A large-scale GPU performance detection method, system and device

Through FPGA and CXL Switch technology, parallel processing of large-scale GPU performance detection is realized, solving the problems of low GPU performance detection efficiency and high CPU load in the existing technology, and improving detection efficiency and user experience.

CN118656285BActive Publication Date: 2024-11-12YUANQIXIN (SHANDONG) SEMICONDUCTOR TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411154465.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-22
Publication Date
2024-11-12
Estimated Expiration
2044-08-22

AI Technical Summary

Technical Problem

The prior art is difficult to detect the performance of a large number of GPUs simultaneously, resulting in an increase in CPU computing power loss, affecting the execution process of other applications, and the detection efficiency is low and the feedback time is long.

Method used

The field programmable logic gate array FPGA is used to participate in GPU performance detection. The detection tasks are allocated to multiple GPUs in parallel by computing the fast link CXL Switch, and the performance detection results are calculated using the preset GPU performance arbitration algorithm to reduce CPU load and improve detection efficiency. .

Benefits of technology

It alleviates CPU computing power loss and realizes the simultaneous detection of multiple GPUs in a single period, significantly improves detection efficiency and user experience, reduces the computing load of the detection terminal, and improves the efficiency and stability of data transmission.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118656285B_ABST
    Figure CN118656285B_ABST
Patent Text Reader

Abstract

The present application provides a large-scale GPU performance detection method, system and device, which belongs to the field of GPU performance detection technology. In this method, based on the received GPU detection request, the performance detection task information of the corresponding target GPU to be detected is determined. The performance detection task information at least includes calculating the quick link CXL address and the detection task. Based on the preset CXL Switch, the performance detection task information is sent to the corresponding CXL address so that the target GPU to be detected performs the detection task; after receiving the task execution result from the target GPU to be detected, the performance detection result corresponding to the target GPU to be detected is determined according to the task execution result and the preset GPU performance arbitration algorithm, and the performance detection result is sent to the user terminal; wherein the task execution result at least includes the measured parameters of the target GPU to be detected performing the detection task.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of graphics processing unit (GPU) performance detection, and in particular to a large-scale GPU performance detection method, system and device. Background Art

[0002] GPU is a hardware accelerator specially designed for processing graphics and image calculations. In recent years, the application of GPU has far exceeded the traditional field of graphics rendering and expanded to a wider range of computing tasks.

[0003] With the rise of artificial intelligence, the number of GPUs used by scholars in scientific research and industrial production is increasing day by day, including complex pattern recognition, intelligent computing, convolutional neural network applications, etc. Therefore, it is also very important to conduct scientific and reasonable testing of the performance of GPU devices.

[0004] However, currently, due to the characteristics of the central processing unit (CPU) of personal computers (PCs), GPU performance testing usually adopts the "single PC testing single GPU" mode, which does not support simultaneous testing of a large number of GPUs in a single PC and a single period of time; and the existing methods control and complete GPU testing tasks through the CPU, which will inevitably cause CPU computing power loss and load increase, thereby affecting the execution process of other important PC applications. In addition, the method of directly testing GPU performance through the CPU contained in the PC is inefficient and has a long feedback time. Summary of the invention

[0005] To solve the above problems, the embodiments of the present application provide a large-scale GPU performance detection method, system and device.

[0006] On the one hand, an embodiment of the present application provides a large-scale GPU performance detection method, which is applied to a large-scale GPU performance detection system, and the method includes:

[0007] Based on the received GPU detection request, determine the performance detection task information of the corresponding target GPU to be detected; the performance detection task information at least includes a calculation fast link CXL address and a detection task; one CXL address corresponds to one GPU;

[0008] Based on a preset computation fast link switch CXL Switch, the performance detection task information is sent to the corresponding CXL address, so that the target GPU to be detected performs the detection task;

[0009] After receiving the task execution result from the target GPU to be detected, the performance detection result corresponding to the target GPU to be detected is determined according to the task execution result and a preset GPU performance arbitration algorithm, and the performance detection result is sent to the user terminal; wherein the task execution result at least includes the measured parameters of the target GPU to be detected executing the detection task.

[0010] In one implementation of the present application, a device for detecting GPU performance includes a field programmable gate array FPGA, and the FPGA deploys multiple groups of GPU performance detection IP cores to perform parallel detection of different target GPUs to be detected through different GPU performance detection IP cores; the GPU performance detection IP core at least includes an input data cache module, a request decoding module, a detection item confirmation module, a memory interface generator MIG memory reading module, a data header verification module, a GPU performance detection example data decomposition module, a data analysis and mining example module, an image analysis and processing example module, an AI deep learning example module, a scientific computing simulation example module, a data merging cache module, an example sending and receiving correctness confirmation finite state machine FSM module, a test result random access memory RAM cache module, a performance test result decomposition module, a GPU performance test arbitration calculation module, and an output data cache module.

[0011] In one implementation of the present application,

[0012] Based on the received GPU detection request, the performance detection task information of the corresponding target GPU to be detected is determined, specifically including:

[0013] Determine the required operation data corresponding to the detection task, and determine the preset test reference coefficient corresponding to the detection task; wherein the preset test parameter coefficient includes at least a multidimensional reference complexity and a multidimensional test item coefficient matching the test category corresponding to the detection task;

[0014] According to the preset example dimension division rule, the demand operation data is divided into corresponding multi-dimensional example data; wherein the preset example dimension division rule is used to identify the corresponding example dimension in the demand operation data; the multi-dimensional example data at least includes data analysis and mining examples, image analysis and processing examples, AI deep learning examples, and scientific computing simulation examples;

[0015] The multi-dimensional example data is subjected to data merging processing to generate the performance detection task information.

[0016] In one implementation of the present application, before determining the performance test result corresponding to the target GPU to be tested according to the task execution result and a preset GPU performance arbitration algorithm, the method further includes:

[0017] Determining whether the task execution result is correct;

[0018] If correct, decompose the task execution result according to the parameter dimensions corresponding to the measured parameters to obtain the measured parameters used for calculation of the preset GPU performance arbitration algorithm;

[0019] Otherwise, a fault code corresponding to the target GPU to be detected is generated and sent to the user terminal.

[0020] In one implementation of the present application, the measured parameters include at least the amount of test data, execution time, CXL transmission bandwidth, and test item coefficient; the test item coefficient corresponds to at least the following dimensions: task complexity, model complexity, and precision complexity.

[0021] In one implementation of the present application, determining the performance test result corresponding to the target GPU to be tested according to the task execution result and a preset GPU performance arbitration algorithm specifically includes:

[0022] Calculate a first division result of the test data volume in the measured parameter and the CXL transmission bandwidth as a first intermediate value;

[0023] The difference between the execution time and the first intermediate value is used as the second intermediate value;

[0024] Calculate a second division result of the test data volume and the second intermediate value, and use the product value of the second division result and the test item coefficient as the performance test arbitration calculation result value to generate the performance detection result according to the performance test arbitration calculation result value.

[0025] In one implementation of the present application, before taking the product value of the second division operation result and the test item coefficient as the performance test arbitration calculation result value, the method further includes:

[0026] Determining a current test category corresponding to the detection task;

[0027] Determine a task complexity coefficient as a division result of an actual task complexity corresponding to the current test category and a preset reference task complexity;

[0028] Determine a division result of the actual model complexity corresponding to the current test category and the preset reference model complexity as a model complexity coefficient;

[0029] Determine a division result of the actual precision complexity corresponding to the current test category and the preset reference precision complexity as the precision complexity coefficient;

[0030] The test item coefficient is determined according to the product and value of the task complexity coefficient, the model complexity coefficient, the precision complexity coefficient and the corresponding preset task complexity weight, preset model complexity weight, and preset precision complexity weight.

[0031] In one implementation of the present application, before determining the performance detection task information of the corresponding target GPU to be detected based on the received GPU detection request, the method further includes:

[0032] Receiving the GPU detection request via a high-speed serial computer expansion bus standard PCIe interface connected to the user terminal;

[0033] Based on a preset CXL Switch, the performance detection task information is sent to the corresponding CXL address so that the target GPU to be detected performs the detection task, specifically including:

[0034] The performance detection task information is sent to the target GPU to be detected through a PCIe interface corresponding to the CXL Switch; wherein the CXL Switch corresponds to multiple GPUs.

[0035] On the other hand, the embodiment of the present application also provides a large-scale GPU performance detection system, the system comprising:

[0036] A first determination module is used to determine the performance detection task information of the corresponding target GPU to be detected based on the received GPU detection request; the performance detection task information at least includes a calculation fast link CXL address and a detection task; one CXL address corresponds to one GPU;

[0037] A sending module, configured to send the performance detection task information to the corresponding CXL address based on a preset computation fast link switch CXL Switch, so that the target GPU to be detected performs the detection task;

[0038] The second determination module is used to determine the performance detection result corresponding to the target GPU to be detected according to the task execution result and a preset GPU performance arbitration algorithm after receiving the task execution result from the target GPU to be detected, and send the performance detection result to the user terminal; wherein the task execution result at least includes the measured parameters of the target GPU to be detected executing the detection task.

[0039] On the other hand, the embodiment of the present application further provides a large-scale GPU performance detection device, the device comprising:

[0040] At least one processor; and a memory in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the large-scale GPU performance detection method described above.

[0041] Compared with the prior art, the present invention has the following significant effects:

[0042] 1) Through the above technical solution, FPGA is used to participate in GPU performance detection, which alleviates the pressure of CPU computing power loss in the current PC. At the same time, the above design can coordinate multiple GPUs to complete performance detection in a single period, thereby improving detection efficiency and user experience.

[0043] 2) The large-scale GPU performance detection requirements that need to be executed by the CPU in the PC of the GPU performance detection terminal are transferred to the GPU heterogeneous detection execution unit to realize parallel detection calculation, which greatly reduces the computing load of the PC of the GPU performance detection terminal and significantly improves the performance detection efficiency of large-scale GPUs.

[0044] 3) This application also uses CXL technology to achieve communication between GPU performance detection terminals and multiple GPUs to be detected, with higher data transmission and communication efficiency and stronger stability. At the same time, this application provides a detailed application and scheduling strategy of large-scale GPU performance detection architecture, which can better achieve unified scheduling control of large-scale GPU performance detection and significantly improve system performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0046] Figure 1 A schematic diagram of a process flow of a large-scale GPU performance detection method in an embodiment of the present application;

[0047] Figure 2 A structural schematic diagram of a system corresponding to a large-scale GPU performance detection method in an embodiment of the present application;

[0048] Figure 3 This is another structural schematic diagram of a system corresponding to a large-scale GPU performance detection method in an embodiment of the present application;

[0049] Figure 4 This is a schematic diagram of the structure of a GPU performance detection IP core in a large-scale GPU performance detection method in an embodiment of the present application;

[0050] Figure 5Another schematic diagram of a large-scale GPU performance detection method in an embodiment of the present application;

[0051] Figure 6 A schematic diagram of the structure of a large-scale GPU performance detection system in an embodiment of the present application;

[0052] Figure 7 This is a schematic diagram of the structure of a large-scale GPU performance detection device in an embodiment of the present application. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the present application clearer, the technical solution of the present application will be clearly and completely described below in combination with the specific embodiments of the present application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present application.

[0054] The embodiments of the present application provide a large-scale GPU performance detection method, system and device to solve the current problem that it is difficult to simultaneously detect a large number of GPUs in a single period of time, which easily puts a burden on the CPU, affects the normal process of the device, has low detection efficiency, and affects the user experience.

[0055] The following describes in detail various embodiments of the present application in conjunction with the accompanying drawings.

[0056] The embodiment of the present application provides a large-scale GPU performance detection method, which is applied to a large-scale GPU performance detection system, and the large-scale characterization GPU number is large. Figure 1 As shown, the method may include steps S101-S103:

[0057] S101, FPGA determines the performance detection task information of the corresponding target GPU to be detected based on the received GPU detection request.

[0058] The performance detection task information includes at least a Compute Express Link (CXL) address and a detection task. One CXL address corresponds to one GPU.

[0059] The device used to detect GPU performance includes a Field Programmable Gate Array (FPGA). In this embodiment, FPGA is used as the execution subject. The hardware device can be any electronic device carrying FPGA. FPGA and GPU are not in the same hardware entity device. The large-scale GPU performance detection system also includes a GPU performance detection control terminal, a number of GPUs to be detected, and a computing fast link switch CXL Switch connecting these components. Figure 2 As shown, the GPU performance detection control terminal includes a PC and an FPGA, which is responsible for the overall management and control of the large-scale GPU performance detection system, and directly connects to the user interface through the software application programming interface (API) to display the performance detection results of each GPU to the user; the user control interface on the PC side is used by the user to send GPU detection requests, and the system is coordinated from the software level, and is responsible for coordinating the exchange of GPU detection requests and detection result information with the detection execution unit FPGA; the GPU detection execution unit is responsible for packaging and sending detection task instructions and detection execution data to the CXL address of the target GPU to be detected through the CXL Switch, and timely collecting the detection results and sending them to the user PC; in addition, the GPU performance detection control terminal is connected to the CXL Switch through CXL3.1 to realize communication with each target GPU to be detected.

[0060] The PC (including CPU and cache) of the GPU performance detection control terminal and the GPU detection execution unit (including FPGA and DDR5) are connected through a high-speed serial computer expansion bus standard (Peripheral Component Interconnectexpress, PCIe) interface to achieve information exchange. FPGA is the core of the GPU detection execution unit, responsible for the allocation and scheduling of large-scale GPU concurrent detection tasks, and sends detection task instructions and detection execution data to the target GPU through CXL Switch, collects detection data in a timely manner, calculates and detects GPU performance through algorithms, and sends the results to the user's PC; the GPU detection execution unit receives and stores the performance detection tasks of the user-controlled PC through the DDR5 peripheral, and sends the tasks to each GPU to be detected according to the detection request; the GPU detection execution unit uses the PCIe interface as the physical carrier and is connected to each GPU to be detected through the CXL Switch. The system architecture corresponding to the large-scale GPU performance detection method is as follows: Figure 3As shown in the figure. After receiving the instruction, the target GPU to be tested deployed on the CXL Switch performs a series of complex data reception, value acquisition, decoding, memory access, operation execution (data analysis and mining, image signal processing, AI deep learning, etc.), result collection and write-back operations according to the detection requirements, and sends the execution results back to the GPU detection execution unit in a timely manner; after receiving the processing results, the FPGA chip of the GPU detection execution unit calculates and detects the GPU performance through algorithms, and sends the results to the user PC.

[0061] In an embodiment of the present application, before the FPGA determines the performance detection task information of the corresponding target GPU to be detected based on the received GPU detection request, the FPGA can receive the GPU detection request through the PCIe interface connected to the user terminal.

[0062] The user can send a GPU detection request through the user control interface on the PC, and the FPGA can receive the GPU detection request through PCIe to distribute it to each target GPU to be detected, and then wait for the response signal of the GPU.

[0063] Among them, FPGA deploys multiple groups of GPU performance detection IP cores, IP cores are intellectual property cores, so that different GPU performance detection IP cores can be used to detect different target GPUs in parallel. For example, FPGA deploys 16 groups of the same GPU performance detection IP cores, and different GPU performance detection IP cores can perform detection tasks on different GPUs in parallel, and each task thread is independent of each other. Figure 4 As shown, the GPU performance detection IP core at least includes an input data cache module, a request decoding module, a detection item confirmation module, a memory interface generator MIG memory reading module, a data header verification module, a GPU performance detection example data decomposition module, a data analysis and mining example module, an image analysis and processing example module, an AI deep learning example module, a scientific computing simulation example module, a data merging and caching module, an example sending and receiving correctness confirmation finite state machine FSM module, a test result random access memory RAM cache module, a performance test result decomposition module, a GPU performance test arbitration calculation module, and an output data cache module.

[0064] In an embodiment of the present application, the FPGA includes at least: a MIG memory reading module for reading required operation data, a GPU performance test arbitration calculation module for running a preset GPU performance arbitration algorithm, and a GPU performance detection example data decomposition module for decomposing the required operation data.

[0065] Among them, the FPGA determines the performance detection task information of the corresponding target GPU to be detected based on the received GPU detection request, specifically including:

[0066] Through the detection item confirmation module, the required operation data corresponding to the detection task is determined, and the preset test reference coefficient corresponding to the detection task is determined, and the preset test reference coefficient is sent to the GPU performance test arbitration calculation module. Among them, the preset test parameter coefficient at least includes a multi-dimensional reference complexity and a multi-dimensional test item coefficient that match the test category corresponding to the detection task. According to the preset example dimension division rule, the required operation data is divided into corresponding multi-dimensional example data, and the preset example dimension division rule is used to identify the corresponding example dimension in the required operation data.

[0067] Specifically, the present application sends the required operation data to the GPU performance detection case data decomposition module through the MIG memory reading module to determine the corresponding multi-dimensional case data. The multi-dimensional case data includes at least the following case dimensions: data analysis and mining case, image analysis and processing case, AI deep learning case, scientific computing simulation case. The multi-dimensional case data is merged and processed to generate performance detection task information. In this way, performance detection task information is efficiently generated for performance detection of the target GPU to be detected.

[0068] Furthermore, when the GPU detection execution unit receives the GPU detection request, the input data cache module temporarily stores the input target GPU address and detection requirement instruction data; the request decoding module decodes the input data cache content to obtain the universal identification code preset for the FPGA execution detection, and transmits the decoding result to the detection item confirmation module; the detection item confirmation module identifies whether the target GPU address exists and confirms whether the required detection item is legal. After confirmation, the DDR5 address where the detection target required operation data is stored is sent to the MIG memory reading module; in addition, the detection item confirmation module also sends the task code (required operation data) corresponding to the detection requirement and preset test parameter coefficients such as task complexity to the GPU performance test arbitration calculation module, so as to facilitate the subsequent algorithm calculation of the actual performance result of the operation. The MIG memory reading module reads all the target demand operation data that the corresponding GPU needs to execute for this detection from DDR5, and sends it to the header verification module to confirm the accuracy of the data. After confirmation, all the target demand operation data are sent to the GPU performance detection case data decomposition module; the GPU performance detection case data decomposition module decomposes the read demand operation data into data analysis and mining case data, image analysis and processing case data, AI deep learning case data, scientific computing simulation case data and other data (which can be added at will according to user needs) and sends the decomposition results to the data analysis and mining case module, image analysis and processing case module, AI deep learning case module, scientific computing simulation case module and other modules for caching; the data merging and caching module reads specific categories of case data from the data analysis and mining case module, image analysis and processing case module, AI deep learning case module, scientific computing simulation case module and other modules according to the user's query requirements for the target GPU, and merges the data (the merged result is the test data that the target GPU actually needs to execute). Subsequently, the merged result is sent to the calculation example receiving and sending correctness confirmation FSM module, and the detection data is sent to the GPU to be detected at the target address through CXL packets in sequence.

[0069] S102, the FPGA sends the performance detection task information to the corresponding CXL address based on the preset CXL Switch, so that the target GPU to be detected executes the detection task.

[0070] Based on the preset CXL Switch, the performance detection task information is sent to the corresponding CXL address so that the target GPU to be detected can perform the detection task, including:

[0071] The performance detection task information is sent to the target GPU to be detected through the PCIe interface corresponding to the CXL Switch, where the CXL Switch corresponds to multiple GPUs.

[0072] Through CXL technology, communication between devices is achieved, and data transmission and communication efficiency is higher and more stable.

[0073] Among them, when the target GPU to be detected performs the detection task, it can perform the following Figure 3 In the process of operation, after receiving the instruction, the target GPU to be tested deployed on the CXL Switch performs a series of complex data reception, value acquisition, decoding, memory access, operation execution (data analysis and mining, image signal processing, AI deep learning, etc.), result collection and write-back, etc. according to the detection requirements, and sends the execution results back to the GPU detection execution unit in time. The target GPU to be tested can also judge whether the detection task can be executed normally during execution, so as to issue fault feedback or execute the detection task normally, so as to complete the GPU performance detection well and avoid the detection task failing to execute normally and delaying the normal detection process of the GPU.

[0074] S103, after receiving the task execution result from the target GPU to be detected, the FPGA determines the performance detection result corresponding to the target GPU to be detected according to the task execution result and a preset GPU performance arbitration algorithm, and sends the performance detection result to the user terminal.

[0075] The task execution result at least includes the measured parameters of the target GPU to be detected for executing the detection task. The measured parameters at least include the test data volume, execution time, CXL transmission bandwidth, and test item coefficient. The test item coefficient corresponds to at least the following dimensions: task complexity, model complexity, and precision complexity.

[0076] In the embodiment of the present application, before determining the performance test result corresponding to the target GPU to be tested according to the task execution result and the preset GPU performance arbitration algorithm, the method further includes:

[0077] Through the correctness confirmation module, determine whether the task execution result is correct. When it is determined that the task execution result is correct, the task execution result is decomposed according to the parameter dimension corresponding to the measured parameter to obtain the measured parameters used for the preset GPU performance arbitration algorithm calculation. Among them, the FPGA realizes parameter decomposition by sending the task execution result to the performance test result decomposition module, and decomposes the task execution result through the performance test result decomposition module to determine the measured parameters. And the measured parameters are input into the GPU performance test arbitration calculation module. Otherwise, a fault code corresponding to the target GPU to be tested is generated and sent to the user terminal.

[0078] In other words, FPGA can verify whether the task execution result is correct or not, so as to achieve performance test result decomposition or fault feedback, efficiently complete GPU performance detection, avoid obtaining erroneous performance detection results when the task execution result is abnormal, and improve the credibility of the performance detection result. The fault code is a custom code indicating the fault type, and this application does not make specific restrictions on this.

[0079] More specifically, refer to Figure 2 , the example receiving and sending correctness confirmation FSM module in FPGA waits for the target GPU to be tested to respond and then receives the task execution result, and the module verifies the result; and records the response time, transmission bandwidth and other information of the target GPU, and sends them together with the task execution result verification status to the test result RAM for storage; after all the test example data of the target GPU are executed, the test result RAM sends the result to the performance test result decomposition module; the performance test result decomposition module decomposes the response time, transmission bandwidth, result accuracy and other parameters required for the subsequent GPU performance test arbitration calculation module to execute the algorithm from the GPU test result; the GPU performance test arbitration calculation module executes the GPU performance arbitration algorithm according to the measured parameters of the target GPU to execute the task, and obtains the comprehensive performance score of the target GPU to execute related operations, and transmits the result to the output data cache module; the output data cache module packages the target GPU address, target GPU score and other data of the detection, and sends them to the user PC through the PCIe interface to complete the detection of the target GPU.

[0080] In the embodiment of the present application, the performance test result corresponding to the target GPU to be tested is determined according to the task execution result and the preset GPU performance arbitration algorithm, specifically including:

[0081] Calculate the first division result of the test data volume and CXL transmission bandwidth in the measured parameters , is the first intermediate value. The difference between the execution time and the first intermediate value is taken as the second intermediate value . Calculate the result of the second division operation of the test data volume and the second intermediate value and taking the product value of the second division operation result and the test item coefficient as the performance test arbitration calculation result value, so as to generate the performance detection result according to the performance test arbitration calculation result value.

[0082] The test item coefficients are determined by the following embodiments:

[0083] Determine the current test category corresponding to the detection task; determine the division result of the actual task complexity corresponding to the current test category and the preset reference task complexity, which is the task complexity coefficient; determine the division result of the actual model complexity corresponding to the current test category and the preset reference model complexity, which is the model complexity coefficient; determine the division result of the actual precision complexity corresponding to the current test category and the preset reference precision complexity, which is the precision complexity coefficient; determine the test item coefficient according to the product and value of the task complexity coefficient, model complexity coefficient, and precision complexity coefficient with the corresponding preset task complexity weight, preset model complexity weight, and preset precision complexity weight.

[0084] Among them, the preset GPU performance arbitration algorithm is as follows:

[0085] ;

[0086] in, The performance test arbitration calculation result value corresponding to the performance detection result. The amount of test data. The execution time can be understood as the total time taken for the parameters corresponding to the performance detection task information to be sent from the FPGA to the target GPU to be detected and for the GPU to return the task execution results. This is the CXL transmission bandwidth. is the test item coefficient corresponding to the current test category corresponding to the detection task. The test data volume can be understood as the amount of image data when executing a convolutional neural network on the pixel data of an image; another example is the amount of raw data from data mining; another example is the amount of raw data from AI deep learning; refer to Figure 3 The execution data on the GPU to be tested. The above current test categories include data analysis and mining, image signal processing, video signal processing, AI deep learning, convolutional neural network training, etc.

[0087] Among them, the corresponding formula of the test item coefficient is as follows:

[0088]

[0089] ;

[0090] in, is the actual task complexity corresponding to the current test category. The preset reference task complexity for the current test category. is the actual model complexity of the current test category. The preset reference model complexity for the current test category. is the actual accuracy complexity of the current test category. The preset reference accuracy complexity for the current test category. is the preset task complexity weight. is the preset model complexity weight. Preset precision complexity weight.

[0091] The preset reference precision complexity is such as the corresponding complexity of single-precision floating point number fp32, double-precision floating point number fp64, 32-bit integer int32, etc. , , , , It is set by the user according to the actual usage scenario, and this application does not make any specific limitations on this.

[0092] Through the above-mentioned preset GPU performance arbitration algorithm, the task execution results of the target GPU to be detected can be used to flexibly and efficiently calculate its corresponding performance detection results.

[0093] Figure 5 Another flowchart of a large-scale GPU performance detection method is shown in FIG. Figure 5 As shown, the following steps are included:

[0094] S501, the user publishes the CXL address of the target GPU to be detected and the detection task in the form of API through the PC terminal software;

[0095] S502, the CPU of the user PC sends a GPU detection request to the detection execution unit FPGA through the PCIe interface;

[0096] S503, the GPU detection execution unit FPGA reads the CXL address of the target GPU to be detected from the PC terminal cache, detailed detection tasks, detection execution data and other detailed information, and starts to execute the query;

[0097] S504, the detection execution unit FPGA packages and sends the detection task and detection execution data to the CXL address of the target GPU to be detected through the CXL Switch;

[0098] S505, the target GPU to be detected starts to execute the detection task instruction and processes the corresponding tasks (such as data analysis and mining, image signal processing, AI deep learning, etc.);

[0099] S506, judging whether the task can be executed normally, if it can be executed normally, jumping to step S507, if not, jumping to step S509;

[0100] S507, the FPGA of the GPU detection execution unit receives the execution result of the target GPU task to be detected;

[0101] S508, judging whether the result is correct, if the result is correct, jumping to step S5010, if not, jumping to step S509;

[0102] S509: The target GPU to be detected has a functional failure, and the detection execution unit FPGA sends a corresponding fault code to the PC end through PCIe;

[0103] S5010, the detection execution unit FPGA records the execution time of the target GPU to be detected, the CXL transmission bandwidth and other data, and combines the information such as the complexity of the detection task to execute the GPU performance arbitration calculation algorithm to calculate the actual performance test arbitration calculation result value of the GPU;

[0104] S5011, the detection execution unit FPGA sends the actual detection performance feedback of the corresponding GPU to the PC end through PCIe to complete the detection.

[0105] In the above process, since the FPGA chip of the GPU detection execution unit deploys 16 groups of identical GPU performance detection IPs, different IPs can execute detection tasks on different GPUs in parallel, and each task thread is independent of each other. The GPU detection execution unit supports simultaneous performance detection of multiple target GPUs in a single period, which has a high degree of parallelism. The PC terminal is a user terminal, which can be understood as the user's mobile phone, computer and other devices, and this application does not make specific restrictions on this.

[0106] Through the above technical solution, FPGA is used to participate in GPU performance testing to alleviate the pressure of CPU computing power loss in the current PC. At the same time, the above design can coordinate multiple GPUs to complete performance testing in a single period, thereby improving detection efficiency and user experience.

[0107] In addition, the large-scale GPU performance testing requirements that originally needed to be executed by the CPU in the PC of the GPU performance testing terminal are transferred to the GPU heterogeneous testing execution unit to realize parallel testing calculations, which greatly reduces the computing load of the PC of the GPU performance testing terminal and significantly improves the performance testing efficiency of large-scale GPUs.

[0108] This application also uses CXL technology to achieve communication between GPU performance detection terminals and multiple GPUs to be detected, with higher data transmission efficiency and greater stability. At the same time, this application provides a detailed application and scheduling strategy of a large-scale GPU performance detection architecture, which can better achieve unified scheduling control of large-scale GPU performance detection and significantly improve system performance.

[0109] Figure 6 A schematic diagram of the structure of a large-scale GPU performance detection system provided in an embodiment of the present application is shown in FIG. Figure 6 As shown, the system 600 includes:

[0110] The first determination module 601 is used to determine the performance detection task information of the corresponding target GPU to be detected based on the received GPU detection request. The performance detection task information includes at least a CXL address and a detection task. One CXL address corresponds to one GPU. The sending module 602 is used to send the performance detection task information to the corresponding CXL address based on a preset CXL Switch so that the target GPU to be detected performs the detection task. The second determination module 603 is used to determine the performance detection result corresponding to the target GPU to be detected according to the task execution result and a preset GPU performance arbitration algorithm after receiving the task execution result from the target GPU to be detected, and send the performance detection result to the user terminal. Among them, the task execution result includes at least the measured parameters of the target GPU to be detected performing the detection task.

[0111] In an embodiment of the present application, the system also includes: a device for detecting GPU performance includes a field programmable gate array FPGA, and the FPGA deploys multiple groups of GPU performance detection IP cores to perform parallel detection of different target GPUs to be detected through different GPU performance detection IP cores. The GPU performance detection IP core includes at least an input data cache module, a request decoding module, a detection item confirmation module, a memory interface generator MIG memory reading module, a data header verification module, a GPU performance detection example data decomposition module, a data analysis and mining example module, an image analysis and processing example module, an AI deep learning example module, a scientific computing simulation example module, a data merging cache module, an example receiving and sending correctness confirmation finite state machine FSM module, a test result random access memory RAM cache module, a performance test result decomposition module, a GPU performance test arbitration calculation module, and an output data cache module.

[0112] In the embodiment of the present application, the first determining module 601 is specifically used to:

[0113] Determine the required operation data corresponding to the detection task, and determine the preset test reference coefficient corresponding to the detection task. Among them, the preset test parameter coefficient at least includes the multidimensional reference complexity and multidimensional test item coefficient that match the test category corresponding to the detection task. According to the preset example dimension division rule, the required operation data is divided into corresponding multidimensional example data. Among them, the preset example dimension division rule is used to identify the corresponding example dimension in the required operation data. The multidimensional example data at least includes data analysis and mining examples, image analysis and processing examples, AI deep learning examples, and scientific computing simulation examples. The multidimensional example data is subjected to data merging and processing to generate performance detection task information.

[0114] The system can also:

[0115] Determine whether the task execution result is correct. If correct, decompose the task execution result according to the parameter dimension corresponding to the measured parameter to obtain the measured parameters used for the preset GPU performance arbitration algorithm calculation. Otherwise, generate a fault code corresponding to the target GPU to be detected and send it to the user terminal.

[0116] The system also includes: the measured parameters include at least the test data volume, execution time, CXL transmission bandwidth, and test item coefficient; the test item coefficient corresponds to at least the following dimensions: task complexity, model complexity, and accuracy complexity.

[0117] The second determination module 603 can specifically:

[0118] A first division result of the test data volume and the CXL transmission bandwidth in the measured parameters is calculated as a first intermediate value; the difference between the execution time and the first intermediate value is used as a second intermediate value; a second division result of the test data volume and the second intermediate value is calculated, and the product value of the second division result and the test item coefficient is used as a performance test arbitration calculation result value, so as to generate a performance detection result according to the performance test arbitration calculation result value.

[0119] The system also includes:

[0120] Determine the current test category corresponding to the detection task; determine the division result of the actual task complexity corresponding to the current test category and the preset reference task complexity as the task complexity coefficient; determine the division result of the actual model complexity corresponding to the current test category and the preset reference model complexity as the model complexity coefficient; determine the division result of the actual precision complexity corresponding to the current test category and the preset reference precision complexity as the precision complexity coefficient; determine the test item coefficient according to the product sum of the task complexity coefficient, the model complexity coefficient, the precision complexity coefficient and the corresponding preset task complexity weight, the preset model complexity weight, and the preset precision complexity weight.

[0121] The system can also:

[0122] The GPU detection request is received through the PCIe interface connected to the user terminal.

[0123] The sending module 602 is specifically used for:

[0124] The performance detection task information is sent to the target GPU to be detected through the PCIe interface corresponding to the CXL Switch, where the CXL Switch corresponds to multiple GPUs.

[0125] Figure 7 Schematic diagram of the structure of a large-scale GPU performance detection device in an embodiment of the present application. Figure 7 As shown, the device includes:

[0126] At least one processor; and a memory in communication with the at least one processor. The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to:

[0127] Based on the received GPU detection request, the performance detection task information of the corresponding target GPU to be detected is determined. The performance detection task information includes at least a CXL address and a detection task. One CXL address corresponds to one GPU. Based on the preset CXLSwitch, the performance detection task information is sent to the corresponding CXL address so that the target GPU to be detected performs the detection task. After receiving the task execution result from the target GPU to be detected, the performance detection result corresponding to the target GPU to be detected is determined according to the task execution result and the preset GPU performance arbitration algorithm, and the performance detection result is sent to the user terminal. Among them, the task execution result includes at least the measured parameters of the target GPU to be detected performing the detection task.

[0128] Each embodiment in this application is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system and device embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0129] The system and equipment provided in the embodiments of the present application correspond one-to-one to the method. Therefore, the system and equipment also have similar beneficial technical effects as the corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the system and equipment will not be repeated here.

[0130] It should also be noted that the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of more restrictions, the elements defined by the sentence "comprises a ..." do not exclude the existence of other identical elements in the process, method, commodity or device including the elements.

[0131] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A large-scale GPU performance detection method, characterized in that: The method is applied to a large-scale GPU performance detection system; the method comprises: Based on the received GPU detection request, determine the performance detection task information of the corresponding target GPU to be detected; the performance detection task information at least includes a calculation fast link CXL address and a detection task; one CXL address corresponds to one GPU; Based on a preset computation fast link switch CXL Switch, the performance detection task information is sent to the corresponding CXL address, so that the target GPU to be detected performs the detection task; After receiving the task execution result from the target GPU to be detected, determining the performance detection result corresponding to the target GPU to be detected according to the task execution result and a preset GPU performance arbitration algorithm, and sending the performance detection result to the user terminal; wherein the task execution result at least includes the measured parameters of the target GPU to be detected executing the detection task; Wherein, based on the received GPU detection request, determining the performance detection task information of the corresponding target GPU to be detected specifically includes: Determine the required operation data corresponding to the detection task, and determine the preset test reference coefficient corresponding to the detection task; wherein the preset test reference coefficient at least includes a multidimensional reference complexity and a multidimensional test item coefficient that match the test category corresponding to the detection task; According to the preset example dimension division rule, the demand operation data is divided into corresponding multi-dimensional example data; wherein the preset example dimension division rule is used to identify the corresponding example dimension in the demand operation data; the multi-dimensional example data at least includes data analysis and mining examples, image analysis and processing examples, AI deep learning examples, and scientific computing simulation examples; Performing data merging processing on the multi-dimensional example data to generate the performance detection task information; The measured parameters include at least the amount of test data, execution time, CXL transmission bandwidth, and test item coefficient; the test item coefficient corresponds to at least the following dimensions: task complexity, model complexity, and precision complexity; According to the task execution result and the preset GPU performance arbitration algorithm, the performance test result corresponding to the target GPU to be tested is determined, specifically including: Calculate a first division result of the test data volume in the measured parameter and the CXL transmission bandwidth as a first intermediate value; The difference between the execution time and the first intermediate value is used as the second intermediate value; Calculate a second division result of the test data volume and the second intermediate value, and use the product value of the second division result and the test item coefficient as a performance test arbitration calculation result value, so as to generate the performance detection result according to the performance test arbitration calculation result value; The preset GPU performance arbitration algorithm is as follows: ; in, The performance test arbitration calculation result value corresponding to the performance detection result; is the test data volume; The execution is time-consuming; is the CXL transmission bandwidth; is the test item coefficient corresponding to the current test category corresponding to the detection task.

2. A large-scale GPU performance detection method according to claim 1, characterized in that: The device for detecting GPU performance includes a field programmable gate array FPGA, and the FPGA deploys multiple groups of GPU performance detection IP cores to parallelly detect different target GPUs to be detected through different GPU performance detection IP cores; the GPU performance detection IP core at least includes an input data cache module, a request decoding module, a detection item confirmation module, a memory interface generator MIG memory reading module, a data header verification module, a GPU performance detection example data decomposition module, a data analysis and mining example module, an image analysis and processing example module, an AI deep learning example module, a scientific computing simulation example module, a data merging cache module, an example sending and receiving correctness confirmation finite state machine FSM module, a test result random access memory RAM cache module, a performance test result decomposition module, a GPU performance test arbitration calculation module, and an output data cache module.

3. A large-scale GPU performance detection method according to claim 1, characterized in that: Before determining the performance test result corresponding to the target GPU to be tested according to the task execution result and the preset GPU performance arbitration algorithm, the method further includes: Determining whether the task execution result is correct; If correct, decomposing the task execution result according to the parameter dimensions corresponding to the measured parameters to obtain the measured parameters used for calculation of the preset GPU performance arbitration algorithm; Otherwise, a fault code corresponding to the target GPU to be detected is generated and sent to the user terminal.

4. The large-scale GPU performance detection method according to claim 1, characterized in that: Before using the product value of the second division operation result and the test item coefficient as the performance test arbitration calculation result value, the method further includes: Determining a current test category corresponding to the detection task; Determine a task complexity coefficient as a division result of an actual task complexity corresponding to the current test category and a preset reference task complexity; Determine a division result of the actual model complexity corresponding to the current test category and the preset reference model complexity as a model complexity coefficient; Determine a division result of the actual precision complexity corresponding to the current test category and the preset reference precision complexity as the precision complexity coefficient; The test item coefficient is determined according to the product and value of the task complexity coefficient, the model complexity coefficient, the precision complexity coefficient and the corresponding preset task complexity weight, preset model complexity weight, and preset precision complexity weight.

5. The large-scale GPU performance detection method according to claim 1, characterized in that: Before determining the performance detection task information of the corresponding target GPU to be detected based on the received GPU detection request, the method further includes: Receiving the GPU detection request via a high-speed serial computer expansion bus standard PCIe interface connected to the user terminal; Based on a preset CXL Switch, the performance detection task information is sent to the corresponding CXL address so that the target GPU to be detected performs the detection task, specifically including: The performance detection task information is sent to the target GPU to be detected through a PCIe interface corresponding to the CXL Switch; wherein the CXL Switch corresponds to multiple GPUs.

6. A large-scale GPU performance detection system, characterized in that: The system comprises: A first determination module is used to determine the performance detection task information of the corresponding target GPU to be detected based on the received GPU detection request; the performance detection task information at least includes a calculation fast link CXL address and a detection task; one CXL address corresponds to one GPU; A sending module, configured to send the performance detection task information to the corresponding CXL address based on a preset computation fast link switch CXL Switch, so that the target GPU to be detected performs the detection task; A second determination module is used to determine, after receiving the task execution result from the target GPU to be detected, the performance detection result corresponding to the target GPU to be detected according to the task execution result and a preset GPU performance arbitration algorithm, and send the performance detection result to the user terminal; wherein the task execution result at least includes the measured parameters of the target GPU to be detected executing the detection task; Wherein, the first determining module is specifically used for: Determine the required operation data corresponding to the detection task, and determine the preset test reference coefficient corresponding to the detection task; wherein the preset test reference coefficient at least includes a multidimensional reference complexity and a multidimensional test item coefficient that match the test category corresponding to the detection task; According to the preset example dimension division rule, the demand operation data is divided into corresponding multi-dimensional example data; wherein the preset example dimension division rule is used to identify the corresponding example dimension in the demand operation data; the multi-dimensional example data at least includes data analysis and mining examples, image analysis and processing examples, AI deep learning examples, and scientific computing simulation examples; Performing data merging processing on the multi-dimensional example data to generate the performance detection task information; The measured parameters include at least the amount of test data, execution time, CXL transmission bandwidth, and test item coefficient; the test item coefficient corresponds to at least the following dimensions: task complexity, model complexity, and precision complexity; The second determination module can specifically: Calculate a first division result of the test data volume in the measured parameter and the CXL transmission bandwidth as a first intermediate value; The difference between the execution time and the first intermediate value is used as the second intermediate value; Calculate a second division result of the test data volume and the second intermediate value, and use the product value of the second division result and the test item coefficient as a performance test arbitration calculation result value, so as to generate the performance detection result according to the performance test arbitration calculation result value; The preset GPU performance arbitration algorithm is as follows: ; in, The performance test arbitration calculation result value corresponding to the performance detection result; is the test data volume; The execution is time-consuming; is the CXL transmission bandwidth; is the test item coefficient corresponding to the current test category corresponding to the detection task.

7. A large-scale GPU performance testing device, characterized in that: The device comprises: 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, and the instructions are executed by the at least one processor so that the at least one processor can execute a large-scale GPU performance detection method as described in any one of claims 1-5 above.