffer in the on-chip memory 1305. The vision processor 1304 can then parse the decoded video and perform preliminary processing operations on the frames of the decoded video to prepare them for processing using a trained image recognition model. For example, the vision processor 1304 can accelerate convolution operations for a CNN used to perform image recognition on high-resolution video data, while back-end model calculations are performed by the GPGPU 1306.

[0216] The multi-core processor 1308 may include control logic for facilitating the sequencing and synchronization of data transfers and shared memory operations performed by the media processor 1302 and the vision processor 1304. The multi-core processor 1308 may also function as an application processor for executing software applications capable of utilizing the inference computing capabilities of the GPGPU 1306. For example, at least a portion of navigation and steering logic may be implemented in software executing on the multi-core processor 1308. Such software may directly issue computational workloads to the GPGPU 1306, or computational workloads may be issued to the multi-core processor 1308, which may offload at least a portion of those operations to the GPGPU 1306.

[0217] GPGPU 1306 may include a compute cluster, such as a low-power configuration of processing clusters 706A-706H within GPGPU 700. The compute cluster within GPGPU 1306 may support instructions specifically optimized to perform inference computations on trained neural networks. For example, GPGPU 1306 may support instructions for performing low-precision computations such as 8-bit and 4-bit integer vector operations. Additional System Overview

[0218] Figure 14 is a block diagram of processing system 1400 . Figure 14 Elements having the same or similar names as elements of any other figures herein describe the same elements as in the other figures, can operate or function in a manner similar to that in the other figures, may include the same components, and may be linked to other entities such as, but not limited to, those described elsewhere herein. Processing system 1400 can be used in: a single-processor desktop computer system, a multi-processor workstation system, or a server system having (one or more) processors 1402 or processor cores 1407. Processing system 1400 can be a processing platform incorporated within a system-on-chip (SoC) integrated circuit for use in a mobile device, a handheld device, or an embedded device, such as for use in an Internet-of-things (IoT) device having wired or wireless connectivity to a local area network or a wide area network.

[0219] The processing system 1400 may be a system having Figure 1 For example, in different configurations, the processor(s) 1402 or processor core 1407 may be associated with the components of the processing system corresponding to those components. Figure 1 The graphics processor(s) 1408 may correspond to the processor(s) 102 of FIG. Figure 1The external graphics processor 1418 may be Figure 1 One of the (one or more) plug-in devices 120.

[0220] The processing system 1400 may include, be coupled with, or be integrated into: a server-based gaming platform; a gaming console, including gaming and media consoles; a mobile gaming console, a handheld gaming console, or an online gaming console. The processing system 1400 may be part of a mobile phone, a smartphone, a tablet computing device, or a mobile internet-connected device, such as a laptop with low internal storage capacity. The processing system 1400 may also include, be coupled with, or be integrated into: a wearable device, such as a smartwatch wearable device; smart glasses or clothing that is enhanced with augmented reality (AR) or virtual reality (VR) features to provide visual, audio, or tactile output to supplement the real-world visual, audio, or tactile experience or otherwise provide text, audio, graphics, video, holographic images or video, or tactile feedback. The processing system 1400 may include, or be part of a television or set-top box device. Processing system 1400 may include, be coupled to, or be integrated within an autonomous vehicle, such as a bus, a tractor-trailer, an automobile, an electric motor or electric power cycle, an airplane, or a glider (or any combination thereof). The autonomous vehicle may use processing system 1400 to process the environment sensed around the vehicle.

[0221] The processor(s) 1402 may include one or more instances of a processor core 1407 to process instructions that, when executed, perform operations for system or user software. At least one of the processor cores 1407 may be configured to process a specific instruction set 1409. The instruction set 1409 may facilitate complex instruction set computing (CISC), reduced instruction set computing (RISC), or computation via very long instruction words (VLIW). In one embodiment, one of the processor cores 1407 may process a different instruction set 1409, which may include instructions for facilitating emulation of other instruction sets. The processor core 1407 may also include other processing devices, such as a digital signal processor (DSP).

[0222] Processor(s) 1402 may include cache memory 1404. Depending on the architecture, processor(s) 1402 may have a single internal cache or multiple levels of internal cache. In some embodiments, cache memory is shared between various components of processor(s) 1402. In some embodiments, processor(s) 1402 also uses an external cache (e.g., a level 3 (L3) cache or a last level cache (LLC)) (not shown), which may be shared between processor cores 1407 using known cache coherence techniques. A register file 1406 may additionally be included in processor(s) 1402 and may include different types of registers (e.g., integer registers, floating point registers, status registers, and an instruction pointer register) for storing different types of data. Some registers may be general purpose registers, while other registers may be specific to the design of processor(s) 1402.

[0223] The processor(s) 1402 may be coupled to one or more interface buses 1410 to transmit communication signals, such as address, data, or control signals, between the processor(s) 1402 and other components in the processing system 1400. In one of these embodiments, the interface bus(es) 1410 may be a processor bus, such as a version of a direct media interface (DMI) bus. However, the processor bus is not limited to a DMI bus and may include one or more peripheral component interconnect buses (e.g., PCI, PCI express), a memory bus, or other types of interface buses. For example, the processor(s) 1402 may include a memory controller 1416 and a platform controller hub 1430. The memory controller 1416 facilitates communication between memory devices and other components of the processing system 1400, while the platform controller hub 1430 provides connections to I / O devices via a local I / O bus.

[0224] Memory device 1420 may be a dynamic random access memory (DRAM) device, a static random access memory (SRAM) device, a flash memory device, a phase change memory device, or some other memory device with suitable performance to function as process memory. Memory device 1420 may, for example, operate as system memory for processing system 1400 to store data 1422 and instructions 1421 for use when processor(s) 1402 execute applications or processes. Memory controller 1416 may optionally be coupled to an external graphics processor 1418, which may communicate with graphics processor(s) 1408 in processor(s) 1402 to perform graphics and media operations. In some embodiments, graphics operations, media operations, and / or compute operations may be assisted by accelerator 1412, which is a coprocessor that can be configured to perform a specialized set of graphics, media, or compute operations. For example, accelerator 1412 may be a matrix multiplication accelerator for optimizing machine learning or compute operations. The accelerator 1412 may be a ray tracing accelerator that can be used to perform ray tracing operations in conjunction with the graphics processor(s) 1408. The accelerator 1412 may also be an AI accelerator or NPU to accelerate neural network training or inference operations. In one embodiment, an external accelerator 1419 may be used in place of the accelerator 1412 or in conjunction with the accelerator 1412. The accelerator 1412 and / or the external accelerator 1419 may have the same Figure 1 The functionality of the accelerator device(s) 130 is similar to the functionality of the accelerator device(s) 130 .

[0225] A display device 1411 may be provided and connected to the processor(s) 1402. The display device 1411 may be one or more of an internal display device, such as in a mobile electronic device or laptop device, or an external display device attached via a display interface (e.g., a display port, etc.). The display device 1411 may be a head mounted display (HMD), such as a stereoscopic display device for use in VR or AR applications.

[0226] The platform controller hub 1430 can enable peripheral devices to connect to the memory device 1420 and the processor(s) 1402 via a high-speed I / O bus. The I / O peripherals include, but are not limited to, an audio controller 1446, a network controller 1434, a firmware interface 1428, a wireless transceiver 1426, a touch sensor 1425, and a data storage device 1424 (e.g., non-volatile memory, volatile memory, hard drive, flash memory, NAND, 3D NAND, 3D XPoint / Optane, etc.). The data storage device 1424 can be connected via a storage interface (e.g., SATA) or via a peripheral bus such as a peripheral component interconnect bus (e.g., PCI, PCI Express). The touch sensor 1425 can include a touch screen sensor, a pressure sensor, or a fingerprint sensor. The wireless transceiver 1426 can be a Wi-Fi transceiver, a Bluetooth transceiver, or a mobile network transceiver such as a 3G, 4G, 5G, or Long-Term Evolution (LTE) transceiver. The firmware interface 1428 enables communication with the system firmware and can be, for example, a unified extensible firmware interface (UEFI). The network controller 1434 can enable network connectivity to a wired network. In some embodiments, a high-performance network controller (not shown) is coupled to the interface bus(es) 1410. The audio controller 1446 can be a multi-channel high-definition audio controller. In some of these embodiments, the processing system 1400 includes an optional legacy I / O controller 1440 for coupling legacy (e.g., Personal System 2 (PS / 2)) devices to the system. The platform controller hub 1430 can also be connected to one or more Universal Serial Bus (USB) controllers 1442 to connect to input devices, such as a keyboard and mouse 1443 combination, a camera 1444, or other USB input devices.

[0227] It will be appreciated that the processing system 1400 shown is exemplary and non-limiting, as other types of data processing systems configured in different manners may also be used. For example, instances of the memory controller 1416 and the platform controller hub 1430 may be integrated into a discrete external graphics processor, such as the external graphics processor 1418. The platform controller hub 1430 and / or the memory controller 1416 may be external to the processor(s) 1402. For example, the memory controller 1416 and the platform controller hub 1430 may be external to the processing system 1400 and configured as a memory controller hub and a peripheral controller hub within a system chipset that communicate with the processor(s) 1402.

[0228] For example, a circuit board ("sled") can be used, on which components (such as a CPU, memory, and other components) are placed, and on which components (such as a CPU, memory, and other components) are designed to achieve improved thermal performance. Processing components such as a processor can be located on the top side of the sled, while nearby memory such as DIMMs is located on the bottom side of the sled. As a result of the enhanced airflow provided by this design, components can operate at higher frequencies and power levels than in typical systems, thereby improving performance. In addition, the sled is configured to blindly mate with power and data communication cables in the rack, thereby enhancing their ability to be quickly removed, upgraded, reinstalled, and / or replaced. Similarly, the various components located on the sled (such as processors, accelerators, memory, and data storage drives) are configured to be easily upgraded due to their increased spacing from each other. In an illustrative embodiment, the components additionally include hardware authentication features for proving their authenticity.

[0229] The data center can utilize a single network architecture ("fabric") that supports multiple other network architectures, including Ethernet and omni-path. The sleds can be coupled to the switches via optical fiber, which provides higher bandwidth and lower latency than typical twisted pair cabling (e.g., Category 5, Category 5e, Category 6, Category 7, Category 8, etc.). Due to the high-bandwidth, low-latency interconnect and network architecture, the data center can, in use, pool physically dispersed resources such as memory, accelerators (e.g., GPUs, graphics accelerators, FPGAs, ASICs, neural network and / or artificial intelligence accelerators, etc.) and data storage drives and provide them to computing resources (e.g., processors) as needed, enabling the computing resources to access these centralized resources as if they were local.

[0230] A power supply or power source can provide voltage and / or current to the processing system 1400 or any component or system described herein. In one example, the power supply includes an AC to DC (alternating current to direct current) adapter for plugging into a wall outlet. Such an AC power source can be a renewable energy source (e.g., solar power). In one example, the power source includes a DC power source, such as an external AC to DC converter. The power source or power supply can also include wireless charging hardware for charging via proximity to a charging field. The power source can include an internal battery, an AC supply, a motion-based power supply, a solar power supply, or a fuel cell source.

[0231] Figures 15A-15C A computing system and a graphics processor are shown. Figure 14 Elements having the same or similar names as elements of any other figures herein describe the same elements as those in the other figures, can operate or function in a similar manner as those in the other figures, may include the same components, and may be linked to other entities such as, but not limited to, those described elsewhere herein.

[0232] Figure 15A1402. FIGURE 1 is a block diagram of processor 1500, which may be a variant of one of processor(s) 1402 and may be used in place of one of those processors. Therefore, any feature disclosed herein in conjunction with processor 1500 also discloses the corresponding combination with processor(s) 1402, but is not limited thereto. Processor 1500 may have one or more processor cores 1502A-1502N, at least one memory controller 1514, and a graphics processor 1508. Graphics processor 1508 may be integrated within processor 1500, within a system chipset, or coupled via a system bus. Processor 1500 may include additional cores, up to and including additional core 1502N, represented by a dashed box. Each of processor cores 1502A-1502N includes one or more internal cache units 1504A-1504N. In some embodiments, each processor core 1502A-1502N also has access to one or more shared cache units 1506. The internal cache units 1504A-1504N and the shared cache unit(s) 1506 represent a cache memory hierarchy within the processor 1500. The cache memory hierarchy may include at least one level of instruction and data cache within each processor core and one or more levels of shared mid-level cache, such as a level 2 (L2), level 3 (L3), level 4 (L4), or other level of cache, with the highest level of cache before external memory being categorized as LLC. In some embodiments, cache coherence logic maintains coherence between the cache units (e.g., the shared cache unit(s) 1506 and the internal cache unit(s) 1504A-1504N).

[0233] The processor 1500 may also include one or more bus controller units 1516 and a system agent core 1510. The one or more bus controller units 1516 manage a collection of peripheral buses, such as one or more PCI buses or PCI Express buses. The one or more bus controller units 1516 may also manage one or more memory buses to various external memory devices (not shown). The system agent core 1510 provides management functionality for various processor components and may include at least one memory controller 1514.

[0234] For example, one or more of the processor cores 1502A-1502N may include support for simultaneous multithreading. The system agent core 1510 includes components for coordinating and operating the cores 1502A-1502N during multithreaded processing. The system agent core 1510 may additionally include a power control unit (PCU) that includes logic and components for regulating the power state of the processor cores 1502A-1502N and the graphics processor 1508.

[0235] Processor 1500 may additionally include a graphics processor 1508 for performing graphics processing operations. In some of these embodiments, graphics processor 1508 is coupled with one or more shared cache units 1506 and a system agent core 1510 including at least one memory controller 1514. System agent core 1510 may also include a display controller 1511 for driving the graphics processor output to one or more coupled displays. Display controller 1511 may also be a separate module coupled to the graphics processor via at least one interconnect, or may be integrated within graphics processor 1508.

[0236] A ring or mesh-based interconnect 1512 may be used to couple the internal components of the processor 1500. However, alternative interconnect means may be used, such as a point-to-point interconnect, a switched interconnect, or other technologies, including those known in the art. In some of these embodiments having a ring or mesh-based interconnect 1512, the graphics processor 1508 is coupled to the ring or mesh-based interconnect 1512 via an I / O link 1513.

[0237] Exemplary I / O link 1513 represents at least one of a variety of I / O interconnects, including an on-package I / O interconnect that facilitates communication between various processor components and a high-performance memory module 1518, such as an embedded DRAM module (eDRAM) or a high-bandwidth memory (HMB) module. Optionally, each of processor cores 1502A-1502N and graphics processor 1508 can utilize high-performance memory module 1518 as a unified memory and / or shared last-level cache when a DRAM memory system is also present. Optionally, one or more accelerators 1515 may also be included within processor 1500, including, for example, an NPU for accelerating certain neural network operations. The NPU may enable lower-power inference operations relative to the use of graphics processor 1508, or may operate in conjunction with graphics processor 1508 to enable higher inference performance relative to a graphics processor 1508 alone. In one embodiment, the NPU within one or more accelerators 1515 may include matrix or tensor acceleration logic and may be used to implement at least some of the computational operations described herein as being implementable via graphics processor 1508.

[0238] The processor cores 1502A-1502N may, for example, be homogeneous cores that execute the same instruction set architecture. Alternatively, the processor cores 1502A-1502N are heterogeneous in terms of instruction set architecture (ISA), wherein one or more of the processor cores 1502A-1502N execute a first instruction set and at least one of the other cores executes a subset of the first instruction set or a different instruction set. The processor cores 1502A-1502N may be heterogeneous in terms of microarchitecture, wherein one or more cores with relatively high power consumption are coupled with one or more power cores with lower power consumption. As another example, the processor cores 1502A-1502N are heterogeneous in terms of computing power. Furthermore, the processor 1500 may be implemented on one or more chips or chiplets, or as a SoC integrated circuit having the components shown in addition to other components. The SoC integrated circuit may be implemented using multiple chiplets.

[0239] Figure 15B is a block diagram of the hardware logic of the graphics processor core block 1519 according to some embodiments described herein. In some embodiments, Figure 15BElements having the same reference numerals (or names) as elements of any other figures herein may operate or function in a manner similar to that described elsewhere herein. In one embodiment, graphics processor core block 1519 is an example of a partition of a graphics processor. Graphics processor core block 1519 may be included in Figure 15A 1508 or a discrete graphics processor, parallel processor, and / or compute accelerator. A graphics processor as described herein may include multiple graphics core blocks based on a target power and performance envelope. Each graphics processor core block 1519 may include a functional block 1530 coupled to multiple graphics cores 1521A-1521F, which may include modular blocks of fixed-function logic and general-purpose programmable logic. The graphics processor core block 1519 also includes a shared / cache memory 1536 accessible by all graphics cores 1521A-1521F, rasterizer logic 1537, and additional fixed-function logic 1538.

[0240] In some embodiments, functional block 1530 includes a geometry / fixed function pipeline 1531 that can be shared by all graphics cores in graphics processor core block 1519. In various embodiments, geometry / fixed function pipeline 1531 includes a 3D geometry pipeline, a video front-end unit, a thread generator, a global thread dispatcher, and a unified return buffer manager that manages the unified return buffer. In one embodiment, functional block 1530 also includes a graphics SoC interface 1532, a graphics microcontroller 1533, and a media pipeline 1534. Graphics SoC interface 1532 provides an interface between graphics processor core block 1519 and other core blocks within a graphics processor or compute accelerator SoC. Graphics microcontroller 1533 is a programmable subprocessor that can be configured to manage various functions of graphics processor core block 1519, including thread dispatching, scheduling, and preemption. Media pipeline 1534 includes logic for facilitating decoding, encoding, pre-processing, and / or post-processing of multimedia data, including image and video data. The media pipeline 1534 implements media operations via requests to computational or sampling logic within the graphics cores 1521A-1521F. One or more pixel backends 1535 may also be included within the functional block 1530. The one or more pixel backends 1535 include buffer memory for storing pixel color values ​​and are capable of performing blending operations and lossless color compression on rendered pixel data.

[0241] In one embodiment, graphics SoC interface 1532 enables graphics processor core block 1519 to communicate with a general-purpose application processor core (e.g., a CPU) and / or other components within the SoC or with a system host CPU coupled to the SoC via a peripheral interface. Graphics SoC interface 1532 also enables communication with off-chip memory hierarchy elements, such as shared last-level cache memory, system RAM, and / or embedded on-chip or on-package DRAM. Graphics SoC interface 1532 may also enable communication with fixed-function devices within the SoC, such as a camera imaging pipeline, and enable use and / or implementation of global memory atomicity, which may be shared between graphics processor core block 1519 and the CPU within the SoC. Graphics SoC interface 1532 may also enable power management control for graphics processor core block 1519 and enable interfacing between the clock domain of graphics processor core block 1519 and other clock domains within the SoC. In one embodiment, graphics SoC interface 1532 enables receiving command buffers from a command streamer and a global thread dispatcher, which are configured to provide commands and instructions to each of one or more graphics cores within a graphics processor. Commands and instructions can be dispatched to media pipeline 1534 when media operations are to be performed, and can be dispatched to geometry and fixed function pipeline 1531 when graphics processing operations are to be performed. When compute operations are to be performed, compute dispatch logic can dispatch commands to graphics cores 1521A-1521F, thereby bypassing the geometry and media pipelines.

[0242] Graphics microcontroller 1533 can be configured to perform various scheduling and management tasks for graphics processor core block 1519. In one embodiment, graphics microcontroller 1533 can execute graphics workloads and / or compute workloads scheduled on the various vector engines 1522A-1522F, 1524A-1524F and matrix engines 1523A-1523F, 1525A-1525F within graphics cores 1521A-1521F. In this scheduling model, host software executing on a CPU core of the SoC that includes graphics processor core block 1519 can submit a workload to one of multiple graphics processor doorbells, which invokes scheduling operations on the appropriate graphics engine. Scheduling operations include determining which workload to run next, submitting the workload to the command stream converter, preempting existing workloads running on the engine, monitoring the progress of the workload, and notifying the host software when the workload is complete. In one embodiment, the graphics microcontroller 1533 is also capable of facilitating a low power or idle state for the graphics processor core block 1519, thereby providing the graphics processor core block 1519 with the ability to save and restore registers within the graphics processor core block 1519 across low power state transitions independent of the operating system and / or graphics driver software on the system.

[0243] Graphics processor core block 1519 may have more or fewer graphics cores 1521A-1521F than shown, up to a maximum of N modular graphics cores. For each set of N graphics cores, graphics processor core block 1519 may also include: shared / cache memory 1536, which may be configured as shared memory or cache memory; rasterizer logic 1537; and additional fixed-function logic 1538 for accelerating various graphics and compute processing operations.

[0244] Within each graphics core 1521A-1521F is a collection of execution resources that can be used to perform graphics operations, media operations, and compute operations in response to requests made by the graphics pipeline, media pipeline, or shader programs. Graphics cores 1521A-1521F include multiple vector engines 1522A-1522F, 1524A-1524F, matrix acceleration units 1523A-1523F, 1525A-1525D, cache / shared local memory (SLM), samplers 1526A-1526F, and ray tracing units 1527A-1527F.

[0245] The vector engines 1522A-1522F, 1524A-1524F are general-purpose graphics processing units capable of performing floating-point and integer / fixed-point logic operations to serve graphics, media, or compute operations (including graphics programs, media programs, or compute / GPGPU programs). The vector engines 1522A-1522F, 1524A-1524F can operate in SIMD, SIMT, or SIMT+SIMD execution modes, with variable vector widths. The matrix acceleration units 1523A-1523F, 1525A-1525D include matrix-matrix and matrix-vector acceleration logic that improves the performance of matrix operations, particularly low-precision and mixed-precision (e.g., INT8, FP16, BF16, FP8, FP4) matrix operations for machine learning. In one embodiment, the matrix acceleration units 1523A-1523F, 1525A-1525D support the microscale (MX) format. In one embodiment, each of the matrix acceleration units 1523A-1523F, 1525A-1525D includes one or more systolic arrays of processing elements capable of performing concurrent matrix multiplication or dot product operations on matrix elements.

[0246] Samplers 1526A-1526F can read media data or texture data into memory and can sample the data in different ways based on the configured sampler state and the texture / media format being read. Threads executing on vector engines 1522A-1522F, 1524A-1524F or matrix acceleration units 1523A-1523F, 1525A-1525D can utilize caches / SLMs 1528A-1528F within each of graphics cores 1521A-1521F. Caches / SLMs 1528A-1528F can be configured as a pool of cache memory or shared memory local to each of the corresponding graphics cores 1521A-1521F. Ray tracing units 1527A-1527F within graphics cores 1521A-1521F include ray traversal / intersection circuitry for performing ray traversals using a bounding volume hierarchy (BVH) and identifying intersections between rays and primitives enclosed within the BVH volume. In one embodiment, ray tracing units 1527A-1527F include circuitry for performing depth testing and culling (e.g., using a depth buffer or similar arrangement). In one implementation, ray tracing units 1527A-1527F perform traversal and intersection operations in conjunction with image denoising, at least in part of which may be performed using associated matrix acceleration units 1523A-1523F, 1525A-1525D.

[0247] Figure 15C 15 is a block diagram of a general-purpose graphics processing unit (GPGPU) 1570, which can be configured as a graphics processor (e.g., graphics processor 1508) and / or a computing accelerator, according to embodiments described herein. GPGPU 1570 can be interconnected with a host processor (e.g., one or more CPUs 1546) and memories 1571 and 1572 via one or more system and / or memory buses. Memory 1571 can be system memory that can be shared with one or more CPUs 1546, while memory 1572 is device memory dedicated to GPGPU 1570. For example, components within GPGPU 1570 and memory 1572 can be mapped to memory addresses accessible by one or more CPUs 1546. Access to memories 1571 and 1572 can be facilitated via a memory controller 1568. Memory controller 1568 can include an internal direct memory access (DMA) controller 1569, or can include logic for performing operations that would otherwise be performed by a DMA controller. In one embodiment, at least one of the one or more CPUs 1546 may include one or more accelerators 1545, including but not limited to a neural network accelerator.

[0248] GPGPU 1570 includes multiple global cache memories, including an L2 cache 1553, an L1 cache 1554, an instruction cache 1555, and a shared memory 1556. At least a portion of the shared memory 1556 may also be partitioned as cache memory. GPGPU 1570 also includes multiple compute units 1560A-1560N. Each compute unit 1560A-1560N includes a set of vector registers 1561, a set of scalar registers 1562, a set of vector logic units 1563, a set of scalar logic units 1564, and a scheduler 1584. Compute units 1560A-1560N may also include a local shared memory 1565 and a local cache memory 1566. Compute units 1560A-1560N may be coupled with a constant cache 1567 that can be used to store constant data, which is data that will not change during the execution of a kernel program or shader program executed on GPGPU 1570. Constant cache 1567 can be a scalar data cache, and the cached data can be directly fetched into scalar registers 1562. In one embodiment, compute units 1560A-1560N additionally include at least one matrix unit 1580 for accelerating matrix, tensor, or artificial intelligence operations and at least one ray tracing unit (RT unit 1582) for accelerating ray tracing operations. The at least one matrix unit 1580 and RT unit 1582 may include functionality similar to the other matrix / tensor accelerators and ray tracing accelerators described herein.

[0249] During operation, one or more CPUs 1546 may write commands to registers in GPGPU 1570 or to memory in GPGPU 1570 that has been mapped into an accessible address space. Command processor 1557 may read commands from registers or memory and determine how to process those commands within GPGPU 1570. Threads may then be dispatched to compute units 1560A-1560N using thread dispatcher 1558 to execute those commands. Each compute unit 1560A-1560N may execute threads independently of the other compute units. Furthermore, each compute unit 1560A-1560N may be independently configured for conditional computations and may conditionally output the results of the computations to memory. Command processor 1557 may interrupt one or more CPUs 1546 when the submitted commands are completed.

[0250] Figure 161 is a block diagram of a graphics processor 1600, which may be a discrete graphics processing unit or a graphics processor integrated with multiple processing cores or other semiconductor devices, such as, but not limited to, memory devices or network interfaces. Elements of graphics processor 1600 having the same or similar names as elements in any other figures herein describe the same elements as those in the other figures, can operate or function in a similar manner, may include the same components, and may be linked to other entities, such as, but not limited to, those described elsewhere in this document. For example, graphics processor 1600 may be a variant of graphics processor 1508 and may be used in place of graphics processor 1508. The graphics processor may communicate via a memory-mapped I / O interface to registers on the graphics processor and using commands placed in processor memory. Graphics processor 1600 may include a memory interface 1614 to access local memory, one or more internal caches, one or more shared external caches, and / or system memory.

[0251] Graphics processor 1600 may include a display controller 1602 to drive display output data to a display device 1618. Display controller 1602 includes hardware for compositing one or more overlay planes and multiple layers of video or user interface elements for the display. Display device 1618 may be an internal or external display device. In one embodiment, display device 1618 is a head-mounted display device, such as a VR or AR display device. The graphics processor 1600 may include a video codec engine 1606 for encoding media into one or more media coding formats, decoding media from one or more media coding formats, or transcoding media between one or more media coding formats, including but not limited to: Moving Picture Experts Group (MPEG) formats (such as MPEG-2), Advanced Video Coding (AVC) formats (such as H.264 / MPEG-4 AVC, H.265 / HEVC, Alliance for Open Media (AOMedia) VP8, VP9), and the Society of Motion Picture & Television Engineers (SMPTE) 421M / VC-1, and Joint Photographic Experts Group (JPEG) formats (such as JPEG and Motion JPEG (MJPEG) formats).

[0252] Graphics processor 1600 may include a block image transfer (BLIT) engine 1603 to perform two-dimensional (2D) rasterizer operations, including, for example, bit-boundary block transfers, or may use one or more components of a graphics processing engine (GPE 1610) to perform 2D graphics operations. GPE 1610 may include a 3D pipeline 1612 for performing 3D operations, such as rendering three-dimensional images and scenes using processing functions that operate on 3D primitive shapes (e.g., rectangles, triangles, etc.). 3D pipeline 1612 includes programmable and fixed-function elements that perform various tasks within the element and / or spawn execution threads to a 3D / media subsystem 1615. While 3D pipeline 1612 may be used to perform media operations, GPE 1610 may also include a media pipeline 1616 specifically for performing media operations, such as image or video decoding, encoding, post-processing, and enhancement. The media pipeline 1616 may include fixed-function or programmable logic units for performing one or more specialized media operations, such as video decoding acceleration, video deinterlacing, and video encoding acceleration, instead of, or on behalf of, the video codec engine 1606. The media pipeline 1616 may additionally include a thread generation unit for generating threads for execution on the 3D / media subsystem 1615. The generated threads perform calculations for the media operations on one or more graphics execution units included in the 3D / media subsystem 1615.

[0253] Figure 17A Graphics processor 1720 is shown. Figure 161600 and may be used in place of graphics processor 1600 and vice versa. Therefore, any feature disclosed herein in conjunction with graphics processor 1600 also discloses the corresponding combination with graphics processor 1720, but is not limited thereto. According to the embodiments described herein, graphics processor 1720 has a sliced ​​architecture. Graphics processor 1720 may include a graphics processing engine cluster 1722 having multiple graphics processing engines within multiple graphics engine slices. Each graphics engine slice 1710A-1710D may be interconnected via a set of slice interconnects 1723A-1723F. Each graphics engine slice 1710A-1710D may also be connected to a memory module or memory device 1726A-1726D via a memory interconnect 1725A-1725D. Memory devices 1726A-1726D may use any graphics memory technology. For example, the memory devices 1726A-1726D may be graphics double data rate (GDDR) memory. The memory devices 1726A-1726D may be high bandwidth memory (HBM) modules that may be on-die with their corresponding graphics engine slices 1710A-1710D. The memory devices 1726A-1726D may be stacked memory devices that may be stacked on top of their corresponding graphics engine slices 1710A-1710D. Each graphics engine slice 1710A-1710D and associated memory devices 1726A-1726D may reside on separate chiplets that are bonded to a base die or base substrate, as in Figures 25A-25B As described in further detail in .

[0254] Graphics processor 1720 may be configured with a non-uniform memory access (NUMA) system in which memory devices 1726A-1726D are coupled to associated graphics engine slices 1710A-1710D. A given memory device may be accessed by a graphics engine slice different from the graphics engine slice to which it is directly connected. However, access latency to memory devices 1726A-1726D may be minimized when accessing the local slice. In one embodiment, a cache coherent NUMA (ccNUMA) system may be implemented that uses slice interconnects 1723A-1723F to enable communication between cache controllers within graphics engine slices 1710A-1710D to maintain a consistent memory image when more than one cache stores the same memory location.

[0255] Graphics processing engine cluster 1722 may be connected to an interconnect fabric 1724, which may interconnect on-chip or on-package structures. In one embodiment, interconnect fabric 1724 includes a network processor, a network on a chip (NoC), or another switching processor that enables interconnect fabric 1724 to function as a packet-switched interconnect fabric for exchanging data packets between components of graphics processor 1720. Interconnect fabric 1724 may enable communication between graphics engine slices 1710A-1710D and components such as video codec engine 1706 and one or more replication engines 1704. One or more copy engines 1704 may be used to move data out of memory devices 1726A-1726D and memory external to graphics processor 1620 (e.g., system memory), to move data into memory devices 1726A-1726D and memory external to graphics processor 1620 (e.g., system memory), and between memory devices 1726A-1726D and memory external to graphics processor 1620 (e.g., system memory). Interconnect structure 1724 may also be used to interconnect graphics engine slices 1710A-1710D. Graphics processor 1720 may optionally include a display controller 1702 to enable connection to display device 1718. The graphics processor may also be configured as a graphics accelerator or a compute accelerator. In an accelerator configuration, display controller 1702 and display device 1718 may be omitted.

[0256] Graphics processor 1720 may be connected to a host system via host interface 1728. Host interface 1728 may enable communication between graphics processor 1720, system memory, and / or other system components. Host interface 1728 may be, for example, a PCI Express bus or another type of host system interface. For example, host interface 1728 may be an NVLink or NVSwitch interface. Host interface 1728 and interconnect structure 1724 may cooperate to enable multiple instances of graphics processor 1720 to function as a single logical device. The cooperation between host interface 1728 and interconnect structure 1724 may also enable each graphics engine slice 1710A-1710D to appear to the host system as a distinct logical graphics device.

[0257] Figure 17B FIG1 shows a computing accelerator 1730 according to an embodiment described herein. The computing accelerator 1730 may include Figure 17BThe computing accelerator 1730 may be configured as an AI accelerator or an NPU. In such an embodiment, the execution logic of the computing engine slices 1740A-1740D may be primarily targeted at matrix or tensor operations and include the tensor cores or matrix engines described herein. The computing engine slices 1740A-1740D may not include fixed-function graphics processing logic, but in some embodiments, one or more of the computing engine slices 1740A-1740D may include logic for performing media acceleration. The computing engine slices 1740A-1740D may be connected to the memory devices 1726A-1726D via the memory interconnects 1725A-1725D. Memory devices 1726A-1726D and memory interconnects 1725A-1725D may be similar technologies as in graphics processor 1720, or may be different technologies. Compute engine slices 1740A-1740D may also be interconnected via a set of slice interconnects 1723A-1723F, and may be connected to and / or interconnected through fabric interconnect 1724. In one embodiment, compute accelerator 1730 includes a large L3 cache 1736 that may be configured as a device-wide cache. Compute accelerator 1730 may also be configured in a manner similar to that of a graphics processor 1720. Figure 17B The graphics processor 1720 is similarly connected to a host processor and memory via a host interface 1728 .

[0258] Computing accelerator 1730 may also include an integrated network interface 1742. In one embodiment, integrated network interface 1742 includes a network processor and controller logic that enables computing engine cluster 1732 to communicate via physical layer interconnect 1744 without requiring data to traverse the host system's memory. In one embodiment, one of computing engine slices 1740A-1740D is replaced by network processor logic, and data to be transmitted or received via physical layer interconnect 1744 can be transmitted directly to or from memory devices 1726A-1726D. Multiple instances of computing accelerator 1730 can be combined into a single logical device via physical layer interconnect 1744. Alternatively, each computing engine slice 1740A-1740D can be presented as a different network-accessible computing accelerator device. Graphics processing resources

[0259] Figures 18A-18CExecution logic including an array of processing elements employed in a graphics processor is shown according to embodiments described herein. Figure 18A A graphics core cluster is shown according to one embodiment. Figure 18B A vector engine of a graphics core is shown according to one embodiment. Figure 18C Shown is a matrix engine of a graphics core according to one embodiment. Figures 18A-18C Elements having the same reference numerals as elements of any other figure herein may operate or function in any manner similar to that described elsewhere herein, but are not limited thereto. For example, Figures 18A-18C The components can be Figure 15B In one embodiment, Figures 18A-18C The components have Figure 15A Graphics processor 1508 or Figure 15C The equivalent components of the GPGPU 1570 have similar functionality.

[0260] like Figure 18A As shown in FIG, in one embodiment, graphics core cluster 1800 includes graphics processor core block 1519, which may include any number of graphics cores (e.g., graphics core 1815A, graphics core 1815B, through graphics core 1815N). Multiple instances of graphics processor core block 1519 may be included. In one embodiment, the elements of graphics cores 1815A-1815N have the same Figure 15B 1806N, ray tracing units 1808A-1808N, and samplers 1810A-1810N. The circuit modules of graphics cores 1815A-1815N may additionally include fixed-function logic 1812A-1812N. The number of vector engines 1802A-1802N and matrix engines 1803A-1803N within a design's graphics cores 1815A-1815N may vary based on the workload, performance, and power targets for the design.

[0261] Referring to graphics core 1815A, vector engine 1802A and matrix engine 1803A can be configured to perform parallel computations on data in various integer and floating-point data formats based on instructions associated with a shader program. Each vector engine 1802A and matrix engine 1803A can act as a programmable general-purpose computing unit capable of executing multiple simultaneous hardware threads while processing multiple data elements in parallel for each thread. Vector engine 1802A and matrix engine 1803A support processing variable-width vectors in various SIMD widths, including but not limited to SIMD8, SIMD16, and SIMD32. Input data elements can be stored in registers as packed data types, and vector engine 1802A and matrix engine 1803A can process each element based on its data size. For example, when operating on a 256-bit wide vector, the bits of the vector are stored in registers, and the vector is processed as four separate 64-bit packed data elements (quad-word (QW) size data elements), eight separate 32-bit packed data elements (double-word (DW) size data elements), sixteen separate 16-bit packed data elements (word (W) size data elements), or thirty-two separate 8-bit data elements (byte (B) size data elements). However, different vector widths and register sizes are possible. In one embodiment, the vector engine 1802A and the matrix engine 1803A can also be configured to perform SIMT operations on various sizes of cell groups and thread groups (e.g., 8, 16, or 32 threads).

[0262] Continuing with the graphics core 1815A, the memory load / store unit 1804A services memory access requests issued by the vector engine 1802A, the matrix engine 1803A, and / or other components of the graphics core 1815A that have access to memory. The memory access request may be processed by the memory load / store unit 1804A to load or store the requested data into a cache or memory, or from a cache or memory into a register file associated with the vector engine 1802A and / or the matrix engine 1803A. The memory load / store unit 1804A may also perform prefetch operations. Also referring to Figure 19In one embodiment, the memory load / store unit 1804A is configured to provide SIMT scatter / gather prefetches or block prefetches for data stored in memory 1910, from memory local to other slices via the slice interconnect 1908, or from system memory. Prefetches can be performed for a specific L1 cache (e.g., data cache / shared local memory 1806A), the L2 cache 1904, or the L3 cache 1906. In one embodiment, a prefetch to the L3 cache 1906 automatically causes the data to be stored in the L2 cache 1904.

[0263] Instruction cache 1805A stores instructions to be executed by graphics core 1815A. In one embodiment, graphics core 1815A also includes instruction fetch and prefetch circuit modules that fetch or prefetch instructions into instruction cache 1805A. Graphics core 1815A also includes instruction decode logic for decoding instructions within instruction cache 1805A. Data cache / shared local memory 1806A can be configured as a data cache managed by a cache controller that implements a cache replacement policy and / or configured as explicitly managed shared memory. Ray tracing unit 1808A includes circuit modules for accelerating ray tracing operations. Sampler 1810A provides texture sampling for 3D operations and media sampling for media operations. Fixed function logic 1812A includes fixed function circuit modules that are shared between instances of vector engine 1802A and matrix engine 1803A. Graphics cores 1815B-1815N can operate in a manner similar to graphics core 1815A.

[0264] The functionality of the instruction caches 1805A-1805N, data cache / shared local memory 1806A-1806N, ray tracing units 1808A-1808N, samplers 1810A-1810N, and fixed function logic 1812A-1812N corresponds to the equivalent functionality in the graphics processor architecture described herein. For example, the instruction caches 1805A-1805N can be used with Figure 15C The data cache / shared local memory 1806A-1806N, ray tracing units 1808A-1808N and samplers 1810A-1810N can operate in a similar manner to the instruction cache 1555 of FIG. Figure 15B The fixed function logic 1812A-1812N may include: Figure 15B In one embodiment, ray tracing units 1808A-1808N include components for performing the ray tracing operations performed by Figure 3 A circuit module for ray tracing acceleration operations performed by the ray tracing core 372.

[0265] like Figure 18B As shown in FIG, in one embodiment, the vector engine 1802 includes an instruction fetch unit 1837, a general register file array (GRF 1824), an architectural register file array (ARF 1826), a thread arbiter 1822, an issue unit 1830, a branch unit 1832, a SIMDFPU 1834, and, in one embodiment, a SIMD ALU 1835. The GRF 1824 and ARF 1826 comprise a set of general register files and architectural register files associated with each hardware thread that can be active in the vector engine 1802. In one embodiment, per-thread architectural state is maintained in the ARF 1826, while data used during thread execution is stored in the GRF 1824. The execution state of each thread, including the instruction pointer for each thread, can be stored in thread-specific registers in the ARF 1826. Register renaming can be used to dynamically assign registers to hardware threads.

[0266] In one embodiment, vector engine 1802 has an architecture that is a combination of simultaneous multi-threading (SMT) and fine-grained interleaved multi-threading (IMT). This architecture has a modular configuration that can be fine-tuned at design time based on the target number of simultaneous threads and the number of registers per graphics core, where graphics core resources are divided across the logic used to execute multiple simultaneous threads. The number of logical threads that can be executed by vector engine 1802 is not limited to the number of hardware threads, and multiple logical threads can be assigned to each hardware thread.

[0267] In one embodiment, vector engine 1802 can collaboratively issue multiple instructions, each of which can be different instructions. Thread arbiter 1822 can dispatch instructions to one of issue unit 1830, branch unit 1832, or SIMDFPU 1834 for execution. Each execution thread can access 128 general-purpose registers within GRF 1824, each of which can store 32 bytes accessible as variable-width vectors of 32-bit data elements. In one embodiment, each thread has access to 4 kilobytes within GRF 1824, but embodiments are not limited thereto, and more or fewer register resources may be provided in other embodiments. In one embodiment, vector engine 1802 is partitioned into seven hardware threads that can independently execute computational operations, but the number of threads per vector engine 1802 may vary depending on the embodiment. For example, in one embodiment, up to 16 hardware threads are supported. In an embodiment where seven threads have access to 4 kilobytes, GRF 1824 can store a total of 28 kilobytes. With 16 threads accessing 4 kilobytes, the GRF 1824 can store a total of 64 kilobytes. Flexible addressing modes allow registers to be addressed together, effectively creating wider registers or representing strided rectangular block data structures.

[0268] In one embodiment, memory operations, sampler operations, and other longer latency system communications are dispatched via "send" instructions executed through message passing send unit 1830. In one embodiment, branch instructions are dispatched to branch unit 1832 to facilitate SIMD scatter and eventual convergence.

[0269] In one embodiment, the SIMDFPU 1834 of the vector engine 1802 performs floating-point operations. In one embodiment, the SIMDFPU 1834 also supports integer calculations. In one embodiment, the SIMD FPU 1834 can perform up to M 32-bit floating-point (or integer) operations, or up to 2M 16-bit integer or 16-bit floating-point operations. In one embodiment, at least one of the FPUs provides extended math capabilities that support high-throughput transcendental math functions and double-precision 64-bit floating-point. In some embodiments, there is also a SIMD ALU 1835 configured to perform 8-bit integer operations and can be specifically optimized to perform operations associated with machine learning calculations. In one embodiment, the SIMDALU 1835 is replaced by a SIMDFPU 1834 that can be configured to perform integer and floating-point operations. In one embodiment, the SIMDFPU 1834 and the SIMDALU 1835 can be configured to execute SIMT programs. In one embodiment, combined SIMD+SIMT operations are supported.

[0270] In one embodiment, an array of multiple instances of vector engine 1802 can be instantiated within a graphics core. For scalability, product architects can choose the exact number of vector engines grouped per graphics core. In one embodiment, vector engine 1802 can execute instructions across multiple execution channels. In another embodiment, each thread executing on vector engine 1802 executes on a different channel.

[0271] like Figure 18C As shown in , in one embodiment, the matrix engine 1803 includes an array of processing elements configured to perform tensor operations, including vector / matrix operations and matrix / matrix operations, such as but not limited to matrix multiplication and / or dot product operations. The matrix engine 1803 can be configured using M rows and N columns of processing elements 1852AA-1852MN, which include multiplier and adder circuit modules organized in a pipelined manner. In one embodiment, the processing elements 1852AA-1852MN form the physical pipeline stage of an N-wide and M-deep systolic array, which can be used to perform vector / matrix operations or matrix / matrix operations in a data-parallel manner, including matrix multiplication, fused multiply-add, dot product or other general matrix-matrix multiplication (GEMM) operations. In one embodiment, the matrix engine 1803 supports 16-bit and 8-bit floating-point operations, as well as 8-bit, 4-bit, 2-bit and binary integer operations. The matrix engine 1803 can also be configured to accelerate specific machine learning operations. In such embodiments, the matrix engine 1803 may be configured with support for a bfloat (brain floating point) 16-bit floating point format, or a tensor float 32-bit floating point format (TF32), having a different number of mantissa and exponent bits relative to the Institute of Electrical and Electronics Engineers (IEEE) 754 format.

[0272] In one embodiment, during each cycle, each stage can add the result of the operation performed in that stage to the output of the previous stage. In other embodiments, after a set of computation cycles, the pattern of data movement between processing elements 1852AA-1852MN can vary based on the instructions or macro-operations being executed. For example, in one embodiment, a partial sum loopback can be implemented, and a processing element can instead add the output of the current cycle to the output generated in the previous cycle. In one embodiment, the final stage of a systolic array can be configured with a loopback to the initial stage of the systolic array. In such embodiments, the number of physical pipeline stages can be decoupled from the number of logical pipeline stages supported by matrix engine 1803. For example, if processing elements 1852AA-1852MN are configured as a systolic array of M physical stages, a loopback from stage M to the initial pipeline stage can enable processing elements 1852AA-1852MN to operate as a systolic array of, for example, 2M, 3M, 4M, and so on, logical pipeline stages.

[0273] In one embodiment, matrix engine 1803 includes memories 1841A-1841N, 1842A-1842M for storing input data in the form of row and column data for input matrices. Memories 1842A-1842M are configurable to store the row elements (A0-Am) of a first input matrix, and memories 1841A-1841N are configurable to store the column elements (B0-Bn) of a second input matrix. Row elements and column elements are provided as input to processing elements 1852AA-1852MN for processing. In one embodiment, the row elements and column elements of the input matrix can be stored in a systolic register file 1840 within matrix engine 1803 before these elements are provided to memories 1841A-1841N, 1842A-1842M. In one embodiment, systolic register file 1840 is excluded and the registers (e.g., Figure 18B 1802 of the vector engine 1802) or other memory of the graphics core including the matrix engine 1803 (e.g., Figure 18A The results generated by the processing elements 1852AA-1852MN are then output to output buffers and / or written to register files (e.g., systolic register file 1840, GRF 1824, data cache / shared local memory 1806A-1806N) for further processing by other functional units of the graphics processor or for output to memory.

[0274] In some embodiments, matrix engine 1803 is configured to support input sparsity, where multiplication operations on sparse regions of input data can be bypassed by skipping multiplication operations with zero-valued operands. In one embodiment, processing elements 1852AA-1852MN are configured to skip the execution of certain operations with zero-valued inputs. In one embodiment, sparsity within the input matrix can be detected, and operations with known zero output values ​​can be bypassed before being submitted to processing elements 1852AA-1852MN. Loading zero-valued operands into processing elements can be bypassed, and processing elements 1852AA-1852MN can be configured to perform multiplication on non-zero-valued input elements. Matrix engine 1803 can also be configured to support output sparsity, so that operations with predetermined zero results are bypassed. In one embodiment, metadata is provided to processing elements 1852AA-1852MN regarding input sparsity and / or output sparsity to indicate which processing elements and / or data channels will be active during a given processing cycle.

[0275] In one embodiment, matrix engine 1803 includes hardware for realizing the operation of sparse data with the compressed representation of sparse matrix, and this sparse matrix stores non-zero value and the metadata of the position of this non-zero value in the matrix of identification.Exemplary compressed representation includes but is not limited to compressed tensor representation, such as, compressed sparse row (CSR) representation, compressed sparse column (CSC) representation, compressed sparse fiber (compressed sparse fiber, CSF) representation.Support for compressed representation enables operation to be performed on the input in compressed tensor format without the need for compressed representation to be decompressed or decoded.In such embodiment, operation can be performed only on non-zero input value, and the non-zero output value obtained can be mapped into output matrix.In some embodiments, hardware support for machine-specific lossless data compression format is also provided, and these machine-specific lossless data compression formats are used when transmitting data in hardware or transmitting data across system bus. Such data may be retained in a compressed format for sparse input data, and the matrix engine 1803 may use compression metadata for the compressed data to enable operations to be performed only on non-zero values ​​or to enable blocks of zero data input to be bypassed for multiplication operations.

[0276] In various embodiments, the input data may be provided by the programmer in a compressed tensor representation, or the codec may compress the input data into a compressed tensor representation or another sparse data encoding. In addition to supporting compressed tensor representation, streaming compression of the sparse input data may be performed before the sparse input data is provided to the processing elements 1852AA-1852MN. In one embodiment, compression is performed on data written to a cache memory associated with the graphics core cluster 1800, where the compression is performed using an encoding supported by the matrix engine 1803. In one embodiment, the matrix engine 1803 includes support for inputs with structured sparsity, in which a predetermined level or predetermined pattern of sparsity is imposed on the input data. The data may be compressed to a known compression ratio, where the compressed data is processed by the processing elements 1852AA-1852MN based on metadata associated with the compressed data.

[0277] Figure 19 A slice 1900 of a multi-slice processor is shown according to one embodiment. In one embodiment, the slice 1900 represents Figure 17A Graphics engine chips 1710A-1710D or Figure 17B Slice 1900 of a multi-slice graphics processor includes an array of graphics core clusters (e.g., graphics core cluster 1800A, graphics core cluster 1800B, through graphics core cluster 1800N), each of which includes an array of graphics cores 515A-515N. Slice 1900 also includes a global dispatcher 1902 for dispatching threads to processing resources of slice 1900.

[0278] Slice 1900 may include or be coupled with an L3 cache 1906 and a memory 1910. In various embodiments, L3 cache 1906 may be excluded, or slice 1900 may include additional levels of cache, such as an L4 cache. In one embodiment, such as Figure 17A and Figure 17B, each instance of slice 1900 in a multi-chip graphics processor is associated with memory 1910. In one embodiment, the multi-chip processor can be configured as a multi-chip module in which the L3 cache 1906 and / or memory 1910 reside on a separate chiplet that is distinct from the graphics core clusters 1800A-1800N. In this context, a chiplet is an at least partially packaged integrated circuit that includes different logic units that can be assembled into a larger package with other chiplets. For example, the L3 cache 1906 can be included in a dedicated cache chiplet, or reside on the same chiplet as the graphics core clusters 1800A-1800N. In one embodiment, the L3 cache 1906 can be included in an active base die or an active interposer.

[0279] Memory fabric 1903 enables communication between graphics core clusters 1800A-1800N, L3 cache 1906, and memory 1910. L2 cache 1904 is coupled to memory fabric 1903 and is configurable to cache transactions executed via memory fabric 1903. Slice interconnect 1908 enables communication with other slices on the graphics processor and may be Figure 17A and Figure 17B 1723F. In embodiments where L3 cache 1906 is excluded from slice 1900, L2 cache 1904 may be configured as a combined L2 / L3 cache. Memory structure 1903 may be configured to route data to L3 cache 1906 or to a memory controller associated with memory 1910 based on the presence or absence of L3 cache 1906 in a particular implementation. L3 cache 1906 may be configured as a per-tile cache that is dedicated to the processing resources of slice 1900 or may be part of a GPU-wide L3 cache.

[0280] Figure 20 is a block diagram illustrating a graphics processor instruction format 2000. A graphics processor execution unit supports an instruction set having instructions in a variety of formats. Solid-line boxes illustrate components that are typically included in an execution unit instruction, while dashed lines include components that are optional or included only in a subset of instructions. In some embodiments, the graphics processor instruction format 2000 described and illustrated is a macroinstruction, as it is an instruction supplied to the execution unit, as opposed to micro-operations that result from instruction decoding once the instruction is processed. Thus, a single instruction can cause the hardware to execute multiple micro-operations.

[0281] As described herein, the graphics processor execution unit can natively support instructions in the 128-bit instruction format 2010. Depending on the selected instruction, instruction options, and number of operands, a compact 64-bit instruction format 2030 can be used for some instructions. The native 128-bit instruction format 2010 provides access to all instruction options, while some options and operations are restricted in the 64-bit instruction format 2030. The native instructions available in the 64-bit instruction format 2030 vary depending on the embodiment. Instructions are partially compressed using a set of index values ​​in the index field 2013. The execution unit hardware references a set of compression tables based on the index values ​​and uses the compression table output to reconstruct the native instruction in the 128-bit instruction format 2010. Instructions of other sizes and formats may be used.

[0282] For each format, the instruction opcode 2012 defines the operation to be performed by the execution unit. The execution unit executes each instruction in parallel across multiple data elements of each operand. For example, in response to an addition instruction, the execution unit performs a synchronous addition operation across each color channel representing a texture element or picture element. By default, the execution unit executes each instruction across all data channels of the operand. The instruction control field 2014 may enable control of certain execution options such as channel selection (e.g., predication) and data channel order (e.g., swizzle). For instructions in the 128-bit instruction format 2010, the execution size field 2016 limits the number of data channels that will be executed in parallel. The execution size field 2016 may not be available for the compact 64-bit instruction format 2030.

[0283] Some execution unit instructions have up to three operands, including two source operands src0 2020 and src1 2022 and one destination operand (dest 2018). Other instructions (such as, for example, data manipulation instructions, dot product instructions, multiply-add instructions, or multiply-accumulate instructions) may have a third source operand (e.g., src2 2024). The instruction opcode 2012 determines the number of source operands. The last source operand of an instruction may be an immediate (e.g., hard-coded) value passed with the instruction. The execution unit may also support multiple destination instructions, where one or more of the destinations are implicit or implicit based on the instruction and / or the specified destination.

[0284] The 128-bit instruction format 2010 may include an access / addressing mode field 2026 that specifies, for example, whether direct register addressing mode or indirect register addressing mode is used. When direct register addressing mode is used, the register address of one or more operands is provided directly by bits in the instruction.

[0285] The 128-bit instruction format 2010 may also include an access / addressing mode field 2026 that specifies the addressing mode and / or access mode of the instruction. The access mode may be used to define the data access alignment of the instruction. Access modes including a 16-byte aligned access mode and a 1-byte aligned access mode may be supported, wherein the byte alignment of the access mode determines the access alignment of the instruction operands. For example, when in the first mode, the instruction may use byte-aligned addressing for source and destination operands, and when in the second mode, the instruction may use 16-byte aligned addressing for all source and destination operands.

[0286] The addressing mode portion of the access / addressing mode field 2026 may determine whether the instruction uses direct or indirect addressing. When direct register addressing mode is used, the bits in the instruction directly provide the register address of one or more operands. When indirect register addressing mode is used, the register address of one or more operands may be calculated based on the address register value and the address immediate field in the instruction.

[0287] Instructions can be grouped based on the instruction opcode 2012 bit fields to simplify opcode decoding 2040. For 8-bit opcodes, bit 4, bit 5, and bit 6 allow execution units to determine the type of opcode. The exact opcode grouping shown is only an example. Move and logic opcode group 2042 can include data movement and logic instructions (e.g., move (mov), compare (cmp)). Move and logic opcode group 2042 can share five least significant bits (least significant bit, LSB), wherein the move (mov) instruction adopts the form of 0000xxxxb, and the logic instruction adopts the form of 0001xxxxb. Flow control instruction group 2044 (e.g., call (call), jump (jmp)) includes the instruction of 0010xxxxb (e.g., 0x20) form. Miscellaneous instruction group 2046 includes a mixture of instructions, including synchronization instructions (e.g., wait (wait), send (send)) of 0011xxxxb (e.g., 0x30) form. The parallel math instruction group 2048 includes component-wise arithmetic instructions (e.g., add, multiply (mul)) of the form 0100xxxxb (e.g., 0x40). The parallel math instruction group 2048 performs arithmetic operations in parallel across data channels. The vector math group 2050 includes arithmetic instructions (e.g., dp4) of the form 0101xxxxb (e.g., 0x50). The vector math group performs arithmetic on vector operands, such as dot product calculations. In one embodiment, opcode decode 2040 can be used to determine which portion of the execution unit will be used to execute the decoded instruction. For example, some instructions can be designated as systolic instructions to be executed by a systolic array. Other instructions, such as ray tracing instructions (not shown), can be routed to a ray tracing core or ray tracing logic within a slice or partition of the execution logic. Graphics pipeline

[0288] Figure 21 is a block diagram of a graphics processor 2100 according to another embodiment. Figure 21 Elements having the same or similar names as elements of any other figures herein describe the same elements as those in the other figures, can operate or function in a similar manner as those in the other figures, may include the same components, and may be linked to other entities such as, but not limited to, those described elsewhere herein.

[0289] The graphics processor 2100 may include various types of graphics processing pipelines, such as a geometry pipeline 2120, a media pipeline 2130, a display engine 2140, thread execution logic 2150, and a render output pipeline 2170. The graphics processor 2100 may be a graphics processor within a multi-core processing system that includes one or more general-purpose processing cores. The graphics processor may be controlled by register writes to one or more control registers (not shown) or by commands issued to the graphics processor 2100 via a ring or mesh interconnect 2102. The ring or mesh interconnect 2102 may couple the graphics processor 2100 to other processing components (such as other graphics processors or general-purpose processors). Commands from the ring or mesh interconnect 2102 are interpreted by a command stream converter 2103, which supplies instructions to various components of the geometry pipeline 2120 or the media pipeline 2130.

[0290] The command stream converter 2103 may direct the operation of the vertex fetcher 2105, which reads vertex data from memory and executes vertex processing commands provided by the command stream converter 2103. The vertex fetcher 2105 may provide the vertex data to the vertex shader 2107, which performs coordinate space transformation and lighting operations on each vertex. The vertex fetcher 2105 and the vertex shader 2107 may execute the vertex processing instructions by dispatching execution threads to the graphics cores 2152A-2152B via the thread dispatcher 2131.

[0291] Graphics cores 2152A-2152B may be arrays of vector processors with instruction sets for performing graphics and media operations. Graphics cores 2152A-2152B may have an attached L1 cache 2151 that is dedicated to each array or shared between arrays. The cache may be configured as a data cache, an instruction cache, or a single cache partitioned to contain data and instructions in different partitions.

[0292] The geometry pipeline 2120 may include a tessellation component for performing hardware-accelerated tessellation of 3D objects. The programmable hull shader 2111 may configure tessellation operations. The programmable domain shader 2117 may provide back-end evaluation of the tessellation output. The tessellator 2113 may operate under the direction of the programmable hull shader 2111 and may contain specialized logic for generating a detailed set of geometric objects based on a coarse geometric model provided as input to the geometry pipeline 2120. Furthermore, if tessellation is not used, the tessellation components (e.g., the programmable hull shader 2111, the tessellator 2113, and the programmable domain shader 2117) may be bypassed. The tessellation components may operate based on data received from the vertex shader 2107.

[0293] The complete geometric object may be processed by the geometry shader 2119 via one or more threads dispatched to the graphics cores 2152A-2152B, or may proceed directly to the clipper 2129. The geometry shader may operate on entire geometric objects, rather than on vertices or patches of vertices as in previous stages of the graphics pipeline. If tessellation is disabled, the geometry shader 2119 receives input from the vertex shader 2107. The geometry shader 2119 may be programmable by a geometry shader program to perform geometry tessellation when the tessellation unit is disabled.

[0294] Before rasterization, the clipper 2129 processes the vertex data. The clipper 2129 can be a fixed-function clipper or a programmable clipper with clipping and geometry shader functionality. The rasterizer and depth test component 2173 in the render output pipeline 2170 can dispatch a pixel shader to convert geometric objects into a pixel-by-pixel representation. The pixel shader logic can be included in the thread execution logic 2150. Alternatively, the application can bypass the rasterizer and depth test component 2173 and access the unrasterized vertex data via the outflow unit 2123.

[0295] The graphics processor 2100 has an interconnect bus, interconnect structure, or some other interconnect mechanism that allows data and messages to be passed between the main components of the processor. In some embodiments, the graphics cores 2152A-2152B and associated logic units (e.g., L1 cache 2151, samplers 2154, texture cache 2158, etc.) are interconnected via data ports 2156 to perform memory accesses and communicate with the processor's rendering output pipeline components. The samplers 2154, L1 cache 2151, texture cache 2158, and graphics cores 2152A-2152B can each have a separate memory access path. Optionally, the texture cache 2158 can also be configured as a sampler cache.

[0296] The render output pipeline 2170 may include a rasterizer and depth test component 2173 that converts vertex-based objects into associated pixel-based representations. The rasterizer logic may include a windower / masker unit for performing fixed-function triangle and line rasterization. In some embodiments, an associated render buffer 2178 and depth buffer 2179 are also available. A pixel operation component 2177 performs pixel-based operations on data, but in some instances, pixel operations associated with 2D operations (e.g., using hybrid bit-block image transfer) are performed by the 2D engine 2141 or, when displayed, by the display controller 2143 using an overlay display plane instead. A shared L3 cache 2175 may be available to all graphics components, allowing data to be shared without using main system memory.

[0297] The media pipeline 2130 may include a media engine 2137 and a video front end 2134. The video front end 2134 may receive pipeline commands from the command streamer 2103. The media pipeline 2130 may include a separate command streamer. The video front end 2134 may process media commands before sending them to the media engine 2137. The media engine 2137 may include thread generation functionality for generating threads for dispatching to the thread execution logic 2150 via the thread dispatcher 2131.

[0298] The graphics processor 2100 may include a display engine 2140. The display engine 2140 may be external to the graphics processor 2100 and may be coupled to the graphics processor via a ring or mesh interconnect 2102, or some other interconnect bus or structure. The display engine 2140 may include a 2D engine 2141 and a display controller 2143. The display engine 2140 may contain dedicated logic capable of operating independently of the 3D pipeline. The display controller 2143 may be coupled to a display device (not shown), which may be a system-integrated display device, such as in a laptop, or an external display device attached via a display device connector.

[0299] The geometry pipeline 2120 and media pipeline 2130 can be configured to perform operations based on multiple graphics and media programming interfaces and are not dedicated to any one application programming interface (API). Driver software for the graphics processor can convert API calls dedicated to a specific graphics or media library into commands that can be processed by the graphics processor. Support can be provided for all of the Open Graphics Library (OpenGL), Open Computing Language (OpenCL), and / or Vulkan graphics and computing APIs from the Khronos Group. Support can also be provided for the Direct3D library from Microsoft. Combinations of these libraries can be supported. Support can also be provided for the Open Source Computer Vision Library (OpenCV). If a mapping can be performed from the pipeline of a future API to the pipeline of the graphics processor, future APIs with compatible 3D pipelines will also be supported. Graphics pipeline programming

[0300] Figure 22A is a block diagram illustrating a graphics processor command format 2200 for programming a graphics processing pipeline, such as, for example, the graphics processing pipeline described herein in conjunction with Figure 16 and Figure 21 Describe the pipeline. Figure 22Bis a block diagram illustrating a graphics processor command sequence 2210 according to one embodiment. Figure 22A Solid-line boxes in show components that are generally included in the graphics commands, while dashed lines include components that are optional or included only in a subset of the graphics commands. Figure 22A The exemplary graphics processor command format 2200 includes fields for identifying the client of the command 2202, a command operation code (opcode) 2204, and a data field 2206. A sub-opcode 2205 and a command size 2208 are also included in some commands.

[0301] Client 2202 may specify a client unit of a graphics device that processes command data. A graphics processor command parser may examine the client field of each command to adjust further processing of the command and route the command data to the appropriate client unit. A graphics processor client unit may include a memory interface unit, a rendering unit, a 2D unit, a 3D unit, and a media unit. Each client unit may have a corresponding processing pipeline for processing commands. Once a command is received by a client unit, the client unit reads the opcode 2204 and sub-opcode 2205 (if present) to determine the operation to be performed. The client unit uses the information in the data field 2206 to execute the command. For some commands, expected command size 2208 explicitly specifies the size of the command. The command parser may automatically determine the size of at least some of the commands based on the command opcode. Commands may be aligned to multiples of double words. Other command formats may also be used.

[0302] Figure 22B 22 illustrates a graphics processor command sequence 2210. Software or firmware of a data processing system featuring an exemplary graphics processor can use a version of the illustrated command sequence to establish, execute, and terminate a set of graphics operations. The sample command sequence is shown and described for illustrative purposes only and is not limited to these specific commands or command sequences. Furthermore, commands can be issued as batches in the command sequence so that the graphics processor will process the command sequence at least partially concurrently.

[0303] Graphics processor command sequence 2210 may begin with a pipeline flush command 2212 to cause any active graphics pipeline to complete currently pending commands for that pipeline. Optionally, 3D pipeline 2222 and media pipeline 2224 may not operate concurrently. A pipeline flush is performed to cause active graphics pipelines to complete any pending commands. In response to a pipeline flush, the command parser for the graphics processor will halt command processing until the active drawing engines complete pending operations and the associated read buffers are invalidated. Optionally, any data marked as "dirty" in the render buffer may be flushed to memory. Pipeline flush command 2212 may be used for pipeline synchronization or before placing the graphics processor into a low-power state.

[0304] When a command sequence requires the graphics processor to explicitly switch between pipelines, a pipeline select command 2213 may be used. A pipeline select command 2213 may be requested only once in an execution context before issuing a pipeline command, unless the context is issuing commands for both pipelines. A pipeline flush command 2212 may be requested immediately before a pipeline switch via a pipeline select command 2213.

[0305] Pipeline control commands 2214 can configure the graphics pipeline for operation and can be used to program the 3D pipeline 2222 and the media pipeline 2224. Pipeline control commands 2214 can configure the pipeline state for the active pipeline. Pipeline control commands 2214 can be used for pipeline synchronization and to flush data from one or more cache memories within the active pipeline before processing a batch of commands.

[0306] Commands associated with return buffer status 2216 may be used to configure a set of return buffers for a corresponding pipeline to write data. Some pipeline operations require the allocation, selection, or configuration of one or more return buffers into which the operation writes intermediate data during processing. A graphics processor may also use one or more return buffers to store output data and perform cross-thread communication. Return buffer status 2216 may include selecting the size and number of return buffers to be used for a set of pipeline operations.

[0307] The remaining commands in the command sequence differ based on the active pipeline for the operation. Based on pipeline decision 2220 , the command sequence is tailored for the 3D pipeline 2222 starting at 3D pipeline state 2230 or the media pipeline 2224 starting at media pipeline state 2240 .

[0308] The commands used to configure the 3D pipeline state 2230 include 3D state setup commands for vertex buffer state, vertex element state, constant color state, depth buffer state, and other state variables that will be configured before processing 3D primitive commands. The values ​​of these commands are determined at least in part based on the specific 3D API in use. The 3D pipeline state 2230 commands can also selectively disable or bypass certain pipeline elements if those elements will not be used.

[0309] 3D primitive 2232 commands can be used to submit 3D primitives for processing by the 3D pipeline. The commands and associated parameters passed to the graphics processor via the 3D primitive 2232 commands are forwarded to the vertex acquisition function in the graphics pipeline. The vertex acquisition function uses the 3D primitive 2232 command data to generate vertex data structures. The vertex data structures are stored in one or more return buffers. The 3D primitive 2232 commands can be used to perform vertex operations on the 3D primitives via the vertex shader. To process the vertex shader, the 3D pipeline 2222 dispatches the shader execution thread to the graphics processor execution unit.

[0310] The 3D pipeline 2222 can be triggered via an execute 2234 command or event. Registers can be written to trigger command execution. Execution can be triggered via a "go" or "kick" command in a command sequence. Command execution can be triggered using pipeline synchronization commands to flush the command sequence through the graphics pipeline. The 3D pipeline will perform geometry processing for 3D primitives. Once the operation is complete, the resulting geometry is rasterized and the pixel engine shades the resulting pixels. Additional commands for controlling pixel shading and pixel backend operations may also be included for those operations.

[0311] When performing media operations, the graphics processor command sequence 2210 may follow the media pipeline 2224 path. Generally speaking, the specific purpose and manner of programming the media pipeline 2224 depends on the media or compute operation to be performed. During media decoding, certain media decoding operations may be migrated to the media pipeline. The media pipeline may also be bypassed and the media decoding may be performed in whole or in part using resources provided by one or more general-purpose processing cores. The media pipeline may also include elements for general-purpose graphics processor unit (GPGPU) operations, where the graphics processor is configured to perform SIMD vector operations using compute shader programs that are not explicitly related to the rendering of graphics primitives.

[0312] The media pipeline 2224 can be configured in a similar manner to the 3D pipeline 2222. A set of commands for configuring the media pipeline state 2240 is dispatched or placed into a command queue before the media object commands 2242. The commands for the media pipeline state 2240 may include data for configuring the media pipeline elements that will be used to process the media objects. This includes data for configuring the video decoding and video encoding logic within the media pipeline, such as encoding or decoding formats. The commands for the media pipeline state 2240 may also support the use of one or more pointers to "indirect" state elements that contain batches of state settings.

[0313] Media object commands 2242 may supply a pointer to a media object for processing by the media pipeline. The media object includes a memory buffer containing the video data to be processed. Optionally, all media pipeline states must be valid before issuing a media object command 2242. Once the pipeline state is configured and the media object command 2242 is queued, the media pipeline 2224 is triggered via an execute command 2244 or an equivalent execute event (e.g., a register write). The output from the media pipeline 2224 may then be post-processed by operations provided by the 3D pipeline 2222 or the media pipeline 2224. GPGPU operations may be configured and executed in a manner similar to media operations. Graphics software architecture

[0314] Figure 23 An exemplary graphics software architecture for data processing system 2300 is shown. Such a software architecture may include a 3D graphics application 2310, an operating system 2320, and a processor 2330. Processor 2330 may include a graphics processor 2332 and one or more general-purpose processor cores 2334. Processor 2330 may be a variant of, and may be used in place of, one of processor(s) 1402 or any other of the processors described herein. Thus, disclosure of any feature in conjunction with processor(s) 1402 or any other of the processors described herein also discloses the corresponding combination with graphics processor 2332, but is not limited thereto. Figure 23 Elements having the same or similar names as elements in any other figures herein describe the same elements in the other figures, can operate or function in a similar manner to the other figures, can include the same components, and can be linked to other entities such as, but not limited to, those described elsewhere herein. 3D graphics application 2310 and operating system 2320 each execute in system memory 2350 of the data processing system.

[0315] 3D graphics application 2310 may include one or more shader programs that include shader instructions 2312. The shader language instructions may be in a high-level shader language, such as Direct3D's High-Level Shader Language (HLSL), OpenGL Shader Language (GLSL), or the like. The application may also include executable instructions 2314 in a machine language suitable for execution by a general-purpose processor core(s) 2334. The application may also include graphics objects 2316 defined by vertex data.

[0316] The operating system 2320 may be from Microsoft Corporation. An operating system, a proprietary UNIX-like operating system, or an open source UNIX-like operating system using a variant of the Linux kernel. The operating system 2320 may support a graphics API 2322, such as the Direct3D API, OpenGL API, or Vulkan API. When the Direct3D API is in use, the operating system 2320 uses a front-end shader compiler 2324 to compile any shader instructions 2312 using HLSL into a lower-level shader language. The compilation may be just-in-time (JIT) compilation or application executable shader precompilation. During the compilation of the 3D graphics application 2310, high-level shaders may be compiled into low-level shaders. The shader instructions 2312 may be provided in an intermediate form, such as a version of the Standard Portable Intermediate Representation (SPIR) used by the Vulkan API.

[0317] The user-mode graphics driver 2326 may include a backend shader compiler 2327 to compile shader instructions 2312 into a hardware-specific representation. When the OpenGL API is in use, shader instructions 2312 in the GLSL high-level language are passed to the user-mode graphics driver 2326 for compilation. The user-mode graphics driver 2326 may use operating system kernel-mode functions 2328 to communicate with the kernel-mode graphics driver 2329. The kernel-mode graphics driver 2329 may communicate with the graphics processor 2332 to dispatch commands and instructions. IP core implementation

[0318] One or more aspects may be implemented by representative code stored on a machine-readable medium that represents and / or defines logic within an integrated circuit (such as a processor). For example, a machine-readable medium may include instructions representing various logic within a processor. When read by a machine, the instructions may cause the machine to fabricate logic for performing the techniques described herein. Such representations (referred to as "IP cores") are reusable units of logic for an integrated circuit that can be stored on a tangible, machine-readable medium as a hardware model describing the structure of the integrated circuit. The hardware model may be supplied to each customer or manufacturing facility that loads the hardware model on a manufacturing machine that manufactures the integrated circuit. The integrated circuit may be manufactured so that the circuit performs the operations described in association with any of the embodiments described herein.

[0319] Figure 24 2 is a block diagram illustrating an IP core development system 2400 that can be used to manufacture an integrated circuit to perform operations according to one embodiment. The IP core development system 2400 can be used to generate modular, reusable designs that can be incorporated into larger designs or used to build entire integrated circuits (e.g., SoC integrated circuits). A design facility 2430 can generate a software simulation 2410 of the IP core design in a high-level programming language (e.g., C / C++). The software simulation 2410 can be used to design, test, and verify the behavior of the IP core using a simulation model 2412. The simulation model 2412 may include functional simulation, behavioral simulation, and / or timing simulation. A register transfer level design (RTL design 2415) can then be created or synthesized from the simulation model 2412. The RTL design 2415 is an abstraction of the behavior of the integrated circuit (including associated logic executed using the modeled digital signals) that models the flow of digital signals between hardware registers. In addition to the RTL design 2415, lower-level designs at the logic or transistor level can also be created, designed, or synthesized. Therefore, the specific details of the initial design and simulation may vary.

[0320] The RTL design 2415 or an equivalent solution can be further synthesized by the design facility into a hardware model 2420, which can be in a hardware description language (HDL) or some other representation of physical design data. The HDL can be further simulated or tested to verify the IP core design. Non-volatile memory 2440 (e.g., a hard disk, flash memory, or any non-volatile storage medium) can be used to store the IP core design for delivery to a manufacturing facility 2465. The manufacturing facility 2465 can be a third-party manufacturing facility. Alternatively, the IP core design can be transmitted via a wired connection 2450 or a wireless connection 2460 (e.g., via the Internet). The manufacturing facility 2465 can then manufacture an integrated circuit based at least in part on the IP core design. The manufactured integrated circuit can be configured to perform operations according to at least one embodiment described herein.

[0321] Figure 25A A cross-sectional side view of a package assembly 2590 of an integrated circuit comprising hardware logic chiplets of multiple units connected to a substrate 2580 (e.g., a base die) is shown. Graphics processing units, parallel processors, and / or compute accelerators as described herein may be composed of various silicon chiplets that are fabricated separately. In this context, a chiplet is an at least partially packaged integrated circuit that includes different logic units that can be assembled into a larger package with other chiplets. Chiplets with various sets of different IP core logic can be assembled into a single device. In addition, chiplets can be integrated into a base die or base chiplet using active interposer technology. The concepts described herein enable interconnection and communication between different forms of IP within a GPU. IP cores can be manufactured using different process technologies and constructed during manufacturing, which avoids the complexity of converging multiple IP into the same manufacturing process (especially on large SoCs with several flavors of IP). Allowing the use of multiple process technologies improves time to market and provides a cost-effective way to create multiple product SKUs. Furthermore, disaggregated IP can be more easily modified to be independently power-gated; components not in use for a given workload can be shut down, reducing overall power consumption.

[0322] In various embodiments, the package assembly 2590 can include a fewer or greater number of components and chiplets interconnected by an interconnect structure 2585 or a bridge structure 2587. The bridge structure 2587 can be used to facilitate point-to-point interconnection between, for example, a logic or I / O chiplet 2574 and a memory chiplet 2575. In some implementations, the bridge structure 2587 can also be embedded within the substrate 2580. The chiplets within the package assembly 2590 can have a 2.5D arrangement using chip-on-wafer-on-substrate (CoWoS) stacking, where multiple dies are stacked side-by-side on a silicon interposer that includes through-silicon vias (TSVs) to couple the chiplets to the substrate 2580, which includes electrical connections to the package interconnects 2583.

[0323] In one embodiment, the silicon interposer is an active interposer 2589 that includes embedded logic in addition to the TSVs. In such an embodiment, the chiplets within the package assembly 2590 are arranged on top of the active interposer 2589 using 3D face-to-face die stacking. In addition to the interconnect structure 2585 and the bridge structure 2587, the active interposer 2589 may also include I / O hardware logic 2591, cache memory 2592, and other hardware logic 2593. The interconnect structure 2585 enables communication between the various logic chiplets within the active interposer 2589. The interconnect structure 2585 may be a NoC interconnect or another form of packet-switched fabric that exchanges data packets between components of the package assembly. For complex components, the interconnect structure 2585 may be a dedicated chiplet that enables communication between the various hardware logic components of the package assembly 2590.

[0324] The hardware logic chiplets may include dedicated hardware logic chiplets 2572, logic or I / O chiplets 2574, and / or memory chiplets 2575. The dedicated hardware logic chiplets 2572 and logic or I / O chiplets 2574 may be implemented at least in part in configurable logic or fixed-functionality logic hardware and may include one or more portions of any of the processor core(s), graphics processor(s), parallel processor(s), or other accelerator devices described herein. The memory chiplets 2575 may be DRAM (e.g., GDDR, HBM) memory or cache (SRAM) memory. The cache memory 2592 within the active interposer 2589 (or substrate 2580) may serve as a global cache for the package assembly 2590, as part of a distributed global cache, or as a dedicated cache for the interconnect structure 2585.

[0325] Each chiplet can be fabricated as a separate semiconductor die and can be coupled to a base die that is embedded within or coupled to a substrate 2580. Coupling to the substrate 2580 can be performed via an interconnect structure 2573. The interconnect structure 2573 can be configured to route electrical signals between the various chiplets and the logic within the substrate 2580. The interconnect structure 2573 can include interconnects such as, but not limited to, bumps or pillars. In some embodiments, the interconnect structure 2573 can be configured to route electrical signals such as, for example, input / output (I / O) signals and / or power or ground signals associated with the operation of the logic, I / O, and memory chiplets. In one embodiment, an additional interconnect structure couples the active interposer 2589 to the substrate 2580.

[0326] Substrate 2580 may be an epoxy-based laminate substrate and / or may also include other suitable types of substrates. Package assembly 2590 may be connected to other electrical devices via package interconnects 2583. Package interconnects 2583 may be coupled to the surface of substrate 2580 to route electrical signals to other electrical devices, such as a motherboard, other chipsets, or multi-chip modules.

[0327] The logic or I / O chiplet 2574 and the memory chiplet 2575 can be electrically coupled via a bridge structure 2587 that is configured to route electrical signals between the logic or I / O chiplet 2574 and the memory chiplet 2575. The bridge structure 2587 can be a dense interconnect structure that provides routing for electrical signals. The bridge structure 2587 can include a bridge substrate composed of glass or a suitable semiconductor material. Circuit features can be formed on the bridge substrate to provide chip-to-chip connections between the logic or I / O chiplet 2574 and the memory chiplet 2575. The bridge structure 2587 can also be referred to as a silicon bridge or an interconnect bridge. For example, the bridge structure 2587 is an embedded multi-die interconnect bridge (EMIB). Alternatively, the bridge structure 2587 can simply be a direct connection from one chiplet to another.

[0328] Figure 25BA package assembly 2594 including interchangeable chiplets 2595 is shown according to one embodiment. The interchangeable chiplets 2595 can be assembled into standardized chiplet slots or chiplet sockets on base chiplets 2596, 2598. The base chiplets 2596, 2598 can be coupled via a bridge interconnect 2597, which can be similar to other bridge interconnects described herein and can be, for example, EMIB. Memory chiplets can also be connected to logic or I / O chiplets via the bridge interconnect. The I / O and logic chiplets can communicate via the interconnect structure. The base chiplets can each support one or more slots in a standardized format for one of logic or I / O or memory / cache.

[0329] The SRAM and power delivery circuit modules can be fabricated into one or more of the base chiplets 2596, 2598, which can be fabricated using a different process technology than the interchangeable chiplets 2595, which are stacked on top of the base chiplets. For example, the base chiplets 2596, 2598 can be fabricated using a larger process technology while the interchangeable chiplets can be fabricated using a smaller process technology. One or more of the interchangeable chiplets 2595 can be memory (e.g., DRAM) chiplets. Different memory densities can be selected for the package assembly 2594 based on the power and / or performance requirements of the product in which the package assembly 2594 is to be used. Additionally, logic chiplets with different numbers of functional units of different types can be selected at assembly time based on the power and / or performance requirements of the product. Furthermore, chiplets containing IP logic cores of different types can be inserted into the interchangeable chiplet slots, enabling hybrid processor designs that can mix and match IP blocks of different technologies. Exemplary System-on-Chip Integrated Circuit

[0330] Figure 26 An exemplary integrated circuit that can be manufactured using one or more IP cores is shown. Other logic and circuitry may be included in addition to what is shown, including additional graphics processors / cores, peripheral interface controllers, or general-purpose processor cores. Figure 26 Elements having the same or similar names as elements of any other figures herein describe the same elements as those in the other figures, can operate or function in a similar manner as those in the other figures, may include the same components, and may be linked to other entities such as, but not limited to, those described elsewhere herein.

[0331] System-on-chip integrated circuit 2600 includes one or more application processors 2605 (e.g., CPUs), a graphics processor 2610, which may be a variant of graphics processor(s) 1408, or may be a variant of any graphics processor described herein and may be used in place of any of the described graphics processors. Thus, any feature disclosed herein in conjunction with a graphics processor also discloses the corresponding combination with graphics processor 2610, but is not limited thereto. System-on-chip integrated circuit 2600 may additionally include an image processor 2615 and / or a video processor 2620, either of which may be modular IP cores from the same design facility or from multiple different design facilities. System-on-chip integrated circuit 2600 may include peripheral or bus logic, including a USB controller 2625, a UART controller 2630, an SPI / SDIO controller 2635, and an I2S / I2C controller 2640. In addition, the integrated circuit may include a display device 2645 coupled to a high-definition multimedia interface (HDMI) controller 2650 and one or more of a reliability, availability, and serviceability engine (RAS engine 2655). The RAS engine 2655 is used to identify potential faults that may occur during device runtime to minimize downtime if those potential faults occur. Storage may be provided by a flash memory subsystem 2660 (including flash memory and a flash memory controller). A memory interface may be provided via a memory controller 2665 to obtain access to SDRAM or SRAM memory devices. Some integrated circuits additionally include an embedded security engine 2670. Graphics core cluster with asynchronous operation support

[0332] Figure 27 A graphics processing system 2700 including a graphics core cluster 2701 is shown according to an embodiment. The graphics core cluster 2701 includes a set of graphics cores 2702A-2702D interconnected by core-to-core data links. The graphics cores 2702A-2702D may include any processing resources or graphics core architectures described herein. For example, the graphics core cluster 2701 may be configured to and / or include Figure 2C The processing cluster 214 and / or Figure 18A In one embodiment, graphics cores 2702A-2702D include Figure 15C In one embodiment, graphics cores 2702A-2702D may individually or collectively include: Figure 2D Graphics multiprocessor 234, Figure 2E Graphics multiprocessor 235, Figure 15BGraphics cores 1521A-1521F and / or Figure 18A Components of graphics cores 1815A-1815N.

[0333] Taking graphics core 0 (graphics core 2702A) as an exemplary graphics core, each of graphics cores 2702A-2702D includes an addressable barrier unit 2704, an SLM 2706, an asynchronous DMA unit 2708, an L1 cache 2711, an asynchronous MMA unit 2712, and processing resource threads 2714A-2714N. In one embodiment, each of processing resource threads 2714A-2714N is a hardware thread configured as a processing resource and / or execution unit of a scalar or vector processing unit. The processing unit can be configured to execute one or more SIMD data channels and / or processing channels. The processing unit can also be configured to execute multiple SIMT threads via the SIMD processing channels. In one embodiment, addressable barrier unit 2704, asynchronous DMA unit 2708, and asynchronous MMA unit 2712 are hardware units that include circuit modules configured to implement addressable barriers, asynchronous direct memory access, and asynchronous matrix multiplication and accumulation operations, respectively.

[0334] The addressable barrier unit 2704 enables hardware synchronization between processing resource threads 2714A-2714N and between processing resource threads 2714A-2714N and asynchronous MMA unit 2712 to coordinate asynchronous computation and data movement operations. In one embodiment, the addressable barrier unit 2704 provides a non-blocking barrier synchronization mechanism to enable processing resource threads 2714A-2714N to synchronize on asynchronous tensor transfers. Addressable barriers can be specified in one of multiple possible scopes, including workgroup and cluster-wide scopes. Addressable barriers are a type of cluster barrier used to synchronize asynchronous operations. Traditional barriers allow threads within a thread group to synchronize and are only released when all threads reach the barrier, while cluster barriers are thread group-based barriers that allow the entire thread group to synchronize and are only released when all thread groups reach the barrier.

[0335] Asynchronous DMA unit 2708 facilitates asynchronous data transfers by processing resource threads 2714A-2714N, which frees up processing resources associated with those threads previously used to transfer data between SLM 2706 within graphics cores 2702A-2702D and global memory 2710. Asynchronous DMA unit 2708 also frees up register file resources by loading data directly from global memory into SLM 2706, thereby bypassing intermediate storage within the register file of the respective graphics core 2702A-2702D.

[0336] The asynchronous MMA unit 2712 facilitates asynchronous matrix multiplication and accumulation operations via the matrix and / or tensor processing resources of the graphics core cluster 2701. The matrix and / or tensor processing resources may include the tensor cores 264, 371 and / or the matrix engine 1803, as described herein. In one embodiment, the asynchronous MMA unit 2712 includes matrix and / or tensor processing resources. In one embodiment, the asynchronous MMA unit 2712 facilitates asynchronous execution of matrix multiplication and accumulation operations via the matrix and / or tensor processing resources included within the graphics cores 2702A-2702D.

[0337] Graphics core cluster 2701 is connected to global memory 2710, which may include GDDR, HBM, and / or other types of memory suitable for use in a graphics processor (e.g., parallel processor memory 222, memory 366, GPU memory 420-423, graphics memory 433-434, ..., M, memory 714A-714B, memory 1726A-1726D, memory 1910, etc.). In one embodiment, global memory 2710 may also include system memory (e.g., system memory 104, memory 366, processor memory 401-402) mapped into the virtual address space used by graphics core cluster 2701. Graphics core cluster 2701 includes a distributed SLM 2716, where the SLM 2706 within each graphics core 2702A-2702D is interconnected and accessible to other graphics cores in the cluster. The shared local memory within each graphics core 2702A-2702D can be mapped to an address space commonly accessible to all graphics cores 2702A-2702D in graphics core cluster 2701, where the SLM 2706 of each graphics core 2702A-2702D is assigned to a specific address range within distributed SLM 2716. During cluster execution 2722 on graphics core cluster 2701, each thread group (e.g., workgroup 2724) executes within its assigned graphics core of graphics cores 2702A-2702D. During execution on its assigned graphics core, the thread group can exchange data via distributed SLM 2716. The portion of distributed SLM 2716 physically associated with the graphics core (e.g., SLM 2706) is referred to as the local SLM, and the SLMs of the other graphics cores are referred to as remote SLMs. Communication and synchronization between cooperating thread groups is performed via distributed SLM 2716, rather than through global memory 2710.

[0338] In one embodiment, asynchronous MMA uses an accumulation matrix C' to perform operations D=A*B or D=C+A*B for different data types. In one embodiment, matrix C' is always in FP32, while matrices A and B can have different data types with different computation formats. The dot product result can be output in FP32, FP16, or BF16, with the final accumulation performed within the graphics core's SLM 2706 via in-memory adder logic. GPU asynchronous direct memory access

[0339] The asynchronous DMA unit 2708 of the graphics core cluster 2701 is used to asynchronously move data from global memory 2710 to the SLM 2706, and asynchronously move data from the SLM 2706 on one graphics core to another graphics core for distributed SLM 2716. Asynchronous DMA operations are performed in the background without blocking the issuing thread and do not require data to be stored in the register file. A thread group cluster hierarchy is employed, in which thread groups execute collaboratively across multiple graphics cores. Collaborating graphics cores within a thread group cluster can synchronize and share data without going through the L2 cache.

[0340] The primary motivation for thread group clustering is to reduce the amount of L2 bandwidth required. With thread group clustering, slices are loaded from the L2 cache to a single graphics core and multicast to other thread groups in the cluster using the new graphics core interconnect. Synchronization across the thread groups in the cluster is required to ensure that all thread groups in the cluster have arrived at the same point in the kernel and are ready to receive data, and a second synchronization point is needed to notify the thread groups that data is available. Traditional barriers allow threads within a thread group to synchronize and are only released when all threads have reached the barrier. The thread group cluster barriers described in this paper allow an entire thread group to synchronize and be released when all thread groups have reached the barrier. Asynchronous DMA also provides a means for evenly distributing L2 accesses across the graphics cores in the graphics core cluster to load balance the L2 interface and maximize available bandwidth.

[0341] Figure 28 A system 2800 is shown that enables data sharing across graphics cores within a graphics core cluster. Graphics cores 2702A-2702D of graphics core cluster 2701 may share a common L2 cache 2802. The L2 cache may be Figure 19 A variant of the L2 cache 1904 and / or including Figure 2B28. The L2 cache 2802 may include various aspects of the L2 cache 221. The L2 cache 2802 may be a distributed cache, with multiple L2 cache nodes 2812A-2812D interconnected via an L2 fabric 2813. The density of the L2 cache nodes 2812A-2812D may vary across embodiments. In one embodiment, the L2 cache 2802 may include multiple L2 cache nodes 2812A-2812D for each graphics core cluster, e.g., each graphics core has an associated L2 cache node. In one embodiment, multiple graphics core clusters share the L2 cache 2802, with each graphics core cluster having at least one L2 cache node.

[0342] Graphics cores 2702A-2702D may be coupled to L2 cache 2802 via an L2 interconnect, which may be associated with a memory crossbar (e.g., memory crossbar 216) or a memory structure (e.g., memory structure 1903). Graphics cores 2702A-2702D may also include cross-core interconnects 2804A-2804C to enable a multicast data sharing mechanism. In various embodiments, cross-core interconnects 2804A-2804C may be a point-to-point interconnect, a data crossbar, or an on-chip network interconnect. Each graphics core 2702A-2702D includes SLMs 2706A-2706D, which may be interconnected into distributed SLMs 2716, such as Figure 27 As shown in .

[0343] A thread group cluster is shown, which includes an NDRange 2820 formed by four 8x8 thread groups. Each thread group is dispatched to processing resources on a different graphics core 2702A-2702D. The thread groups all access a surface 2803 from the L2 cache 2802, which has been prefetched to each local load / store / cache unit (LSC 2806A-2806D). The LSCs 2806A-2806D include a load / store circuit module to perform memory operations, including load, store, atomic, and prefetch operations between the global memory and the SLMs 2706A-2706D. The LSCs 2806A-2806D also include a cache control circuit module to manage cache control operations associated with these load, store, atomic, and prefetch operations. In one embodiment, the LSCs 2806A-2806D include an L1 cache memory and / or a cache control circuit module for the L1 cache memory. In one embodiment, LSCs 2806A-2806D may be coupled to an interface to L2 cache 2802 and interface with a separate L2 cache control circuit module.

[0344] Without a multicast mechanism, data must be transferred across the L2 interface between each LSC 2806A-2806D and the L2 cache 2802, and possibly across the internal L2 structure 2813, depending on which L2 nodes 2812A-2812D hold the data. This can lead to a bottleneck at the memory banks of the L2 cache 2802. This bottleneck can occur because one or more memory banks of the memory surface 2803 are accessed four times (e.g., by each graphics core 2702A-2702D), requiring four times the L2 and memory arbiter bandwidth. With the multicast feature, one graphics core (e.g., graphics core 2702A) performs a prefetch operation to bring data from the L2 cache 2802 into the graphics core's SLM (2706A). Graphics core 2702A then multicasts the data to the other graphics cores 2702B-2702D across the cross-core interconnect 2804A-2804C. The multicast feature reduces pressure on the L2 nodes 2812A and L2 fabric 2813 of the memory surface 2803. The multicast feature also frees up three of the four L2 interconnects to perform other accesses. Those accesses can be performed on other data, which can then be multicast to the other graphics cores 2702A-2702D within the graphics core cluster 2701. If the algorithm pipeline is deep enough to have four iterations in flight, all four graphics cores 2702A-2702D can have prefetches in flight, thereby maximizing the available L2 cache 2802 and memory arbiter bandwidth.

[0345] Figure 29 A system 2900 is shown including a graphics core cluster 2701 with asynchronous DMA and MMA circuit modules. In one embodiment, Figure 28 The multicast feature of the system 2800 can be implemented via asynchronous DMA units 2708A-2708D within the graphics cores 2702A-2702D to facilitate loading and storing matrix data for MMA operations to be performed via asynchronous MMA units 2712A-2712D. Instead of each graphics core 2702A-2702D reading all necessary tensor data separately (which may result in repeated reads of the L2 cache), different graphics cores 2702A-2702D can obtain different blocks or fragments of the input tensor data, which can then distribute the input tensor data to other graphics cores in the graphics core cluster 2701.

[0346] In one embodiment, asynchronous DMA unit 2708A of graphics core 2702A can retrieve and transfer input data from asynchronous MMA unit 2712A of graphics core 2702A and asynchronous MMA units 2712B-2712C of adjacent graphics cores (e.g., graphics core 2702B and graphics core 2702C). Similarly, asynchronous DMA unit 2708B of graphics core 2702B can retrieve and transfer input data from asynchronous MMA unit 2712B of graphics core 2702B and asynchronous MMA units 2712A and 2712D of adjacent graphics cores (e.g., graphics core 2702A and graphics core 2702D). Similarly, asynchronous DMA unit 2708C of graphics core 2702C can retrieve and transfer input data from asynchronous MMA unit 2712C of graphics core 2702B and asynchronous MMA units 2712A and 2712B of adjacent graphics instances of the graphics core (graphics cores 2702A and 2702B). Asynchronous DMA unit 2708D of graphics core 2702D can retrieve and transfer input data from asynchronous MMA unit 2712D of graphics core 2702B and asynchronous MMA units 2712B and 2712C of adjacent graphics instances of the graphics core (e.g., graphics cores 2702B and 2702C).

[0347] Figure 30 1 shows a graphics core component capable of asynchronous DMA operations according to an embodiment. The asynchronous DMA feature described herein frees up processing element hardware thread and register file resources by loading data directly from global memory into shared local memory. The graphics core cluster 2701 described herein includes multiple graphics cores. The graphics core cluster 2701 may include four graphics cores 2702A-2702D (e.g., Figure 27-Figure 29 ), or eight graphics cores (as shown in Figure 30 Each graphics core has an asynchronous operation circuit module, which includes the asynchronous DMA unit 2708 and the asynchronous MMA unit 2712, as well as the SLM 2706 (as shown in FIG. Figure 27-28 ), and L1 cache 2709 (as shown in Figure 27 ).

[0348] According to one embodiment, graphics core 2702 includes processing resources 3002, a load / store / cache front end (LSCFE 3004), and an asynchronous front end (ASYNCFE 3006). Processing resources 3002 (e.g., execution units, GPGPU core 262, graphics core 370, compute units 1560A-1560N, etc.) are configured to execute instructions associated with graphics and compute programs executed by the GPU. In one embodiment, processing resources 3002 are multi-threaded, each supporting concurrent execution of multiple hardware threads (e.g., PR threads 2714A-2714N). Each hardware thread can execute multiple SIMD channels and / or multiple SIMT threads. The asynchronous DMA operations described herein can be used to asynchronously load input data associated with graphics and compute programs from global memory into shared local memory and store output data back to global memory, traversing various cache memory levels (e.g., L1 cache 2709 and L2 cache 2802) based on the cacheability settings associated with the load and store operations. In response to an instruction to transfer data, processing resource 3002 may submit a request to LSCFE 3004, which in turn sends a request to the LSC (e.g., Figure 28 The LSCFE 3004 provides a front-end interface to the LSCs 2806A-2806D (for graphics core 2702). In response to a request to perform an asynchronous memory transfer, the LSCFE 3004 can submit the request to the asynchronous DMA unit 2708 via the ASYNCFE 3006. The ASYNCFE 3006 also provides an interface for performing asynchronous MMA operations via the asynchronous MMA unit 2712. The LSC crossbar switch (LSC_XBAR 3016) enables access to the SLM 2706 and L1 cache 2709 to satisfy LSC operations with a source or destination including the LSC 2806 or L1 cache 2709, and facilitates access to the LSC 2806 and L1 cache 2709 by the asynchronous MMA unit 2712 and the asynchronous DMA unit 2708. Requests for access to global memory from multiple requesters within the graphics core 2702 are arbitrated by the memory arbiter 3014.

[0349] In one embodiment, the asynchronous DMA unit 2708 is physically adjacent to the SLM 2706 of the graphics core 2702 and has direct access to the memory banks of the SLM 2706. The asynchronous DMA unit 2708 processes asynchronous requests from threads of the processing resource 3002 and communicates with the addressable barrier (A-barrier) logic within the gateway 3008 for synchronization. The gateway 3008 transmits synchronization messages between graphics cores via the cluster control structure port 3010, which provides access to control structures within the graphics core cluster 2701. The asynchronous DMA unit 2708 connects to the graphics core interconnect via the cluster data structure port 3012, which provides access to the graphics core interconnect network (e.g., the cross-core interconnects 2804A-2804C) for peer-to-peer data sharing and data multicast operations.

[0350] The embodiments described herein include hardware support for the GEMM programming model for matrix multiplication operations. The GEMM programming model is implemented via thread group clusters and uses asynchronous DMA unit 2708 and asynchronous MMA features enabled via asynchronous MMA unit 2712. Synchronization of asynchronous DMA operations can be performed via A-barrier operations, which are managed via gateway 3008.

[0351] Table 5 shows example traffic data flow paths within a cluster when performing asynchronous DMA operations in service of GEMM operations via asynchronous MMA unit 2712 or otherwise in service of load / store / atomic operations performed by processing resources 3002 of graphics core 2702 within graphics core cluster 2701 . Table 5 – Service data flow paths

[0352] Figure 31 2708. The circuit module 3100 associated with the asynchronous DMA unit 2708 includes circuit components external to the asynchronous DMA unit 2708 (these circuit components are also in Figure 30), and internal circuit modules of the asynchronous DMA unit 2708. In one embodiment, the internal circuit modules of the asynchronous DMA unit 2708 include a message tracker 3102, a request tracker 3104, a contention buffer 3106, a transposition unit 3108, an address generator unit (AGU) / hash unit 3110, and a control circuit module 3111 for managing the operation of the internal circuit modules. The asynchronous DMA unit 2708 may also include a cache 3112, which, in one embodiment, is used to cache tensor descriptors of tensor data to be transferred. The circuit module 3100 is configured to enable asynchronous transfer of data to and from LSC-managed memory banks (LSC memory banks), which in various embodiments include memory banks of the SLM 2706, memory banks of the L1 cache 2709, and remote SLM memory banks associated with the distributed SLM 2716 of the graphics core cluster 2701.

[0353] During operation, messages initiating asynchronous DMA operations are received at ASYNCFE 3006 and routed to the message interface of asynchronous DMA unit 2708. Message tracker 3102 includes circuitry for tracking incoming messages received at the message interface until those messages are processed. Control circuitry module 3111 can process the messages and submit requests to the graphics processor and cluster circuitry to perform operations associated with those requests. In one embodiment, outstanding requests are tracked via request tracker 3104. Requests to access global memory are arbitrated via memory arbiter 3014. Asynchronous DMA unit 2708 can submit requests and / or commands to memory arbiter 3014 on the data path, which may include write data for write operations. For read operations, the data path from memory arbiter 3014 carries the data back.

[0354] In one embodiment, a fork circuit module 3114 on the data return path from the memory arbiter 3014 can send data back to the LSC memory bank or to the asynchronous DMA unit 2708 for buffering (e.g., via the contention buffer 3106 in the event of memory bank contention). A circuit module similar to the fork circuit module 3114 also resides on the data path from the cluster data fabric ports 3012A-3012B to route commands or data to the LSC memory bank or the asynchronous DMA unit 2708.

[0355] The asynchronous DMA unit 2708 can also process the returned data via the transpose unit 3108 to perform inline transpose operations on tensor data. The AGU / hash unit 3110 is used to perform address generation for memory accesses to tensor data in slice memory format, supporting up to five-dimensional tensors. The hash portion of the AGU / hash unit 3110 performs a bank hash calculation on tensor data stored in the LSC memory bank. The asynchronous DMA unit 2708 can access the LSC_XBAR 3016 to write or read data directly to the SLM 2706.

[0356] The asynchronous DMA unit 2708 synchronizes asynchronous DMA operations via the A-barrier object using the barrier interface at the gateway 3008. The gateway 3008 is coupled to the cluster control structure port 3010 so that an inter-cluster barrier can be implemented via the inter-cluster barrier interface. The asynchronous DMA unit 2708 can be configured to issue a complete_tx operation to the A-barrier attached to the asynchronous operation after the operation is completed. The A-barrier can be located in a local barrier at the gateway 3008, or in a remote barrier managed by a remote gateway, as determined by the source workgroup_id sent over the interface and the destination workgroup_id decoded from the barrier address. Requests for A-barriers in remote gateways are always sent to the local gateway first and then forwarded to the remote gateway via the control structure (which is accessible via the cluster control structure port 3010).

[0357] In general, the asynchronous DMA unit 2708 can implement a variety of functionalities, including: bulk and tensor loads from and stores to the L2 cache, tensor atomics to the L2 cache, multicasting L2 return data to distributed shared memory within the cluster, prefetching into the L2 cache, and copying data across distributed shared memory within the cluster. Completion support is provided via A-barriers at the graphics core or graphics core cluster level. After the DMA operation completes, threads that will use the data copied by the DMA operation to perform compute operations will explicitly wait on the A-barrier. Those threads can then start.

[0358] The asynchronous DMA unit 2708 supports concurrent processing of multiple commands, including concurrent loading of multiple input matrices so that data returns are interleaved to increase interconnect BW utilization. For batch and tensor loads, up to 5D tensors are supported using linear or sheet-like (e.g., tensor) copies of data with linear or sheet-like memory layouts. The asynchronous DMA unit 2708 supports SLM bank hashing to increase the efficiency of the matrix unit during asynchronous MMA operations. In one embodiment, zero and / or NaN padding on loads and data discarding on stores are supported for out-of-bounds elements. Byte alignment is supported for global memory accesses. In one embodiment, asynchronous DMA operations cannot be preempted.

[0359] Units such as the gateway 3008 and the asynchronous DMA unit 2708 can send requests to the control structure via the cluster control structure port 3010, or to the data structure via the cluster data structure ports 3012A-3012B. Those requests are routed to the remote graphics core via the control or data structures within the graphics core cluster 2701. The original request contains the source and destination logical workgroup IDs, while the interconnect structure uses the physical graphics core ID for routing. Therefore, mapping tables are defined in these units to obtain the physical IDs and prepare the packets. The mapping table contains the logical workgroup IDs in the cluster and the physical graphics cores to which they are dispatched. An example definition is shown below, where the workgroup_id array can be used to obtain the physical graphics core ID based on the logical workgroup ID. Table 6 – Graphics core mapping table 01 structgraphics_core_mapping_table{ 02 / / Indicates whether the workgroup / graphics core is part of a cluster 03 boolcluster_en; 04 / / Cluster ID; for debugging purposes 05 uint8_tcluster_id; 06 / / core_valid[n] == true means a work group is assigned to core#n 07 boolcore_valid[MAX_CORES]; 08 / / workgroup_id[n] is the logical workgroup id assigned to core #n 09 uint8_tworkgroup_id[MAX_CORES]; 10 }

[0360] The mapping table is set up during dispatch.When the compute front end (CFE) dispatches a cluster's workgroup to a collection of graphics, it sends these (wg_id, core_id) dispatch results to the asynchronous DMA unit 2708 and gateway 3008 in each graphics core via the dispatch structure. Tensor descriptors for asynchronous tensor transfers

[0361] In BERT training, the sequence lengths of each batch are typically different. In the multi-head attention (MHA) layer, a batch GEMM layer, a softmax layer, and another batch GEMM layer are typically fused into one kernel. MHA is performed per head and per batch. The different sequence lengths of each batch result in different tensor dimension sizes for each batch GEMM. Creating tensor descriptors on the host side requires creating many tensor descriptors. An additional way is provided to program tensor descriptors, where tensor descriptors are constructed on-the-fly during device kernel execution, which involves stateless tensor access. The benefits of tensor descriptor construction in the air are as follows:

[0362] 1. There is no catastrophic performance degradation because the number of tensor descriptors is slightly higher than the surface state cache size, which may cause problems for compute and graphics operations.

[0363] 2. Tensor descriptors are constructed on the fly in the device kernel. Therefore, a tensor descriptor can support as many tensors as the application requires.

[0364] In various embodiments, stateful or stateless tensor access is supported, with at least one embodiment supporting both stateful and stateless tensor access.

[0365] Multiple methods can be applied to construct or update tensor descriptors on the fly. In one method, tensor descriptors are stored in a register file across multiple SIMT channels. In another method, tensor descriptors are stored in a SIMT channel register or in a systolic register file (SRF) (e.g., systolic register file 1840). In a third method, tensor descriptors are stored in shared local memory. In each method, descriptor data is automatically populated by a predefined compute routine (which is executed by a compute resource of a processing resource). In one embodiment, multiple execution channels (e.g., SIMT channels / SIMD channels) are used to populate descriptor fields, where each channel populates different information for the tensor descriptor. In one embodiment, a single execution channel is used to populate all tensor descriptor fields. When tensor descriptors are stored in shared local memory, tensor descriptor creation and updates can be performed using one channel in a thread, multiple channels in a thread, or multiple threads. Tensor descriptors can be reused and updated as needed. Tensor descriptor updates can be performed by a single channel. After constructing or updating the tensor descriptor, the tensor descriptor is written to the shared local memory. A shared local memory fence operation is used to ensure that the tensor descriptor is completely written before issuing an asynchronous DMA command to copy the tensor data. If multiple threads in a thread group use the same tensor descriptor, a thread group synchronization operation can be performed. In one embodiment, the tensor descriptor is stored by the DMA unit in a DMA cache (e.g., cache 3112) for later reuse. After the tensor descriptor is updated, a DMA cache invalidation is performed on the updated tensor descriptor. In one embodiment, this invalidation is explicitly programmed by the kernel programmer. The SLM pointer of the tensor descriptor can be used to transfer the completed or updated tensor descriptor through a single channel.

[0366] API commands are provided to support cluster-based GEMM programming. The specific API commands implemented may vary across embodiments. Example API commands are as follows: 1. Cluster::kernel launch(): Launches a kernel for execution on the cluster. 2.Cluster::work_group_id(): Query the thread group id in the cluster. 3.Cluster::sync(): Synchronizes all threads in the cluster. 4. Async_copy(global_descriptorg_desc, slm_descriptor slm_desc, barrierbar, uINT32_tmulti_cast_mask): This API is updated to include multi_cast_mask, indicating which thread groups the DMA will copy data to and perform asynchronous reach operations on the barriers (same shared local memory address) in that thread group. 5. async_mma(slm_descriptorD_desc, slm_descriptor C_desc, slm_descriptorA_desc, slm_descriptorB_desc, barrierbar_a, uINT32_t, wg_mask_a, barrierbar_b, uINT32_twg_mask_b, barrierbar_d): This API is updated to include the wg_masks for matrices A and B, indicating the thread groups that the MMA must notify via the barrier complete-tx message so that the buffer refill is ready.

[0367] Figure 32 A method 3200 for performing asynchronous DMA operations on a graphics processor is shown. In various embodiments, the method 3200 is performed via circuit modules within and associated with the asynchronous DMA unit described herein, including but not limited to: Figure 30 and Figure 31 The circuit module shown in .

[0368] Method 3200 includes receiving a request to perform an asynchronous direct memory access (DMA) operation at an asynchronous front-end circuit module of a graphics core of a graphics core cluster (3202). In one embodiment, an asynchronous front-end circuit module (e.g., ASYNCFE 3006) provides a front end to the asynchronous operation circuit module of the graphics core, including asynchronous DMA unit 2708 and asynchronous MMA unit 2712. The asynchronous front-end circuit module may then route the request to a circuit module of an asynchronous DMA controller (e.g., asynchronous DMA unit 2708) within the graphics core (3204). The asynchronous DMA controller may then determine parameters of the data to be transferred, including determining a memory layout of the data to be transferred via the DMA operation (3206).

[0369] In one embodiment, the memory layout can be a linear layout (which uses a linear address mode) or a tensor layout (which is a slab memory layout for use with multi-dimensional tensor data). The memory layout can be determined via a tensor descriptor, which defines the parameters of the tensor data to be copied asynchronously. The tensor descriptor can be stored in a cache (e.g., cache 3112) within the asynchronous DMA controller. Based on the memory layout (3207) of the data to be copied, the address generator of the asynchronous DMA controller can be configured for a linear or slab layout. For a linear layout, the address generator will generate addresses for data transfers based on the linear memory layout (3208). For a slab layout, the address generator will generate addresses for data transfers based on the slab memory layout (3210). The asynchronous DMA controller will then use the generated addresses to submit a memory access request to the target memory device (3212). For loads from global memory to shared local memory, the target memory device is a bank of memory devices mapped to the graphics processor's global memory, which may include device local memory and system memory. The global memory may be backed by an L2 cache (eg, L2 cache 1904, L2 cache 2802) or an L3 cache (eg, L3 cache 1906).

[0370] Figure 33 A method 3300 for performing coordinated asynchronous DMA operations on a surface in memory is shown. Instead of each graphics core loading the entire memory surface, the coordinated asynchronous DMA operations distribute global memory and L2 cache accesses for the surface across graphics cores within the cluster. This distribution reduces the overall pressure on the memory arbiter, memory interface, L2 cache, L2 cache interface, and (potentially) the internal L2 structure of the distributed L2 cache memory.

[0371] In one embodiment, an asynchronous DMA circuit module may receive a request to perform an asynchronous direct memory access (DMA) operation to transfer a surface from global memory to a distributed shared local memory (SLM) within a graphics core cluster (3302). An operation is performed to partition the surface into a plurality of regions (3304). In various embodiments, this operation may be performed by the asynchronous DMA circuit module, the graphics core circuit module, or the graphics core cluster circuit module. In one embodiment, the partitioning is predetermined by a programmer or by program code executed by a graphics processor or a general-purpose processor. The regions of the surface are loaded into a distributed SLM, which includes an SLM for each graphics core within the graphics core cluster.

[0372] Method 3000 further includes performing an operation to load a first region of the surface into an SLM of a first graphics core of the graphics core cluster via a first DMA engine within the first graphics core (3306). Performing an additional operation to load a second region of the surface into an SLM of a second graphics core of the graphics core cluster via a second DMA engine within the second graphics core. The loading may be performed from a global memory (e.g., global memory 2710) via an L2 cache (e.g., L2 cache 2802).

[0373] The DMA engines can then relay data to be used by other graphics cores to the SLMs of those graphics cores via asynchronous DMA operations to the SLMs of the remote graphics cores. For example, method 3000 includes an operation (3308) of transferring a first region of a surface to an SLM within a second graphics core via a first DMA engine. Method 3000 also includes an operation (3310) of transferring a second region of the surface to the SLM within the first graphics core via a second DMA engine. Each graphics core can perform load and transfer operations for a separate region of the surface, just as Figure 28 System 2800 and Figure 29 The method 3000 also includes using a gateway circuit module (e.g. Figure 30 and Figure 31 Gateway 3008) synchronizes the asynchronous DMA operation via an addressable barrier object (3312).

[0374] Figure 34A method 3400 for synchronizing asynchronous DMA operations via an addressable barrier object is shown. Method 3400 is performed via a processing resource of a graphics processor as described herein, wherein the processing resource is an asynchronous DMA engine. In one embodiment, method 3400 includes executing a producer thread and a consumer thread on the processing resource of a graphics processor core, wherein the producer thread and the consumer thread are synchronized via the addressable barrier object (3402). Method 3400 further includes programming the asynchronous DMA engine via the producer thread to copy data from global memory to shared local memory of the graphics processor core (3404). In one embodiment, programming the DMA engine includes automatically generating tensor descriptors so that tensor data is asynchronously copied to the shared local memory. Method 3400 further includes updating the producer addressable barrier object upon completion of the asynchronous copy (3406). In various embodiments, the producer addressable barrier object can be updated by the producer thread or the asynchronous DMA engine. Method 3400 further includes executing a compute operation at the consumer thread in response to the update of the producer addressable barrier object (3408). The computational operation may be, for example, an asynchronous matrix multiplication operation, which is orchestrated via the asynchronous MMA unit 2712 and executed via the matrix accelerator and / or tensor processing resources within the graphics core cluster. Data for other computational and / or graphics operations may also be asynchronously copied before those operations are executed. Output data may also be stored to global memory via the asynchronous copy operation. Method 3400 may then be repeated for other sets of inputs and outputs. Additional Exemplary Computing Devices

[0375] Figure 35 35 is a block diagram of a computing device 3500 including a graphics processor 3504, according to one embodiment. Versions of the computing device 3500 may be or be included in a communication device such as a set-top box (e.g., an internet-based cable TV set-top box), a global positioning system (GPS)-based device, or the like. The computing device 3500 may also be or be included in a mobile computing device such as a cellular phone, a smartphone, a personal digital assistant (PDA), ...

Claims

1. A graphics processor, comprising: a base die comprising a plurality of chiplet sockets; and a plurality of chiplets coupled to the plurality of chiplet sockets, a chiplet in the plurality of chiplets comprising: A graphics core cluster, wherein the graphics core cluster includes multiple graphics cores; a distributed shared local memory, the distributed shared local memory comprising a shared local memory within each of the plurality of graphics cores; as well as A direct memory access engine is provided within each of the plurality of graphics cores, the direct memory access engine being configured to asynchronously copy data from a memory device to the distributed shared local memory.

2. The graphics processor of claim 1 , wherein the chiplet of the plurality of chiplets comprises a matrix accelerator within each of the plurality of graphics cores, the matrix accelerator having a circuit module for asynchronously performing matrix multiplication and accumulation operations.

3. The graphics processor according to claim 2, wherein: Each of the plurality of graphics cores includes an asynchronous front-end circuit module coupled with the direct memory access engine and the matrix accelerator.

4. The graphics processor according to claim 3, wherein: The asynchronous front-end circuit module is configured to receive a request for performing an asynchronous operation, determine a target of the request from a set of targets including the direct memory access engine and the matrix accelerator, and route the request to the target.

5. The graphics processor of claim 4 , wherein the direct memory access engine is configured to: determining a memory layout of the data in the memory device; generating an address for a data transfer associated with the data based on the memory layout; and A plurality of memory access requests are submitted to read data at the generated addresses associated with the memory device.

6. The graphics processor according to claim 5, wherein: The data includes tensor data having a slice-like memory layout, and the direct memory access engine is configured to generate slice-like addresses for the tensor data.

7. The graphics processor according to any one of claims 1 to 7, wherein: Determining the memory layout of the data includes reading the memory layout from a tensor descriptor of the data, the tensor descriptor stored in a cache memory of the direct memory access engine.

8. The graphics processor according to claim 1, wherein: The direct memory access engine of a first graphics core among the plurality of graphics cores is configured to load a first region of tensor data from the memory device into a first shared local memory of the first graphics core, and to transfer the first region of tensor data to a second shared local memory of a second graphics core, the first shared local memory and the second shared local memory being located within the distributed shared local memory.

9. The graphics processor according to claim 8, wherein: The direct memory access engine of the second graphics core of the plurality of graphics cores is configured to load a second region of tensor data from the memory device into the second shared local memory of the second graphics core, and to transfer the second region of tensor data to the first shared local memory of the first graphics core.

10. The graphics processor according to claim 9, wherein: Each of the plurality of graphics cores includes a gateway circuit module configured to track addressable barrier objects associated with the graphics core cluster, the addressable barrier objects configured to facilitate synchronization between the first graphics core and the second graphics core.

11. A method comprising: receiving a request to perform an asynchronous direct memory access (DMA) operation at an asynchronous front-end circuit module of a graphics core within a graphics core cluster; A circuit module for routing the request to an asynchronous DMA controller within the graphics core; determining a memory layout of data to be transferred via the asynchronous DMA operation; generating, via an address generator of the asynchronous DMA controller, an address for data transmission based on the memory layout; as well as A memory access request is submitted by the asynchronous DMA controller to a target memory device using the address generated by the address generator.

12. The method of claim 11, wherein: The data to be transferred via the asynchronous DMA operation includes multi-dimensional tensor data, and determining the memory layout of the data to be transferred via the asynchronous DMA operation includes reading a memory layout from a tensor descriptor associated with the multi-dimensional tensor data.

13. The method of claim 12, further comprising: submitting the memory access request to a memory device associated with the graphics core cluster, the memory device being mapped as a global memory device; and The memory access request is satisfied via cache memory associated with the graphics core cluster.

14. The method of claim 12, wherein: The multidimensional tensor data comprises a plurality of regions of tensor data, and the method further comprises: reading a first region of tensor data via the asynchronous DMA controller; storing the first region of tensor data in shared local memory within the graphics core; as well as The first region of tensor data is transferred to the shared local memory of another graphics core within the graphics core cluster.

15. The method of claim 11, further comprising synchronizing threads having dependencies on the data to be transferred by the asynchronous DMA operation via an addressable barrier object.

16. A data processing system comprising: a base die comprising a plurality of chiplet sockets; and a plurality of chiplets coupled to the plurality of chiplet sockets, the plurality of chiplets comprising: a first chiplet comprising a memory device; and A second chiplet, comprising: a graphics core cluster comprising a plurality of graphics cores, the graphics core cluster coupled to the memory device of the first chiplet; a distributed shared local memory, the distributed shared local memory comprising a shared local memory within each of the plurality of graphics cores; as well as A direct memory access engine is provided within each of the plurality of graphics cores, the direct memory access engine being configured to asynchronously copy data from a memory device to the distributed shared local memory.

17. The data processing system of claim 16, wherein the second chiplet comprises a matrix accelerator within each of the plurality of graphics cores, the matrix accelerator having a circuit module for asynchronously performing matrix multiplication and accumulation operations.

18. The data processing system according to claim 17, wherein: Each of the plurality of graphics cores includes an asynchronous front-end circuit module coupled to the direct memory access engine and the matrix accelerator, the asynchronous front-end circuit module being configured to receive a request for performing an asynchronous operation, determine a target of the request from a set of targets, and route the request to the target, the set of targets including the direct memory access engine and the matrix accelerator.

19. The data processing system of claim 18, wherein the direct memory access engine is configured to: determining a memory layout of the data in the memory device via tensor descriptors stored in a cache memory of the direct memory access engine; generating an address for a data transfer associated with the data based on the memory layout; as well as A plurality of memory access requests are submitted to read data at the generated addresses associated with the memory device.

20. The data processing system according to any one of claims 16 to 19, wherein: The direct memory access engine of a first graphics core among the multiple graphics cores is configured to load a first region of tensor data from the memory device into a first shared local memory of the first graphics core, and transfer the first region of tensor data to a second shared local memory of a second graphics core, the first shared local memory and the second shared local memory being located within the distributed shared local memory, and wherein the direct memory access engine of a second graphics core among the multiple graphics cores is configured to load a second region of tensor data from the memory device into the second shared local memory of the second graphics core, and transfer the second region of tensor data to the first shared local memory of the first graphics core.

21. The data processing system of claim 20, wherein: Each of the plurality of graphics cores includes a gateway circuit module configured to track addressable barrier objects associated with the graphics core cluster, the addressable barrier objects configured to facilitate synchronization between the first graphics core and the second graphics core.

Images