A method and apparatus for improving the scalability of GPU architecture

By distributing workloads to the Hyper-Extended Processing Group (HPC) and the SPC through a hierarchical GPU architecture, the bottleneck problem caused by the connection pressure between common modules and the SPC in the GPU architecture is solved, thereby improving the scalability and performance of the GPU system.

CN123118339APending Publication Date: 2026-11-17BEIJING FENGHUA CHUANGZHI TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202510625941.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-05-15
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

As the number of shared common modules (SPCs) in existing GPU architectures increases, the connection pressure between the common modules and SPCs increases significantly, leading to bottlenecks in scalability and performance.

Method used

The total workload is distributed to the hyperscale processing group, and then further distributed to SPCs within the hyperscale processing group. A layered architecture is adopted to reduce the amount of control logic and connections between common modules and SPCs. The load distribution is optimized through a hierarchical distribution mechanism at the system level and SSG common modules.

Benefits of technology

It simplifies GPU hardware implementation, reduces connectivity issues, breaks through scalability and performance bottlenecks, and greatly improves the scalability of GPU systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118339A_ABST
    Figure CN123118339A_ABST
Patent Text Reader

Abstract

This invention relates to the field of high-performance computing technology, and provides a method and apparatus for improving the scalability of GPU architecture. A system-level common module divides the total workload into multiple first-level workloads, allocating these first-level workloads to corresponding super-expansion processing groups (SPCs). The SSG common module of each SPC further divides these first-level workloads into multiple second-level workloads, allocating these second-level workloads to corresponding SPCs, so that each SPC can process the received second-level workloads. This invention simplifies GPU hardware implementation by dividing the workload into multiple levels of common modules, reduces the pressure on the common modules and the connections between the common modules and each SPC, thereby overcoming the scalability and performance bottlenecks of GPUs and significantly improving the scalability of GPU systems.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of high-performance computing technology, and in particular to a method and apparatus for improving the scalability of GPU architecture. Background Technology

[0002] Graphics Processing Unit (GPU) enables high-performance parallel computing and is therefore widely used to accelerate computer graphics applications.

[0003] In existing technologies, to improve computing performance and scalability, a GPU architecture often includes multiple Shader Processing Clusters (SPCs), and each SPC may contain multiple Shader Execution Units (SEUs). Multiple SPCs share a common module, and through the connection between the common module and each SPC, they receive input data, send output images, and access system-level caches.

[0004] However, when the number of SPCs sharing a common module increases to a certain threshold, the pressure on the common module and the connections between the common module and each SPC will also increase significantly, resulting in bottlenecks in the scalability and performance of the GPU.

[0005] Therefore, overcoming the shortcomings of the existing technology is an urgent problem to be solved in this technical field. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method and apparatus for improving the scalability of GPU architecture. The purpose is to first distribute the total workload to each hyperscale processing group, and then distribute the corresponding first-level workload to each SPC within the hyperscale processing group. This solves the problem of GPU performance bottleneck caused by the large number of SPCs sharing common modules.

[0007] The present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for improving the scalability of GPU architecture, comprising:

[0009] The system-level common module breaks down the total workload into multiple first-level workloads and distributes the first-level workloads to the corresponding hyper-expansion processing groups;

[0010] The SSG common module of the super-expanded processing group splits the first-level workload into multiple second-level workloads and distributes the second-level workloads to the corresponding SPCs so that each SPC can process the received second-level workload.

[0011] Furthermore, the system-level common module divides the total workload into multiple first-level workloads, and allocates the first-level workloads to corresponding hyper-expansion processing groups, including:

[0012] When the system-level entry module of the system-level common module receives the start rendering signal, the system-level entry module splits the total workload into multiple first-level workloads of the same size;

[0013] Based on the tile group credit values ​​from each hyperscale processing group, the multiple first-level workloads of the same size are sequentially distributed to each hyperscale processing group.

[0014] Furthermore, during initial allocation, when the system-level entry module of the system-level common module receives the start rendering signal, the system-level entry module splits the total workload into multiple first-level workloads of equal size, including:

[0015] The system-level entry module obtains the number of SPCs in each hyperscaled processing group and the number of SEUs in each SPC; it calculates the product of the number of SPCs and the corresponding number of SEUs to obtain the total number of pipelines in the hyperscaled processing group.

[0016] Obtain the number of tiles corresponding to the total workload; calculate the ratio of the total number of pipelines to the number of tiles to obtain the initial allocation lower limit;

[0017] The total workload is divided into multiple first-level workloads of the same size that are greater than or equal to the initial allocation lower limit.

[0018] Further, the step of sequentially allocating the plurality of first-level workloads of the same size to each hyper-expansion processing group based on the tile group credit value from each hyper-expansion processing group includes:

[0019] Determine the maximum value of the received tile group credit value, and allocate the first-level workload to the super-expansion processing group corresponding to the maximum value of the tile group credit.

[0020] Transmit the tile group position offset corresponding to the first-level workload to the assigned super-expansion processing group.

[0021] Furthermore, the tile group credit value is the sum of the credit values ​​of all tiles in the super-extended processing group;

[0022] Alternatively, the tile group credit value may be the maximum value among all tile credit values ​​in the super-extended processing group.

[0023] Furthermore, when the first-level workload allocated by the system-level common module to the hyper-expansion processing group corresponds to at least one unassigned tile, the SSG common module of the hyper-expansion processing group splits the first-level workload into multiple second-level workloads, and allocates the second-level workloads to the corresponding SPCs, including:

[0024] When the SSG entry module of the SSG common module receives the tile group position offset, it determines at least one tile to be allocated corresponding to the first level workload based on the tile group position offset.

[0025] Obtain tile credit values ​​from each SPC, determine the maximum tile credit value among the obtained tile credit values, and assign the at least one tile to be assigned to the SPC corresponding to the maximum tile credit value, so as to allocate the second-level workload to the corresponding SPC.

[0026] Furthermore, when the second-level workload corresponds to at least one skipped block, the second-level workload received by each SPC process includes:

[0027] In each super-expansion processing group, at least one skipped tile corresponding to the first-level workload is identified, and a first merge signal for the skipped tile is generated; when each SPC completes processing the received tile to be assigned, a second merge signal for the tile to be assigned is generated; when all the first merge signals and the second merge signals are received, a tile group end signal is generated.

[0028] When all tile group end signals of the hyper-expansion processing groups are received, an overall end signal is generated to complete the response to the total workload.

[0029] In a second aspect, the present invention also provides an apparatus for improving the scalability of a GPU architecture, for implementing the method for improving the scalability of a GPU architecture as described in the first aspect, wherein the apparatus for improving the scalability of a GPU architecture comprises a system-level common module and multiple hyper-expansion processing groups; the hyper-expansion processing group includes an SSG common module and multiple SPCs;

[0030] The system-level common module is used to split the total workload into multiple first-level workloads and allocate the first-level workloads to the corresponding SSG common modules of the hyper-expansion processing group.

[0031] The SSG common module is used to split the first-level workload into multiple second-level workloads and allocate the second-level workloads to the corresponding SPCs so that each SPC can process the received second-level workloads.

[0032] Furthermore, the SSG common module includes an SSG entry module and an SSG cache;

[0033] The SSG entry module is used to determine at least one map block to be allocated corresponding to the first level workload based on the map block group position offset when it receives the map block group position offset.

[0034] The SSG entry module is also used to obtain the tile credit value from each SPC, determine the maximum tile credit value among the obtained tile credit values, and assign the at least one tile to be assigned to the SPC corresponding to the maximum tile credit value, so as to allocate the second-level workload to the corresponding SPC.

[0035] The SSG cache is used to cache the data generated by the second-level workload contained in the SSG during each SPC operation.

[0036] Furthermore, the system-level entry module includes a system-level allocation module and a system-level event merging module; the SSG entry module includes an SSG request module, an SSG allocation module, and an SSG event merging module.

[0037] The system-level allocation module is used to split the total workload into multiple first-level workloads of the same size; and to allocate the multiple first-level workloads of the same size to each SSG request module in sequence according to the tile group credit value from each hyper-expansion processing group.

[0038] The SSG request module is used to determine at least one skip block corresponding to the first level workload in each hyper-expansion processing group, generate a first merge signal for the skip block, and send the first merge signal to the SSG event merging module.

[0039] The SSG allocation module is used to generate a second merging signal for the received blocks to be allocated when each SPC completes the processing of the blocks to be allocated.

[0040] The SSG event merging module is used to generate a tile group end signal when all the first merging signals and second merging signals are received, and send the tile group end signal to the system-level event merging module.

[0041] The system-level event merging module is used to generate an overall termination signal when it receives the tile group termination signals of all hyper-expansion processing groups, so as to complete the response to the total workload.

[0042] Thirdly, the present invention also provides an apparatus for improving the scalability of GPU architecture, comprising:

[0043] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor for performing the method for improving the scalability of the GPU architecture as described in the first aspect.

[0044] Fourthly, the present invention also provides a non-volatile computer storage medium storing computer-executable instructions that are executed by one or more processors to perform the method for improving the scalability of GPU architecture described in the first aspect.

[0045] Fifthly, a chip is provided, comprising: a processor and an interface for calling and running a computer program stored in memory, performing methods for improving the scalability of GPU architecture as described in the first aspect.

[0046] In a sixth aspect, a computer program product containing instructions is provided that, when executed on a computer or processor, causes the computer or processor to perform methods for improving the scalability of a GPU architecture as described in the first to fourth aspects and any one thereof.

[0047] In a seventh aspect, a system for improving GPU architecture scalability is provided, comprising means for improving GPU architecture scalability as described in the second aspect, and using the method for improving GPU architecture scalability as described in the first aspect to perform interaction with the means for improving GPU architecture scalability as described in the second aspect.

[0048] Unlike existing technologies, the present invention has at least the following beneficial effects:

[0049] This invention provides a hierarchical GPU architecture that breaks down the task of allocating computational workload to each SPC into multi-level subtasks for execution. First, the total workload is allocated to each hyperscale processing group, and then the corresponding first-level workload is allocated to each SPC within the hyperscale processing group. This minimizes the control logic and number of connections between the common module and the SPCs and hyperscale processing groups, thereby simplifying the GPU hardware implementation, reducing the connection pressure between the common module and each SPC, and thus breaking through the scalability and performance bottlenecks of the GPU, greatly improving the scalability of the GPU system. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention will be briefly described below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0051] Figure 1 This is a schematic diagram of a prior art GPU architecture provided by an embodiment of the present invention;

[0052] Figure 2 This is a flowchart illustrating a method for improving the scalability of GPU architecture provided by an embodiment of the present invention;

[0053] Figure 3 This is a schematic diagram of a layered GPU architecture according to an embodiment of the present invention;

[0054] Figure 4 This is a flowchart illustrating step 10 provided in an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of another layered GPU architecture provided in this embodiment of the invention;

[0056] Figure 6 This is a schematic diagram of the structure and connection of a system-level entry module provided in an embodiment of the present invention;

[0057] Figure 7 This is a flowchart illustrating step 101 provided in an embodiment of the present invention;

[0058] Figure 8 This is a flowchart illustrating step 102 provided in an embodiment of the present invention;

[0059] Figure 9 This invention provides a schematic diagram of the structure and connection of an SSG entry module.

[0060] Figure 10 This is a flowchart illustrating step 20 provided in an embodiment of the present invention;

[0061] Figure 11 This is a flowchart illustrating another method for improving the scalability of GPU architecture provided by an embodiment of the present invention;

[0062] Figure 12 This is a schematic diagram of an architecture for improving the scalability of GPU architecture provided by an embodiment of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0064] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0065] Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as openly inclusive, meaning "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples" are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this disclosure. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics mentioned may be included in any suitable manner in any one or more embodiments or examples; that is, although they may be incorporated into embodiments or examples using the above terms for reasons such as order and position, it does not limit them to be incorporated in combination by a single embodiment or example.

[0066] In the description of this invention, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this disclosure and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this disclosure.

[0067] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, for example, the description may use the prefix "A" or "B" to describe the same type of nouns as two independent entities. In this case, the corresponding features defined with "A" and "B" are used only to distinguish between similar entities and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features.

[0068] In describing some embodiments, the terms "coupled," "coupled," and "connected," and their derivative expressions, may be used. For example, the term "connected" may be used in describing some embodiments to indicate that two or more components have direct physical or electrical contact with each other. Similarly, the term "coupled" may be used in describing some embodiments to indicate that two or more components have direct physical or electrical contact. However, the terms "connected" or "coupled" may also refer to two or more components that do not have direct contact with each other but still cooperate or interact with each other, such as "optical coupling," "wireless connection," etc. The embodiments disclosed herein are not necessarily limited to the scope of this invention.

[0069] In the description of this invention, the expression “A and / or B” (where A and B are used to formally represent specific features) will be used. The corresponding expression includes the following three combinations: only A, only B, and a combination of A and B.

[0070] As used in this invention, “about,” “approximately,” or “approximately” includes the stated value and the average value within an acceptable range of deviation from a particular value, wherein the acceptable range of deviation is determined by a person skilled in the art taking into account the measurement under discussion and the error associated with the measurement of the particular quantity (i.e., the limitations of the measurement system).

[0071] GPUs are widely used to accelerate computer graphics applications, such as rendering tools for computer games and computer-aided design; GPUs are also widely used in general computing applications, such as the currently popular Large Language Model (LLM) artificial intelligence applications.

[0072] Computer graphics processing systems process graphics data (including raw data and vertex data) from computer graphics applications (such as computer games) and generate rendered images through a rendering pipeline.

[0073] In computer graphics applications, 3D models are set up using primitives, which include, but are not limited to, triangles, lines, and points. These primitives are defined by their vertex, ray, and shadow attributes, respectively, based on their position in 3D space. In the 3D model, the geometric data of the primitives and vertices is sent as input data to the computer graphics processing system for geometric processing. This processing includes converting the primitives to screen space and removing primitives that are not visible in screen space. The geometrically processed data is then sent to the fragment processing pipeline for rendering in the computer graphics processing system. After rendering, the 3D model generates an output image, which is displayed on a display unit (e.g., a screen).

[0074] Computer graphics processing systems have many different implementations, including Immediate-Mode Rendering (IMR) and tile-based rendering. In IMR, primitives are geometrically processed and then directly sent to the fragment processing pipeline. Therefore, during fragment processing, processed primitives can be located anywhere on the screen. In tile-based rendering, the screen is divided into multiple rectangular tiles. After geometric processing, primitives are assigned to tiles at specific locations on the screen, and fragment processing is performed separately within each tile.

[0075] The computer graphics workload from the application is sent to the GPU hardware via the graphics control flow. This includes rendering state, geometry data, shader code, and data, so that the geometry data can be processed, the shader code can be executed, and the final output image can be generated in the SEU of the GPU hardware.

[0076] In existing technologies, to improve computing performance and scalability, multiple SEUs can be used to form a graphics pipeline, with each SPC group containing a corresponding graphics pipeline. To achieve different performance targets, a GPU architecture can contain multiple SPC groups when configuring a GPU system.

[0077] like Figure 1 The diagram illustrates a GPU architecture where each SPC group comprises a scheduler, a graphics pipeline consisting of four SEUs, and an SPC cache. The four SPC groups share a common module, which includes an entry module and a system-level cache, each connected to one of the SPC groups. The graphics control flow of the workload is distributed from the entry module to each SPC group as input data. The scheduler in each SPC then distributes and schedules the data to the respective SEUs in the graphics pipeline for processing. Specifically, the entry module retrieves necessary cached data from the system-level cache when distributing workloads, or the SPCs retrieve necessary cached data from the SPC cache when processing received workloads. After the SPC finishes processing, it sends an end signal to the entry module. When the entry module receives the end signals from all the assigned task SPCs, it merges them to obtain a single end signal. The configuration register corresponding to the workload will then update the relevant parameters. In the above process, when the total number of SEUs in the SPC is large, the hardware cost, the clock cycle for tile allocation, and the clock cycle for merging the end signal will all increase, which may affect the processing performance.

[0078] When the number of SPCs connected to a shared common module increases (for example, when 16 SPCs share a common module), the pressure on the common module and the connections between the common module and the SPCs also increases. The difficulty and complexity of hardware implementation and control logic are also high, which leads to bottlenecks in the scalability and performance of the GPU.

[0079] To solve the above problems, such as Figure 2 As shown, this embodiment of the invention provides a method for improving the scalability of GPU architecture, including:

[0080] Step 10: The system-level common module splits the total workload into multiple first-level workloads and allocates the first-level workloads to the corresponding hyper-expansion processing groups.

[0081] like Figure 3 As shown, this invention proposes a device for improving the scalability of GPU architecture. The device includes a system-level common module and N+1 Super Scalable Groups (SSGs). Each Super Scalable Group includes M+1 SPCs and an SSG common module. The M+1 SPCs are connected to the SSG common module. Each SPC includes multiple SEUs, a scheduler, and SPC caches, among other common modules. The common modules within each SPC are as follows... Figure 1 As shown.

[0082] The number of SPCs in a hyper-expansion processing group and the number of hyper-expansion processing groups in a GPU core can be configured according to different performance goals of the GPU architecture. For example, a system-level common module is connected to each SSG common module in N+1 hyper-expansion processing groups. Each hyper-expansion processing group includes four SPCs and one SSG common module, with the four SPCs connected to the SSG common module.

[0083] Step 20: The SSG common module of the super-expanded processing group splits the first-level workload into multiple second-level workloads and distributes the second-level workloads to the corresponding SPCs so that each SPC can process the received second-level workload.

[0084] This invention provides a hierarchical GPU architecture that splits the task of allocating computational workload to each SPC into two levels of subtasks for execution. First, the total workload is allocated to each hyperscale processing group, and then the corresponding first-level workload is allocated to each SPC within the hyperscale processing group. This minimizes the control logic and number of connections between the two-level common modules and the SPCs and hyperscale processing groups, thereby simplifying the GPU hardware implementation, reducing the pressure on the common modules and the connections between the common modules and each SPC, and thus breaking through the scalability and performance bottlenecks of the GPU, greatly improving the scalability of the GPU system.

[0085] To further address the aforementioned issues, the following section will use tile-based rendering as an example to illustrate the process of handling the total workload from the application. Figure 4 As shown, step 10 includes:

[0086] Step 101: When the system-level entry module of the system-level common module receives the start rendering signal, the system-level entry module splits the total workload into multiple first-level workloads of the same size.

[0087] like Figure 5 As shown, the system-level common module of this embodiment includes a system-level entry module and a system-level cache, and the SSG common module includes an SSG entry module and an SSG cache. This embodiment allocates the total workload in two levels: First, the total workload is divided into multiple first-level workloads and scheduled from the system-level entry module to the SSG entry module. Then, the SSG entry module divides the first-level workload allocated to itself into multiple second-level workloads and allocates them to the SPCs in the corresponding SSGs, and further assigns them to the SEUs in the graphics pipeline within the SPCs, so that the SEUs can process the second-level workloads. Specifically, during the processing of the first-level workloads in each SSG, the system-level common module retrieves the required cached data from the system-level cache; during the processing of the second-level workloads in each SPC, the system-level common module retrieves the required cached data from the SSG cache.

[0088] It should be noted that the method of distributing the total workload to two levels in this embodiment of the invention is only a specific example of the method for improving the scalability of GPU architecture in this embodiment of the invention, and is not intended to limit the invention. Those skilled in the art can refer to this embodiment of the invention to divide the workload into multiple levels of common modules for hierarchical processing.

[0089] The allocation process of the SSG allocation module will be described below; the allocation process of the system-level allocation module will then be described below:

[0090] Total workload refers to the rendering tasks required for this process, which include the tile groups to be rendered and the rendering parameters. For example... Figure 6As shown, the configuration register corresponding to the total workload sends a start rendering signal to the system-level entry module. Upon receiving the start rendering signal, the system-level request module in the system-level entry module first obtains the tile groups required to process the total workload. The system-level request module requests and obtains the tile groups in the form of blocks on the screen to be rendered, such as the screen coordinates of the tile groups and the base addresses of the geometric data of the primitives to be processed in the tile groups. After obtaining the tile groups, they are sent to the system-level allocation module, which then allocates them to different hyper-expansion processing groups to split the total workload into multiple first-level workloads and allocate them. The corresponding first workload is the sub-rendering task to be processed in this instance, and each sub-rendering task also includes the tiles to be rendered and the rendering parameters. During allocation, the system-level allocation module can adopt different allocation strategies, such as round-robin mode and demand mode. Round-robin mode refers to allocating tile groups to each SSG in a fixed order. In this mode, the current actual performance of the SSG is not considered, and the first-level workload will be allocated to each SSG in turn until all first-level workloads are allocated. Demand mode refers to dynamically allocating the first-level workload based on the current actual performance of each group of SSGs. In this mode, tasks are assigned to SSGs with higher performance and lower load each time to shorten processing time and improve overall resource utilization and efficiency.

[0091] In an optional embodiment, each hyperscale processing group can be assigned a tile group of equal size; for example, each hyperscale processing group can be assigned a tile group of k×k tiles. The tile group size can be set by the software driver via configuration registers. To better balance the computational workload of SPC and SEU in the GPU system, an optimized tile group size can be obtained through experimentation, thereby improving the overall rendering performance of the GPU system.

[0092] Step 102: Based on the tile group credit values ​​from each hyper-expansion processing group, the multiple first-level workloads of the same size are sequentially distributed to each hyper-expansion processing group.

[0093] The system-level allocation module receives tile group credit values ​​from each hyper-expansion processing group and allocates tile group position offsets to each currently available hyper-expansion processing group. The tile group credit value is calculated from the tile credit values ​​of the SPCs in each hyper-expansion processing group, where the tile group credit value is the sum of all tile credit values ​​in the hyper-expansion processing group; or, the tile group credit value is the maximum value among all tile credit values ​​in the hyper-expansion processing group. The tile credit value can be defined as the SPC's current workload capacity to receive and process new tiles, or it can be defined as the number of tiles being processed in the SPC.

[0094] In the initial allocation, such as Figure 7As shown, step 101 includes:

[0095] Step 1011: The system-level entry module obtains the number of SPCs in each hyper-expansion processing group and the number of SEUs in each SPC; calculates the product of the number of SPCs and the corresponding number of SEUs to obtain the total number of pipelines in the hyper-expansion processing group.

[0096] The initial number of tiles allocated should be equal to or greater than the total number of graphics pipelines in each hyperscale processing group. For example, in a GPU system, if each hyperscale processing group is configured with 4 SPCs and each SPC has 4 SEUs, then the total number of pipelines in that hyperscale processing group is 16.

[0097] Step 1012: Obtain the number of tiles corresponding to the total workload; calculate the ratio of the total number of pipelines to the number of tiles to obtain the initial allocation lower limit.

[0098] For example, when the tile group size is 2×2 tiles, the total workload corresponds to 4 tiles. Since the total number of pipelines is 16, the initial allocation lower limit is 4, meaning that at least 4 tile groups are allocated to this super-expansion processing group in the initial stage. Alternatively, when the tile group size is 4×4 tiles, the total workload corresponds to 16 tiles. Since the total number of pipelines is 16, the initial allocation lower limit is 1, meaning that at least 1 tile group is allocated to this super-expansion processing group in the initial stage.

[0099] Step 1013: Divide the total workload into multiple first-level workloads of the same size that are greater than or equal to the initial allocation lower limit.

[0100] Further tile group allocation methods can be implemented based on the tile group credit value fed back by the SSG allocation module. For example... Figure 8 As shown, step 102 includes:

[0101] Step 1021: Determine the maximum value of the received tile group credit value, and allocate the first-level workload to the super-expansion processing group corresponding to the maximum value of the tile group credit value.

[0102] The system-level allocation module dynamically allocates tile groups to each hyperscale processing group based on the tile group credit value. In one embodiment, the allocation strategy is to allocate the next tile group to the hyperscale processing group with the highest tile group credit value each time, that is, to allocate the tile group to the hyperscale processing group with the highest processing capacity each time. In an optional embodiment, other allocation strategies (such as round-robin mode) can also be used to allocate tile groups to improve GPU performance.

[0103] Step 1022: Transmit the tile group position offset corresponding to the first level workload to the allocated super-expansion processing group.

[0104] The tile group position offset corresponding to the first-level workload refers to the position offset of the tile group to be rendered on the screen for processing the first-level workload. In this embodiment of the invention, the base address is added to the tile group position offset to obtain the position of the tile group in memory. The tile group is then retrieved from memory through the system-level request module and sent to the allocated super-extension processing group through the system-level allocation module.

[0105] In an optional embodiment, in order to reduce the latency from the system-level allocation module to the SSG request module in the SSG entry module, the tile group position offset can be sent sequentially to the allocated super-expansion processing group through a separate first-in-first-out queue / module in the system-level allocation module.

[0106] Compared to existing technologies, the number of hyper-expansion processing groups in this embodiment of the invention is several times that of SPCs in existing technologies. Through hyper-expansion processing groups, the connection pressure between system-level common modules, between system-level common modules and SSG entry modules of hyper-expansion processing groups, and between SSG entry modules and SPCs is kept at a low level, which greatly reduces the difficulty and complexity of hardware implementation and control logic, and reduces the delay in searching for the next graphics pipeline during the tile group allocation process.

[0107] To further explain the allocation process of the SSG entry module, such as Figure 10 As shown, when the first-level workload allocated by the system-level common module to the hyper-expansion processing group corresponds to at least one tile to be allocated, step 20 includes:

[0108] Step 201: When the SSG entry module of the SSG common module receives the tile group position offset, it determines at least one tile to be allocated corresponding to the first level workload based on the tile group position offset.

[0109] The tiles to be assigned corresponding to the first-level workload refer to the tiles that the sub-rendering tasks of this first-level workload need to render. For example... Figure 9 As shown, the SSG request module retrieves the corresponding tiles to be allocated from memory according to the tile group position offset and sends them to the SSG allocation module. This process is similar to the interaction between the system-level request module and the system-level allocation module.

[0110] Step 202: Obtain the tile credit value from each SPC, determine the maximum tile credit value among the obtained tile credit values, and assign the at least one tile to be assigned to the SPC corresponding to the maximum tile credit value, so as to allocate the second-level workload to the corresponding SPC.

[0111] The SSG allocation module dynamically allocates tiles to be assigned to each SPC based on the tile credit value. The corresponding allocation strategy can be the same as the allocation strategy of the system-level allocation module for allocating tile groups to each hyper-expansion processing group.

[0112] like Figure 9 As shown, after the tiles to be assigned are allocated to the SPC, the SPC allocates and schedules them to different SEUs in the graphics pipeline for rendering based on the tile data acquisition module in the SEU. During the allocation to the SEU, the tile position offset of the tiles to be assigned is sent to the tile data acquisition module in the corresponding SEU. The tiles to be assigned are then transferred from the tile data acquisition module to the Raster Image Processing (RAST) module for rendering. Figure 9 In this context, "SSG SEUs" refers to all SEU modules within an SSG.

[0113] To illustrate the process after rendering, such as Figure 11 As shown, when the second-level workload corresponds to at least one skipped block, the process following the second-level workload received by each SPC process includes:

[0114] In step 301, in each super-expansion processing group, at least one skipped tile corresponding to the first level workload is determined, and a first merge signal for the skipped tile is generated; when each SPC completes processing the received tiles to be assigned, a second merge signal for the tiles to be assigned is generated; when all the first merge signals and the second merge signals are received, a tile group end signal is generated.

[0115] In this context, the skipped tiles corresponding to the first-level workload refer to tiles that do not need to be rendered when processing the sub-rendering tasks of that first-level workload. During the graphics rendering process, it is often necessary to skip certain tiles according to a certain strategy, that is, to skip rendering these tiles. These tiles are called skipped tiles.

[0116] For a hyper-expansion processing group, when the SSG request module retrieves the last skipped tile required for processing by the first-level workload from memory, it generates a first merge signal and sends it to the SSG event merge module. When the corresponding raster image processing module tile in the SPC finishes rendering for the assigned module, it generates a second merge signal and sends it to the SSG event merge module. After receiving the first and second merge signals, the SSG event merge module generates a tile group end signal and sends it to the system-level event merge module in the system-level entry module.

[0117] In one alternative embodiment, each SSG allocation module can repeatedly retrieve the tiles to be allocated and the skipped tiles corresponding to the first-level workload allocated to the hyperscale processing group. Although the corresponding read bandwidth will increase, it will be easier to process rendering end events based on the position of the tiles already processed in the SSG allocation module.

[0118] The number of graphics pipelines in this embodiment of the invention is several times that in the prior art, thus reducing the complexity of hardware implementation and reducing the delay in searching for the optimal graphics pipeline during tile allocation.

[0119] In step 302, when all tile group end signals of the super-extended processing groups are received, an overall end signal is generated to complete the response to the total workload.

[0120] like Figure 6 As shown, when the system-level event merging module receives the tile group end signal of all allocated hyper-expansion processing groups, it generates an overall end signal and sends it to the configuration register corresponding to the total workload.

[0121] It should be noted that the method for improving GPU architecture scalability using the embodiments of the present invention has a limited effect on improving GPU performance, but it can greatly simplify hardware implementation and reduce the latency of tile allocation in GPU systems with a large number of SPCs and SEUs, thus significantly improving the scalability of the GPU architecture and laying a necessary foundation for breaking through performance bottlenecks when designing subsequent GPU function extensions.

[0122] In one alternative embodiment, such as Figure 3 As shown, this embodiment of the invention also provides an apparatus for improving the scalability of GPU architecture, including a system-level common module and multiple hyper-expansion processing groups; the hyper-expansion processing group includes an SSG common module and multiple SPCs.

[0123] The system-level common module is used to split the total workload into multiple first-level workloads and allocate the first-level workloads to the corresponding SSG common modules of the hyper-expansion processing group.

[0124] The SSG common module is used to split the first-level workload into multiple second-level workloads and allocate the second-level workloads to the corresponding SPCs so that each SPC can process the received second-level workloads.

[0125] In one embodiment, such as Figure 6As shown, the system-level common modules include a system-level entry module and a system-level cache. The system-level cache refers to a high-speed cache memory (cache), and it is a system-level general-purpose module. The SSG common modules include an SSG entry module and an SSG cache; the SSG cache refers to a cache, and it is an SSG-level general-purpose module.

[0126] The SSG entry module is used to determine at least one map block to be allocated corresponding to the first level workload based on the map block group position offset when it receives the map block group position offset.

[0127] The SSG entry module is also used to obtain the tile credit value from each SPC, determine the maximum tile credit value among the obtained tile credit values, and assign the at least one tile to be assigned to the SPC corresponding to the maximum tile credit value, so as to allocate the second-level workload to the corresponding SPC.

[0128] The SSG cache is used to cache the data generated by the second-level workload contained in the SSG during each SPC operation.

[0129] In one embodiment, such as Figure 6 As shown, the system-level entry module includes a system-level allocation module and a system-level event merging module; as Figure 9 As shown, the SSG entry module includes an SSG request module, an SSG allocation module, and an SSG event merging module;

[0130] The system-level allocation module is used to split the total workload into multiple first-level workloads of the same size; and to allocate the multiple first-level workloads of the same size to each SSG request module in sequence according to the tile group credit value from each hyper-expansion processing group.

[0131] The SSG request module is used to determine at least one skip block corresponding to the first level workload in each hyper-expansion processing group, generate a first merge signal for the skip block, and send the first merge signal to the SSG event merging module.

[0132] The SSG allocation module is used to generate a second merging signal for the received blocks to be allocated when each SPC completes the processing of the blocks to be allocated.

[0133] The SSG event merging module is used to generate a tile group end signal when all the first merging signals and second merging signals are received, and send the tile group end signal to the system-level event merging module.

[0134] The system-level event merging module is used to generate an overall termination signal when it receives the tile group termination signals of all hyper-expansion processing groups, so as to complete the response to the total workload.

[0135] In another alternative embodiment, such as Figure 12 The diagram shown is an architectural schematic of a device for improving GPU architecture scalability according to an embodiment of the present invention. The device for improving GPU architecture scalability in this embodiment includes one or more processors 21 and a memory 22. Figure 12 Take a processor 21 as an example.

[0136] Processor 21 and memory 22 can be connected via a bus or other means. Figure 12 Taking the example of a connection between China and Israel via a bus.

[0137] Memory 22, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs and non-volatile computer-executable programs, such as the method for improving GPU architecture scalability in this embodiment. Processor 21 executes the method for improving GPU architecture scalability by running the non-volatile software programs and instructions stored in memory 22.

[0138] Memory 22 may include high-speed random access memory, and may also include non-volatile memory, such as at least one disk storage device, flash memory device, or other non-volatile solid-state storage device. In some embodiments, memory 22 may optionally include memory remotely located relative to processor 21, which can be connected to processor 21 via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0139] The program instructions / modules are stored in the memory 22. When executed by one or more processors 21, they perform the method for improving GPU architecture scalability in the above embodiments, for example, performing each step of the method for improving GPU architecture scalability described in the embodiments of the present invention.

[0140] This invention also provides a non-volatile computer storage medium storing computer-executable instructions that are executed by one or more processors, for example... Figure 12 A processor 21 may enable one or more of the processors to execute the method for improving GPU architecture scalability in the specific embodiments of the present invention, for example, to execute the various steps of the method for improving GPU architecture scalability described in the embodiments of the present invention above; it may also implement Figure 12The various modules and units described above; or the method for improving GPU architecture scalability in specific embodiments of the present invention, for example, executing the various steps of the method for improving GPU architecture scalability described in the embodiments of the present invention above; can also be implemented. Figure 12 The various modules and units mentioned above.

[0141] It is worth noting that the information interaction and execution process between the modules and units in the above-mentioned device and system are based on the same concept as the processing method embodiment of the present invention. For details, please refer to the description in the method embodiment of the present invention, and will not be repeated here.

[0142] Those skilled in the art will understand that all or part of the steps in the various methods of the embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc.

[0143] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for improving the scalability of GPU architecture, characterized in that, include: The system-level common module breaks down the total workload into multiple first-level workloads and distributes the first-level workloads to the corresponding hyper-expansion processing groups; The SSG common module of the super-expanded processing group splits the first-level workload into multiple second-level workloads and distributes the second-level workloads to the corresponding SPCs so that each SPC can process the received second-level workload.

2. The method for improving GPU architecture scalability according to claim 1, characterized in that, The system-level common module splits the total workload into multiple first-level workloads, and allocates the first-level workloads to corresponding hyper-expansion processing groups, including: When the system-level entry module of the system-level common module receives the start rendering signal, the system-level entry module splits the total workload into multiple first-level workloads of the same size; Based on the tile group credit values ​​from each hyperscale processing group, the multiple first-level workloads of the same size are sequentially distributed to each hyperscale processing group.

3. The method for improving GPU architecture scalability according to claim 2, characterized in that, During initial allocation, when the system-level entry module of the system-level common module receives the start rendering signal, the system-level entry module splits the total workload into multiple first-level workloads of equal size, including: The system-level entry module obtains the number of SPCs in each hyperscaled processing group and the number of SEUs in each SPC; it calculates the product of the number of SPCs and the corresponding number of SEUs to obtain the total number of pipelines in the hyperscaled processing group. Obtain the number of tiles corresponding to the total workload; calculate the ratio of the total number of pipelines to the number of tiles to obtain the initial allocation lower limit; The total workload is divided into multiple first-level workloads of the same size that are greater than or equal to the initial allocation lower limit.

4. The method for improving GPU architecture scalability according to claim 2, characterized in that, The step of sequentially allocating the multiple first-level workloads of the same size to the respective hyper-expansion processing groups based on the tile group credit values ​​from each hyper-expansion processing group includes: Determine the maximum value of the received tile group credit value, and allocate the first-level workload to the super-expansion processing group corresponding to the maximum value of the tile group credit. Transmit the tile group position offset corresponding to the first-level workload to the assigned super-expansion processing group.

5. The method for improving GPU architecture scalability according to claim 4, characterized in that, The tile group credit value is the sum of the credit values ​​of all tiles in the super-extended processing group; Alternatively, the tile group credit value may be the maximum value among all tile credit values ​​in the super-extended processing group.

6. The method for improving GPU architecture scalability according to claim 1, characterized in that, When the first-level workload allocated by the system-level common module to the hyper-expansion processing group corresponds to at least one unassigned tile, the SSG common module of the hyper-expansion processing group splits the first-level workload into multiple second-level workloads, and allocates the second-level workloads to the corresponding SPCs, including: When the SSG entry module of the SSG common module receives the tile group position offset, it determines at least one tile to be allocated corresponding to the first level workload based on the tile group position offset. Obtain tile credit values ​​from each SPC, determine the maximum tile credit value among the obtained tile credit values, and assign the at least one tile to be assigned to the SPC corresponding to the maximum tile credit value, so as to allocate the second-level workload to the corresponding SPC.

7. The method for improving GPU architecture scalability according to claim 6, characterized in that, When the second-level workload corresponds to at least one skipped block, the process following the second-level workload received by each SPC includes: In each super-expansion processing group, at least one skipped tile corresponding to the first-level workload is identified, and a first merge signal for the skipped tile is generated; when each SPC completes processing the received tile to be assigned, a second merge signal for the tile to be assigned is generated; when all the first merge signals and the second merge signals are received, a tile group end signal is generated. When all tile group end signals of the hyper-expansion processing groups are received, an overall end signal is generated to complete the response to the total workload.

8. An apparatus for improving the scalability of GPU architecture, characterized in that, It includes a system-level common module and multiple hyper-expansion processing groups; the hyper-expansion processing group includes an SSG common module and multiple SPCs; The system-level common module is used to split the total workload into multiple first-level workloads and allocate the first-level workloads to the corresponding SSG common modules of the hyper-expansion processing group. The SSG common module is used to split the first-level workload into multiple second-level workloads and allocate the second-level workloads to the corresponding SPCs so that each SPC can process the received second-level workloads.

9. The apparatus for improving GPU architecture scalability according to claim 8, characterized in that, The SSG common module includes an SSG entry module and an SSG cache; The SSG entry module is used to determine at least one map block to be allocated corresponding to the first level workload based on the map block group position offset when it receives the map block group position offset. The SSG entry module is also used to obtain the tile credit value from each SPC, determine the maximum tile credit value among the obtained tile credit values, and assign the at least one tile to be assigned to the SPC corresponding to the maximum tile credit value, so as to allocate the second-level workload to the corresponding SPC. The SSG cache is used to cache the data generated by the second-level workload contained in the SSG during each SPC operation.

10. The apparatus for improving GPU architecture scalability according to claim 9, characterized in that, The system-level entry module includes a system-level allocation module and a system-level event merging module; the SSG entry module includes an SSG request module, an SSG allocation module, and an SSG event merging module. The system-level allocation module is used to split the total workload into multiple first-level workloads of the same size; and to allocate the multiple first-level workloads of the same size to each SSG request module in sequence according to the tile group credit value from each hyper-expansion processing group. The SSG request module is used to determine at least one skip block corresponding to the first level workload in each hyper-expansion processing group, generate a first merge signal for the skip block, and send the first merge signal to the SSG event merging module. The SSG allocation module is used to generate a second merging signal for the received blocks to be allocated when each SPC completes the processing of the blocks to be allocated. The SSG event merging module is used to generate a tile group end signal when all the first merging signals and second merging signals are received, and send the tile group end signal to the system-level event merging module. The system-level event merging module is used to generate an overall termination signal when it receives the tile group termination signals of all hyper-expansion processing groups, so as to complete the response to the total workload.