General detection method for GPU (Graphics Processing Unit) chip

Abstract

The invention provides a general detection method for a GPU (Graphics Processing Unit) chip, and relates to the field of GPU chip detection. The method comprises the following steps: adjusting a general detection process for a to-be-detected GPU chip according to GPU chip data, and applying the adjusted general detection process to the to-be-detected GPU chip; according to the general detection process, algorithm correctness detection, algorithm performance detection, interface test, access test, firmware test, environment detection test, data read-write detection, storage space detection and chip function test are executed on GPU chips one by one, and all the detection processes are easy to implement and suitable for detection of various GPU chips; besides, the setting sequence of each detection process accords with the sequence of objects related to the working process of the GPU chip, and the sequence of test execution and the test objects are reasonably set, so that the time of the test process can be effectively shortened, and the test efficiency can be improved; therefore, normal operation, function realization and data security of the GPU chip are guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of GPU chip detection, and in particular to a general detection method for GPU chips. Background Art

[0002] In today's digital age, computer technology is developing rapidly, and various computing-intensive applications are emerging, such as artificial intelligence, machine learning, and big data processing. With the popularity of these applications, GPU (graphics processing unit) plays an increasingly important role in computer systems. The high parallel computing power and efficient data processing capabilities of GPU make it the preferred device for computing-intensive applications.

[0003] However, with the popularity of GPUs, some security issues have also arisen. As an important device to protect data and computing processes from malicious attacks and unauthorized access, GPU chips play a key role in various computing-intensive applications, which also makes GPUs an important target for malicious attacks and unauthorized access. If a GPU chip fails or is abnormal, it may lead to security issues such as data leakage, data corruption, and system crashes. Therefore, the normal operation and function of GPU chips are the prerequisites for achieving data security and privacy protection.

[0004] Therefore, it is necessary to propose an efficient and reliable universal detection method for GPU chips to ensure normal operation. Summary of the invention

[0005] To solve the above technical problems, the present invention provides a general detection method for GPU chips, which adjusts the general detection process for the CPU chip to be detected according to the GPU chip data, and uses the adjusted general detection process for the GPU chip to be detected; wherein the general detection process performs algorithm correctness detection, algorithm performance detection, interface test, access test, firmware test, environment detection test, data read and write detection, storage space detection and chip function test on the CPU chip one by one, and during the execution of each test process, determines whether the GPU chip fails the test or is unsafe; if any test fails or is unsafe, the GPU chip is marked and the corresponding test process information is output.

[0006] As a further solution, before executing the algorithm correctness test, the supported algorithms are identified based on the GPU chip information, and the corresponding input data and expected results are predefined for each algorithm; when executing the algorithm correctness test, the algorithm is implemented on the GPU chip and the corresponding input data is imported, waiting for the GPU chip to execute the algorithm and obtain the output result, and verifying whether the output result is consistent with the expected result. If the result of the GPU chip calculation is inconsistent with the expected result, the output test fails.

[0007] As a further solution, the algorithms supported by the GPU chip are identified based on the GPU chip data and the algorithm performance is tested. The algorithm performance test uses a fixed data set to test the execution time of the algorithm. After repeating the performance test multiple times, the average execution time of the algorithm is calculated, and the computing efficiency of the GPU chip is evaluated by the average execution time. If the computing efficiency of the GPU chip is lower than the normal computing efficiency, the output test fails.

[0008] As a further solution, the interface provided by the GPU chip is obtained according to the GPU chip data and the interface test is performed; the interface test includes: Test the working voltage of the pins of the interface; if the working voltage of the pins is not within the specified range, the output test fails; Measure the resistance between the interface pins; if there is a short circuit or open circuit, the output test fails; Send and receive data through the interface in a specific data mode; if data cannot be sent or received normally, the output test fails; According to the protocol standard of the interface, check whether the communication behavior of the chip interface complies with the specifications of the protocol standard; if it is inconsistent with the specifications of the protocol standard, the output test fails.

[0009] As a further solution, the access control mechanism is obtained according to the GPU chip data and corresponding access tests are performed, including access limitation test and access protection test; Access restriction test: Access sensitive data when access permissions are met; if normal access is not possible, the output test fails; Access sensitive data when access rights are not met; confirm that the GPU chip starts the protection mechanism and cannot be accessed normally, otherwise the output test fails; Protection mechanism test: Set specific users or programs to perform GPU-sensitive operations based on access rights; if the user or program is allowed to fail to execute normally, or is not allowed to execute normally, the output is unsafe; When the protection mechanism is started, the source code is checked to determine whether the GPU sensitive information is stored encrypted; if storage encryption is not performed, the output is unsafe.

[0010] As a further solution, the firmware supported by the GPU chip is identified according to the GPU chip data and firmware testing is performed, including firmware execution testing, firmware anomaly testing, firmware integrity testing, and firmware startup mode testing; Firmware execution test: execute the firmware function; if the firmware function cannot be fully executed, the output test fails; Firmware exception test: Check the effectiveness of the firmware exception handling mechanism; if the firmware exception handling mechanism cannot be implemented normally, the output test fails; Firmware integrity test: Check the integrity protection mechanism of the firmware; if there is no integrity protection mechanism, the output is unsafe; Firmware startup method test: Verify the firmware startup method to confirm that the firmware loaded when the GPU chip starts is authenticated and not tampered with, otherwise the output is unsafe.

[0011] As a further solution, the working environment parameters and their normal range values, maximum range values ​​and minimum range values ​​are set according to the GPU chip data and the environmental detection test is performed; among them, If the working environment parameters are within the normal range; if the GPU chip works abnormally, the output test fails; Iteratively increase the pressure of the working environment parameters according to the pressure increase strategy until the highest range value is reached; when the highest range value is reached, if the GPU chip still remains in the running state, the output test fails; The working environment parameters are iteratively decompressed according to the decompression strategy until the lowest range value is reached; when the lowest range value is reached, if the GPU chip still remains in operation, the output test fails.

[0012] As a further solution, the data erasure mode is set according to the GPU chip information and data read and write detection is performed; data read and write detection: the GPU chip is under high temperature conditions, and the GPU chip is cyclically written-read-erase operations are performed on the GPU chip to write the specified verification data into the GPU chip; the GPU chip is kept under high temperature conditions until the high temperature test time, the data in the GPU chip is read and compared with the specified verification data, if the data is inconsistent, the output test fails.

[0013] As a further solution, the Flash space size of the GPU chip is obtained according to the GPU chip data and storage space detection is performed. Storage space detection: prepare the data volume corresponding to the Flash space size, download the data to the Flash space for storage, and verify whether the data volume after writing is consistent with the Flash space size of the GPU chip; if not, the output test fails.

[0014] As a further solution, the operating environment of the GPU chip is obtained according to the GPU chip data and the chip function test is performed. The chip function test: select the CNN model and standard data set, and train the model through the GPU chip; among them, Observe the changes in the loss and accuracy of the CNN model during model training. If the changes in the loss and accuracy of the CNN model are abnormal, the output test fails. Observe the memory usage of the GPU chip during model training. If the memory usage does not increase during model training, the output test fails. Use the same CNN model and standard data set to train the model on the CPU chip and GPU chip respectively, and compare the accuracy of the CNN models trained by the two. If there is a difference in accuracy, the output test fails.

[0015] Compared with the related art, the GPU chip universal detection method provided by the present invention has the following beneficial effects: The present invention adjusts a general detection process for a GPU chip to be detected according to GPU chip data, and uses the adjusted general detection process for the GPU chip to be detected; wherein the general detection process performs algorithm correctness detection, algorithm performance detection, interface test, access test, firmware test, environment detection test, data read and write detection, storage space detection and chip function test on the GPU chip one by one, each detection process is easy to implement and is applicable to the detection of various GPU chips; in addition, the setting sequence of each detection process is consistent with the sequence of objects involved in the work flow of the GPU chip, and the sequence of test execution and the test objects are reasonably set, which can effectively reduce the test process time and improve the test efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of the steps of a general detection method for GPU chips provided by the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0018] Example 1 This embodiment provides a general GPU chip detection method. For its processing flow, please refer to Figure 1 , adjust the general detection process for the GPU chip to be detected according to the GPU chip data, and use the adjusted general detection process for the GPU chip to be detected; wherein, the general detection process performs algorithm correctness detection, algorithm performance detection, interface test, access test, firmware test, environment detection test, data read and write detection, storage space detection and chip function test on the GPU chip one by one, and during the execution of each test process, judge whether the GPU chip fails the test or is unsafe; if any test fails or is unsafe, mark the GPU chip and output the corresponding test process information.

[0019] It should be noted that: this embodiment detects the correctness of the GPU chip algorithm, interface testing, access control, firmware testing, environmental testing, data reading and writing security, etc. Each detection process is easy to implement and is applicable to the detection of various GPU chips; the setting order of each detection process is consistent with the order of objects involved in the workflow of the GPU chip. Reasonable setting of the order of test execution and test objects can effectively reduce the test process time and improve test efficiency.

[0020] As a further solution, before executing the algorithm correctness test, the supported algorithms are identified based on the GPU chip information, and the corresponding input data and expected results are predefined for each algorithm; when executing the algorithm correctness test, the algorithm is implemented on the GPU chip and the corresponding input data is imported, waiting for the GPU chip to execute the algorithm and obtain the output result, and verifying whether the output result is consistent with the expected result. If the result of the GPU chip calculation is inconsistent with the expected result, the output test fails.

[0021] As a further solution, the algorithms supported by the GPU chip are identified based on the GPU chip data and the algorithm performance is tested. The algorithm performance test uses a fixed data set to test the execution time of the algorithm. After repeating the performance test multiple times, the average execution time of the algorithm is calculated, and the computing efficiency of the GPU chip is evaluated by the average execution time. If the computing efficiency of the GPU chip is lower than the normal computing efficiency, the output test fails.

[0022] As a further solution, the interface provided by the GPU chip is obtained according to the GPU chip data and the interface test is performed; the interface test includes: Test the working voltage of the pins of the interface; if the working voltage of the pins is not within the specified range, the output test fails; Measure the resistance between the interface pins; if there is a short circuit or open circuit, the output test fails; Send and receive data through the interface in a specific data mode; if data cannot be sent or received normally, the output test fails; According to the protocol standard of the interface, check whether the communication behavior of the chip interface complies with the specifications of the protocol standard; if it is inconsistent with the specifications of the protocol standard, the output test fails.

[0023] As a further solution, the access control mechanism is obtained according to the GPU chip data and corresponding access tests are performed, including access limitation test and access protection test; Access restriction test: Access sensitive data when access permissions are met; if normal access is not possible, the output test fails; Access sensitive data when access rights are not met; confirm that the GPU chip starts the protection mechanism and cannot be accessed normally, otherwise the output test fails; Protection mechanism test: Set specific users or programs to perform GPU-sensitive operations based on access rights; if the user or program is allowed to fail to execute normally, or is not allowed to execute normally, the output is unsafe; When the protection mechanism is started, the source code is checked to determine whether the GPU sensitive information is stored encrypted; if storage encryption is not performed, the output is unsafe.

[0024] As a further solution, the firmware supported by the GPU chip is identified according to the GPU chip data and firmware testing is performed, including firmware execution testing, firmware anomaly testing, firmware integrity testing, and firmware startup mode testing; Firmware execution test: execute the firmware function; if the firmware function cannot be fully executed, the output test fails; Firmware exception test: Check the effectiveness of the firmware exception handling mechanism; if the firmware exception handling mechanism cannot be implemented normally, the output test fails; Firmware integrity test: Check the integrity protection mechanism of the firmware; if there is no integrity protection mechanism, the output is unsafe; Firmware startup method test: Verify the firmware startup method to confirm that the firmware loaded when the GPU chip starts is authenticated and not tampered with, otherwise the output is unsafe.

[0025] As a further solution, the working environment parameters and their normal range values, maximum range values ​​and minimum range values ​​are set according to the GPU chip data and the environmental detection test is performed; among them, If the working environment parameters are within the normal range; if the GPU chip works abnormally, the output test fails; Iteratively increase the pressure of the working environment parameters according to the pressure increase strategy until the highest range value is reached; when the highest range value is reached, if the GPU chip still remains in the running state, the output test fails; The working environment parameters are iteratively decompressed according to the decompression strategy until the lowest range value is reached; when the lowest range value is reached, if the GPU chip still remains in operation, the output test fails.

[0026] It should be noted that the working environment parameters mainly include working voltage, working frequency and working temperature; among them, Working voltage boost strategy: The GPU chip runs the test program under normal working voltage, increasing the working voltage by 0.2V until it exceeds the maximum working voltage of the chip.

[0027] Working voltage decompression strategy: The GPU chip runs the test program under normal working voltage, reducing the working voltage by 0.2V until it exceeds the minimum working voltage of the chip.

[0028] Operating frequency boost strategy: The GPU chip runs the test program at its normal operating frequency, increasing the operating frequency by 500KHz until it exceeds the chip's maximum operating frequency.

[0029] Operating frequency decompression strategy: The GPU chip runs the test program at its normal operating frequency, reducing the operating frequency by 500KHz until it exceeds the chip's minimum operating frequency.

[0030] Operating temperature boost strategy: The GPU chip runs the test program at normal operating temperature, increasing the operating temperature by 5°C until it exceeds the maximum operating temperature of the chip.

[0031] Operating temperature decompression strategy: The GPU chip runs the test program at normal operating temperature, reducing the operating temperature by 5°C until it exceeds the minimum operating temperature of the chip.

[0032] As a further solution, the data erasure mode is set according to the GPU chip information and data read and write detection is performed; data read and write detection: the GPU chip is under high temperature conditions, and the GPU chip is cyclically written-read-erase operations are performed on the GPU chip to write the specified verification data into the GPU chip; the GPU chip is kept under high temperature conditions until the high temperature test time, the data in the GPU chip is read and compared with the specified verification data, if the data is inconsistent, the output test fails.

[0033] Among them, the high temperature conditions and normal operating temperature of each GPU chip are different, but they can be determined by consulting the GPU chip manual or related standard materials. No specific restrictions are made here and the actual query results shall prevail.

[0034] As a further solution, the Flash space size of the GPU chip is obtained according to the GPU chip data and storage space detection is performed. Storage space detection: prepare the data volume corresponding to the Flash space size, download the data to the Flash space for storage, and verify whether the data volume after writing is consistent with the Flash space size of the GPU chip; if not, the output test fails.

[0035] As a further solution, the operating environment of the GPU chip is obtained according to the GPU chip data and the chip function test is performed. The chip function test: select the CNN model and standard data set, and train the model through the GPU chip; among them, Observe the changes in the loss and accuracy of the CNN model during model training. If the changes in the loss and accuracy of the CNN model are abnormal, the output test fails. Observe the memory usage of the GPU chip during model training. If the memory usage does not increase during model training, the output test fails. Use the same CNN model and standard data set to train the model on the CPU chip and GPU chip respectively, and compare the accuracy of the CNN models trained by the two. If there is a difference in accuracy, the output test fails.

[0036] It should be noted that GPU chips are suitable for a large number of floating-point operations and have a faster speed when training models than CPU chips, but the computing architecture of CPU chips is more rigorous and strictly follows logical operations; therefore, when comparing the accuracy, we use the CNN model trained by the CPU chip as the standard model to verify whether the accuracy of the GPU chip meets the standard.

[0037] The above are only embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A general detection method for GPU chips, characterized in that: The general detection process for the GPU chip to be detected is adjusted according to the GPU chip data, and the adjusted general detection process is used for the GPU chip to be detected; wherein the general detection process performs algorithm correctness detection, algorithm performance detection, interface test, access test, firmware test, environment detection test, data read and write detection, storage space detection and chip function test on the GPU chip one by one, and during the execution of each test process, it is determined whether the GPU chip fails the test or is unsafe; if any test fails or is unsafe, the GPU chip is marked and the corresponding test process information is output.

2. A general detection method for GPU chips according to claim 1, characterized in that: Before executing the algorithm correctness test, identify the supported algorithms based on the GPU chip information, and predefine the corresponding input data and expected results for each algorithm; when executing the algorithm correctness test, implement the algorithm into the GPU chip and import the corresponding input data, wait for the GPU chip to execute the algorithm and obtain the output result, and verify whether the output result is consistent with the expected result. If the result of the GPU chip calculation is inconsistent with the expected result, the output test fails.

3. A general detection method for GPU chips according to claim 1, characterized in that: According to the GPU chip data, the algorithms supported by the GPU chip are identified and the algorithm performance is tested. Among them, the algorithm performance test uses a fixed data set to test the execution time of the algorithm. After repeating the performance test multiple times, the average execution time of the algorithm is calculated, and the computing efficiency of the GPU chip is evaluated by the average execution time. If the computing efficiency of the GPU chip is lower than the normal computing efficiency, the output test fails.

4. The general detection method for GPU chips according to claim 1, characterized in that: According to the GPU chip data, obtain the interface provided by the GPU chip and perform interface testing; the interface testing includes: Test the working voltage of the pins of the interface; if the working voltage of the pins is not within the specified range, the output test fails; Measure the resistance between the interface pins; if there is a short circuit or open circuit, the output test fails; Send and receive data through the interface in a specific data mode; if data cannot be sent or received normally, the output test fails; According to the protocol standard of the interface, check whether the communication behavior of the chip interface complies with the specifications of the protocol standard; if it is inconsistent with the specifications of the protocol standard, the output test fails.

5. A general detection method for GPU chips according to claim 1, characterized in that: Obtain access control mechanism based on GPU chip data and perform corresponding access tests, including access limitation test and access protection test; Access restriction test: Access sensitive data when access permissions are met; if normal access is not possible, the output test fails; Access sensitive data when access rights are not met; confirm that the GPU chip starts the protection mechanism and cannot be accessed normally, otherwise the output test fails; Protection mechanism test: Set specific users or programs to perform GPU-sensitive operations based on access rights; if the user or program is allowed to fail to execute normally, or is not allowed to execute normally, the output is unsafe; When the protection mechanism is started, the source code is checked to determine whether the GPU sensitive information is stored encrypted; if storage encryption is not performed, the output is unsafe.

6. A general detection method for GPU chips according to claim 1, characterized in that: According to the GPU chip data, the firmware supported by the GPU chip is identified and firmware tests are performed, including firmware execution test, firmware exception test, firmware integrity test and firmware startup mode test; among them, Firmware execution test: execute the firmware function; if the firmware function cannot be fully executed, the output test fails; Firmware exception test: Check the effectiveness of the firmware exception handling mechanism; if the firmware exception handling mechanism cannot be implemented normally, the output test fails; Firmware integrity test: Check the integrity protection mechanism of the firmware; if there is no integrity protection mechanism, the output is unsafe; Firmware startup method test: Verify the firmware startup method to confirm that the firmware loaded when the GPU chip starts is authenticated and not tampered with, otherwise the output is unsafe.

7. A general detection method for GPU chips according to claim 1, characterized in that: According to the GPU chip data, set the working environment parameters and their normal range values, maximum range values ​​and minimum range values ​​and perform environmental detection tests; among them, If the working environment parameters are within the normal range; if the GPU chip works abnormally, the output test fails; Iteratively increase the pressure of the working environment parameters according to the pressure increase strategy until the highest range value is reached; when the highest range value is reached, if the GPU chip still remains in the running state, the output test fails; The working environment parameters are iteratively decompressed according to the decompression strategy until the lowest range value is reached; when the lowest range value is reached, if the GPU chip still remains in operation, the output test fails.

8. A general detection method for GPU chips according to claim 7, characterized in that: Set the data erase mode according to the GPU chip information and perform data read and write detection; Data read and write detection: Under high temperature conditions, the GPU chip is cyclically written-read-erase operations on the GPU chip to write the specified verification data into the GPU chip; Keep the GPU chip under high temperature conditions until the high temperature test time, read the data in the GPU chip and compare it with the specified verification data, if the data is inconsistent, the output test fails.

9. A general detection method for GPU chips according to claim 1, characterized in that: Get the Flash space size of the GPU chip according to the GPU chip data and perform storage space detection. Storage space detection: prepare the data volume corresponding to the Flash space size, download the data to the Flash space for storage, and verify whether the data volume after writing is consistent with the Flash space size of the GPU chip; if not, the output test fails.

10. A general detection method for GPU chips according to claim 1, characterized in that: According to the GPU chip data, the operating environment of the GPU chip is obtained and the chip function test is performed. The chip function test: select the CNN model and standard data set, and perform model training through the GPU chip; among them, Observe the changes in the loss and accuracy of the CNN model during model training. If the changes in the loss and accuracy of the CNN model are abnormal, the output test fails. Observe the memory usage of the GPU chip during model training. If the memory usage does not increase during model training, the output test fails. Use the same CNN model and standard data set to train the model on the CPU chip and GPU chip respectively, and compare the accuracy of the CNN models trained by the two. If there is a difference in accuracy, the output test fails.

Images