Temperature control method, device and equipment for artificial intelligence chip test, storage medium and program product

Abstract

The invention relates to a temperature control method and device for an artificial intelligence chip test, equipment, a storage medium and a program product, and relates to the technical field of artificial intelligence chips. According to the invention, the accuracy of artificial intelligence chip testing can be effectively guaranteed. The method comprises the following steps: acquiring a current voltage and a current frequency for a current test of the artificial intelligence chip, and inputting the current voltage and the current frequency into a first prediction model to output current predicted power consumption and a current predicted load rate of the artificial intelligence chip by the first prediction model, and inputting the current voltage, the current frequency, the current predicted power consumption, the current predicted load rate, the current initial temperature of the temperature control module and the current test duration used for the current test of the artificial intelligence chip into a second prediction model, so that the second prediction model outputs the current predicted temperature of the artificial intelligence chip. And sending a temperature adjustment instruction to the temperature control module according to the current predicted temperature, the test demand temperature for the current test of the artificial intelligence chip and the current initial temperature.

Description

Temperature control methods, devices, equipment, storage media, and software products for testing artificial intelligence chips. Technical Field

[0001] This application relates to the field of artificial intelligence chip technology, and in particular to a temperature control method, apparatus, equipment, storage medium, and program product for testing artificial intelligence chips. Background Technology

[0002] Voltage-frequency (VF) parameterization testing of AI chips is a crucial step in the evaluation and verification of AI chips. AI chips are hardware chips specifically designed and optimized for AI tasks, including but not limited to GPUs (Graphics Processing Units), NPUs (Neural Network Processing Units), and GPGPUs (General-Purpose Computing on Graphics Processing Units).

[0003] Voltage-frequency parameterization testing of artificial intelligence chips requires precise temperature control to ensure test accuracy.

[0004] In the current automated verification process for artificial intelligence chips, the temperature control method typically involves changing the load on the AI ​​chip and then setting the ambient temperature. However, this method can easily lead to distortion of the test data for the AI ​​chip, making it difficult to guarantee the accuracy of the test. Summary of the Invention

[0005] Therefore, it is necessary to provide a temperature control method, apparatus, equipment, storage medium, and program product for testing artificial intelligence chips to address the aforementioned technical problems.

[0006] Firstly, this application provides a temperature control method for testing artificial intelligence chips, including:

[0007] Obtain the current voltage and frequency used for the current test of the artificial intelligence chip;

[0008] The current voltage and current frequency are input into the first prediction model; the first prediction model is used to output the current predicted power consumption and current predicted load rate of the artificial intelligence chip based on the current voltage and current frequency.

[0009] The current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature of the temperature control module, and current test duration for the current test of the artificial intelligence chip are input into the second prediction model; the second prediction model is used to output the current predicted temperature of the artificial intelligence chip based on the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature, and current test duration.

[0010] Before performing the current test, a temperature adjustment command for the current test of the artificial intelligence chip is sent to the temperature control module based on the current predicted temperature, the test requirement temperature, and the current initial temperature; the test requirement temperature is the required temperature for the current test of the artificial intelligence chip.

[0011] In one embodiment, sending a temperature adjustment command for the current testing of the artificial intelligence chip to the temperature control module based on the current predicted temperature, the required test temperature, and the current initial temperature includes: obtaining a temperature adjustment amount for the current testing of the artificial intelligence chip based on the current predicted temperature and the required test temperature; and sending the temperature adjustment command for the current testing of the artificial intelligence chip to the temperature control module based on the temperature adjustment amount and the current initial temperature.

[0012] In one embodiment, the method further includes:

[0013] Based on the sample dataset, model training samples are obtained; the sample dataset includes historical test data of AI chip samples and historical temperature data of corresponding temperature control module samples; the model training samples include voltage samples, frequency samples, power consumption samples, load rate samples, test duration samples, temperature samples of AI chip samples and initial temperature samples of corresponding temperature control module samples; the first prediction model is trained based on the voltage samples, frequency samples, power consumption samples, and load rate samples of the AI ​​chip samples; the second prediction model is trained based on the voltage samples, frequency samples, predicted power consumption samples and predicted load rate samples output by the first prediction model, and the initial temperature samples and test duration samples.

[0014] In one embodiment, training the first prediction model based on voltage samples, frequency samples, power consumption samples, and load rate samples of the AI ​​chip sample includes: inputting the voltage and frequency samples of the AI ​​chip sample into the first prediction model to obtain predicted power consumption samples and predicted load rate samples output by the first prediction model; training the first prediction model based on the error between the predicted power consumption sample and the power consumption sample, and the error between the predicted load rate sample and the load rate sample; and / or, training the second prediction model based on the voltage and frequency samples of the AI ​​chip sample, the predicted power consumption samples and predicted load rate samples output by the first prediction model, and the initial temperature sample and test duration sample includes: inputting the voltage and frequency samples of the AI ​​chip sample, the predicted power consumption samples and predicted load rate samples output by the first prediction model, and the initial temperature sample and test duration sample into the second prediction model to obtain predicted temperature samples output by the second prediction model; training the second prediction model based on the error between the predicted temperature sample and the temperature sample.

[0015] In one embodiment, after sending a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module based on the current predicted temperature, the test requirement temperature, and the current initial temperature, the method further includes: recording the test data of the current test; if the number of tests meets the target condition, adding the test data recorded each time to the sample dataset, and retraining the first prediction model and the second prediction model based on the sample dataset.

[0016] In one embodiment, the method further includes: collecting the voltage, frequency, power consumption, load rate, temperature of the artificial intelligence chip and the temperature of the temperature control module at sampling time points; generating a load curve based on the sampling time points and the voltage, frequency, power consumption, load rate, temperature of the artificial intelligence chip and the temperature of the temperature control module; and outputting load warning information of the artificial intelligence chip.

[0017] Secondly, this application also provides a temperature control device for testing artificial intelligence chips, comprising:

[0018] The acquisition module is used to acquire the current voltage and current frequency for the current test of the artificial intelligence chip;

[0019] The first prediction module is used to input the current voltage and current frequency into the first prediction model; the first prediction model is used to output the current predicted power consumption and current predicted load rate of the artificial intelligence chip based on the current voltage and current frequency.

[0020] The second prediction module is used to input the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature of the temperature control module, and current test duration for the current test of the artificial intelligence chip into the second prediction model; the second prediction model is used to output the current predicted temperature of the artificial intelligence chip based on the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature, and current test duration.

[0021] The sending module is used to send a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module before executing the current test, based on the current predicted temperature, the test requirement temperature, and the current initial temperature; the test requirement temperature is the required temperature for the current test of the artificial intelligence chip.

[0022] Thirdly, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method.

[0023] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method.

[0024] Fifthly, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method.

[0025] The aforementioned temperature control method, apparatus, device, storage medium, and program product for testing artificial intelligence chips acquire the current voltage and frequency for the current test of the artificial intelligence chip, input the current voltage and frequency into a first prediction model, the first prediction model outputs the current predicted power consumption and current predicted load rate of the artificial intelligence chip based on the current voltage and frequency, input the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature of the temperature control module, and current test duration for the current test of the artificial intelligence chip into a second prediction model, the second prediction model outputs the current predicted temperature of the artificial intelligence chip based on the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature, and current test duration, and before executing the current test, sends a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module based on the current predicted temperature, the test requirement temperature for the current test of the artificial intelligence chip, and the current initial temperature. This solution can predict the temperature of the AI ​​chip in each test before each test based on the first and second prediction models. Combined with the required temperature for AI chip testing, it can accurately adjust the temperature of the temperature control module, thereby enabling the temperature control module to actively adapt to the temperature requirements of the AI ​​chip, effectively ensuring the accuracy of AI chip testing, improving AI chip testing efficiency, and reducing hardware risks. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 shows the application environment of a temperature control method for testing an artificial intelligence chip in one embodiment.

[0028] Figure 2 is a flowchart illustrating a temperature control method for testing an artificial intelligence chip in one embodiment;

[0029] Figure 3 is a flowchart illustrating the steps of sending a temperature adjustment command in one embodiment;

[0030] Figure 4 is a flowchart illustrating the steps of model training in one embodiment;

[0031] Figure 5 is a structural block diagram of a temperature control device for testing an artificial intelligence chip in one embodiment;

[0032] Figure 6 is an internal structure diagram of a computer device in one embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0034] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various objects, but these objects are not limited by these terms. These terms are only used to distinguish the first object from the second object. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the solutions, or any combination of multiple solutions.

[0035] Voltage-frequency parameterization testing of AI chips such as GPUs requires precise temperature control to ensure test accuracy. Current automated verification processes for AI chips typically employ a temperature control method where the ambient temperature is set only after the AI ​​chip's load has changed.

[0036] However, this method suffers from significant lag in temperature control. It requires waiting for the temperature sensor to detect that the AI ​​chip's temperature has exceeded the limit before passively adjusting the temperature of the temperature control module. The process of "AI chip temperature exceeding the limit → detection → temperature control module temperature adjustment → AI chip temperature drop" may have a delay of 5 to 10 seconds. During this period, the AI ​​chip may have already triggered thermal throttling, resulting in distortion of the voltage-frequency-power (VFP) test data. Furthermore, the longer the high temperature lasts, the higher the risk of hardware damage. Furthermore, this approach fails to consider the synergistic relationship between voltage, frequency, and power consumption, as well as load characteristics. It relies solely on the passive temperature regulation of the temperature control module, neglecting the dynamic correlation between voltage, frequency, power consumption, and AI chip temperature. It also fails to account for the load differences of the AI ​​chip at different stages, such as model training (high load fluctuations) and inference (low load, short response). For instance, at the same voltage and frequency, a high load during training may cause a sudden temperature rise in the AI ​​chip, while the temperature change is gradual under low load during inference. This approach cannot proactively predict and adjust the temperature control module's temperature; it can only passively adapt to the AI ​​chip's temperature changes, thus failing to fully utilize the precise temperature control potential of the module. Current temperature control methods used in automated AI chip verification processes easily lead to distorted test data, making it difficult to guarantee test accuracy.

[0037] In response, the temperature control method for testing artificial intelligence chips in this application can perform voltage-frequency-power consumption-load rate collaborative prediction based on a first prediction model and a second prediction model. Before each test, the temperature of the artificial intelligence chip in each test is predicted. Combined with the required temperature for artificial intelligence chip testing, the temperature of the temperature control module is precisely adjusted in advance. This enables the temperature control module to actively adapt to the temperature requirements of the artificial intelligence chip, solves the problem of prominent lag in current temperature control methods, effectively ensures the accuracy of artificial intelligence chip testing, improves the efficiency of artificial intelligence chip testing, and reduces hardware risks.

[0038] The temperature control method for testing artificial intelligence chips provided in this application embodiment can be applied to the application environment shown in Figure 1. This application environment may include an artificial intelligence chip, a temperature control module, and a computer device. The computer device can communicate with both the artificial intelligence chip and the temperature control module. The artificial intelligence chip can be an AI chip such as a GPU under test. The temperature control module can be used to control the temperature of the artificial intelligence chip, and can be a silicon thermal head or similar temperature control module. The computer device can be a terminal or a server. The terminal can be, but is not limited to, various personal computers, laptops, and tablets. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0039] In an exemplary embodiment, as shown in FIG2, a temperature control method for testing an artificial intelligence chip is provided. This method can be applied to the computer device shown in FIG1, and the method may include the following steps:

[0040] Step S201: Obtain the current voltage and current frequency for the current test of the artificial intelligence chip.

[0041] In this step, the AI ​​chip can undergo multiple tests (also known as multiple sets, multiple rounds, etc.), and the test to be performed at the moment can be called the current test. Before performing the current test of the AI ​​chip, the computer device can obtain the voltage and frequency used for the current test of the AI ​​chip, which can be recorded as the current voltage and current frequency.

[0042] Step S202: Input the current voltage and current frequency into the first prediction model.

[0043] In this step, the first prediction model is used to predict the power consumption and load rate of the AI ​​chip. These power consumption and load rate can be referred to as predicted power consumption and predicted load rate. The computer device can input the current voltage and frequency used for the current test of the AI ​​chip into the first prediction model. The first prediction model then outputs the predicted power consumption and predicted load rate of the AI ​​chip in the current test based on the current voltage and frequency. These predicted power consumption and predicted load rate can be denoted as current predicted power consumption and current predicted load rate. In other words, the first prediction model is used to output the current predicted power consumption and current predicted load rate of the AI ​​chip based on the current voltage and current frequency.

[0044] Step S203: Input the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature of the temperature control module, and current test duration for the current test of the artificial intelligence chip into the second prediction model.

[0045] In this step, the computer device can obtain the current initial temperature of the temperature control module, which can be the temperature before the current test of the AI ​​chip. The computer device can also obtain the test duration for the current test of the AI ​​chip, which can be recorded as the current test duration. In this step, the second prediction model is used to predict the temperature of the AI ​​chip, which can be called the predicted temperature. Therefore, the computer device inputs the current voltage, current frequency, current predicted power consumption, current predicted load rate, the current initial temperature of the temperature control module, and the current test duration for the current test of the AI ​​chip into the second prediction model. The second prediction model then outputs the predicted temperature of the AI ​​chip during the current test based on the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature, and current test duration. This predicted temperature can be recorded as the current predicted temperature. In other words, the second prediction model is used to output the current predicted temperature of the AI ​​chip based on the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature, and current test duration.

[0046] Step S204: Before executing the current test, send a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module based on the current predicted temperature, the required test temperature, and the current initial temperature.

[0047] In this step, the computer device can acquire the required test temperature, which is the temperature required for the current test of the AI ​​chip. In some embodiments, the required temperature of the AI ​​chip in different tests can be the same or different. In this step, before the AI ​​chip performs the current test, the computer device can send a temperature adjustment command to the temperature control module based on the current predicted temperature, the required test temperature, and the current initial temperature, so that the temperature control module can adjust the temperature before the AI ​​chip performs the current test. Then, the AI ​​chip can perform the current test. Thus, the computer device can instruct the temperature control module to adjust the temperature before each test of the AI ​​chip, and then the AI ​​chip can perform the test, thereby ensuring the accuracy of each test of the AI ​​chip.

[0048] The temperature control method for testing AI chips in this embodiment obtains the current voltage and frequency for the current test of the AI ​​chip. The current voltage and frequency are input into a first prediction model, which outputs the current predicted power consumption and current predicted load rate of the AI ​​chip based on these parameters. The current voltage, frequency, predicted power consumption, predicted load rate, the current initial temperature of the temperature control module, and the current test duration for the current test of the AI ​​chip are input into a second prediction model. The second prediction model outputs the current predicted temperature of the AI ​​chip based on these parameters. Before executing the current test, a temperature adjustment command is sent to the temperature control module based on the current predicted temperature, the required test temperature for the current test of the AI ​​chip, and the current initial temperature. This solution, based on the first and second prediction models, can predict the temperature of the AI ​​chip for each test before execution. Combined with the required test temperature, it accurately adjusts the temperature of the temperature control module, thereby enabling the temperature control module to proactively adapt to the temperature requirements of the AI ​​chip, effectively ensuring the accuracy of AI chip testing, improving testing efficiency, and reducing hardware risks.

[0049] In an exemplary embodiment, as shown in FIG3, step S204, which involves sending a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module based on the current predicted temperature, the required test temperature, and the current initial temperature, may include:

[0050] Step S301: Based on the current predicted temperature and the required test temperature, obtain the temperature adjustment amount for the current test of the artificial intelligence chip.

[0051] In this step, the computer equipment can determine the difference between the current predicted temperature and the required test temperature as the temperature adjustment amount for the current test of the artificial intelligence chip. This temperature adjustment amount refers to the adjustment of the current initial temperature of the temperature control module.

[0052] Step S302: Based on the temperature adjustment amount and the current initial temperature, send a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module.

[0053] In this step, the computer device can obtain the target temperature required by the temperature control module for the current test based on the sum of the current initial temperature and the temperature adjustment amount. This target temperature can be recorded as the current target temperature. Based on the current target temperature, the temperature control module sends a temperature adjustment command for the current test of the artificial intelligence chip. This temperature adjustment command can be used to instruct the temperature control module to adjust the current initial temperature to the current target temperature.

[0054] The solution in this embodiment can achieve precise and advance adjustment of the target temperature of the temperature control module to meet the current testing needs of the artificial intelligence chip and ensure the accuracy of the current test data.

[0055] In an exemplary embodiment, as shown in FIG4, the method of this application embodiment may further include the following steps:

[0056] Step S401: Obtain model training samples based on the sample dataset.

[0057] In this step, the computer equipment can acquire a sample dataset and obtain model training samples based on the sample dataset. The sample dataset may include historical test data of the AI ​​chip samples and historical temperature data of the corresponding temperature control module samples. The number of AI chip samples can be multiple, including samples from different production batches. The historical test data of the AI ​​chip samples refers to the data from historical testing of the AI ​​chip samples, which may include (real-time) power consumption (e.g., peak power consumption, average power consumption), load rate (e.g., core utilization, %), (real-time) temperature (chip core temperature), test duration, etc., of the AI ​​chip samples under different combinations of voltage (e.g., 0.8~1.2V, step size 0.05V) and frequency (e.g., 0.8~1.2GHz, step size 0.1GHz) combinations. The corresponding temperature control module samples refer to the temperature control module samples corresponding to the AI ​​chip samples. The historical temperature data of the temperature control module samples are the temperature data of the temperature control module samples in the historical testing of the AI ​​chip samples, such as initial temperature, target temperature (e.g., 40℃~65℃), etc.

[0058] The model training samples refer to the samples used to train the first prediction model and the second prediction model. These samples can include voltage, frequency, power consumption, load rate, test duration, and temperature samples of the AI ​​chip samples, as well as the initial temperature samples of the corresponding temperature control module samples. Specifically, the voltage, frequency, power consumption, load rate, test duration, and initial temperature samples of the corresponding temperature control module samples can be selected from the aforementioned AI chip samples under different combinations of voltage and frequency, focusing on power consumption, load rate, temperature, test duration, and the initial temperature of the corresponding temperature control module samples. In detail, the computer equipment can remove invalid data caused by test interruptions or equipment malfunctions from the sample dataset. Through correlation analysis, voltage and frequency are identified as primary driving features, power consumption and load rate as secondary intermediate features, and test duration and the initial temperature of the temperature control module as environmental features. Ultimately, voltage, frequency, power consumption, load rate, test duration, and the initial temperature of the temperature control module can be determined as model input features, thus forming the model training samples.

[0059] Step S402: Train the first prediction model based on the voltage sample, frequency sample, power consumption sample, and load rate sample of the artificial intelligence chip sample.

[0060] Step S403: Train the second prediction model based on the voltage sample, frequency sample, predicted power consumption sample and predicted load rate sample output by the first prediction model, as well as the initial temperature sample and test duration sample of the artificial intelligence chip sample.

[0061] Steps S402 and S403 involve training the first prediction model and the second prediction model to be trained based on the corresponding model input features, resulting in trained first prediction models and second prediction models, which are used for predicting power consumption and load rate, and predicting the temperature of the artificial intelligence chip, respectively.

[0062] In some exemplary embodiments, step S402, training the first prediction model based on the voltage, frequency, power consumption, and load rate samples of the AI ​​chip sample, may include: inputting the voltage and frequency samples of the AI ​​chip sample into the first prediction model to obtain the predicted power consumption and predicted load rate samples output by the first prediction model; and training the first prediction model based on the errors between the predicted power consumption samples and the predicted load rate samples. And / or, step S403, training the second prediction model based on the voltage and frequency samples of the AI ​​chip sample, the predicted power consumption and predicted load rate samples output by the first prediction model, the initial temperature sample, and the test duration sample, may include: inputting the voltage and frequency samples of the AI ​​chip sample, the predicted power consumption and predicted load rate samples output by the first prediction model, the initial temperature sample, and the test duration sample into the second prediction model to obtain the predicted temperature sample output by the second prediction model; and training the second prediction model based on the errors between the predicted temperature samples and the temperature samples.

[0063] For training the first prediction model, the computer device can input voltage and frequency samples from the AI ​​chip sample into the first prediction model to obtain predicted power consumption samples and predicted load rate samples output by the first prediction model. Then, the first prediction model can be trained based on the errors between the predicted power consumption samples and the predicted load rate samples. A gradient boosting tree algorithm can be used for training to ensure that the errors between the predicted power consumption samples and the predicted load rate samples are less than or equal to a power consumption error threshold, and to ensure that the errors between the predicted load rate samples and the predicted load rate samples are less than or equal to a load rate error threshold, thus obtaining the trained first prediction model. As an example, the error between the predicted power consumption samples and the predicted power consumption samples can be represented as "|predicted power consumption sample - power consumption sample| / power consumption sample × 100%" (the percentage corresponding to the ratio of the absolute value of the difference between the predicted power consumption sample and the predicted power consumption sample to the power consumption sample), and the power consumption error threshold can be set to 3%. As an example, the error between the predicted load rate sample and the load rate sample can be represented by "|predicted load rate sample - load rate sample| / load rate sample × 100%" (the percentage corresponding to the ratio of the absolute value of the difference between the predicted load rate sample and the load rate sample to the load rate sample), and the load rate error threshold can be 5%.

[0064] For training the second prediction model, the computer device can acquire the predicted power consumption samples and predicted load rate samples output by the first prediction model. It then inputs the voltage and frequency samples of the AI ​​chip sample, the predicted power consumption and predicted load rate samples output by the first prediction model, as well as the initial temperature sample and test duration sample, into the second prediction model to obtain the predicted temperature sample output by the second prediction model. The second prediction model can then be trained based on the error between the predicted temperature sample and the actual temperature sample. Specifically, the second prediction model can be trained by capturing the temporal correlation between voltage-frequency-power consumption-load rate and the temperature of the AI ​​chip, ensuring that the error between the predicted temperature sample and the actual temperature sample is less than or equal to a temperature error threshold. This results in the trained second prediction model. For example, the error between the predicted temperature sample and the actual temperature sample can be represented by "|predicted temperature sample - actual temperature sample|" (the absolute value of the difference between the predicted temperature sample and the actual temperature sample), and the temperature error threshold can be set to 2℃.

[0065] Therefore, by integrating the first and second prediction models, a collaborative prediction model of voltage-frequency-power consumption-load rate-temperature can be formed, which can achieve end-to-end prediction of "voltage-frequency input → predicted power consumption, predicted load rate → predicted temperature", providing an accurate basis for adjusting the target temperature of the temperature control module in advance and solving the temperature lag problem of passive adaptation in traditional solutions.

[0066] In an exemplary embodiment, further, after sending a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module based on the current predicted temperature, the required test temperature, and the current initial temperature in step S204, the method may further include:

[0067] Record the test data of the current test; if the number of tests meets the target condition, add the test data recorded each time to the sample dataset, and retrain the first prediction model and the second prediction model based on the sample dataset.

[0068] In this embodiment, during the current test of the AI ​​chip, the computer device can collect real-time data such as the AI ​​chip's actual power consumption, actual temperature, actual load rate, and the actual target temperature of the temperature control module (which may also include dynamically changing temperatures during the temperature adjustment process) to form the current test data. This test data can be recorded in a MySQL database. The AI ​​chip can undergo multiple tests, and the test data for each test can be recorded. If the number of tests meets a target condition, for example, reaching a threshold (e.g., 50), the recorded test data from each test can be added to a sample dataset, and the first and second prediction models can be retrained based on this sample dataset. Therefore, when the AI ​​chip completes, for example, 50 sets of tests, the recorded test data (voltage, frequency, actual power consumption, actual temperature, actual load rate, test duration, initial temperature, target temperature, etc.) can be added to the sample dataset. This allows the model training samples to be obtained based on the supplemented sample dataset. The first and second prediction models can then be retrained based on the model training samples. This ensures that the error between the AI ​​chip's predicted temperature and the actual temperature is less than or equal to the temperature error threshold (which can be 2°C). At the same time, it can optimize the matching relationship between the AI ​​chip's predicted temperature and the target temperature of the temperature control module, further improving the temperature control accuracy.

[0069] In one exemplary embodiment, the method of this application may further include the following steps:

[0070] The system collects data on the voltage, frequency, power consumption, load rate, and temperature of the AI ​​chip, as well as the temperature of the temperature control module, at the sampling time points. Based on these data, a load curve is generated, and load warning information for the AI ​​chip is output.

[0071] In this embodiment, the computer device can collect data on the voltage, frequency, power consumption, load rate, and temperature of the AI ​​chip and the temperature control module during testing, according to sampling time points. Based on the sampling time points and the data of the AI ​​chip and the temperature control module, a load curve can be generated with time as the horizontal axis and six dimensions (voltage, frequency, power consumption, load rate, temperature of the AI ​​chip, and temperature of the temperature control module) as the vertical axis. The device can also output load warning information for the AI ​​chip, such as displaying load warning information on the load curve. This load warning information can include load peak warnings and high load sample data. For example, a load peak warning could predict that the load rate will rise to 90% in 5 minutes. High load sample data can include data such as the voltage, frequency, power consumption, load rate, temperature of the AI ​​chip and the temperature control module when the load rate is greater than or equal to 60%. Therefore, the load curve and load warning information can provide a basis for optimizing the target temperature benchmark of the temperature control module.

[0072] In one exemplary embodiment, a temperature control method for testing an artificial intelligence chip is also provided, which can be executed by a computer device. The method may include a model building phase and a testing and iterative optimization phase.

[0073] In the model building phase, a model for the coordinated prediction of voltage, frequency, power consumption, load, and temperature of an artificial intelligence chip can be constructed.

[0074] First, multi-dimensional historical test data can be collected, covering, for example, three or more production batches of AI chips. This includes recording (real-time) power consumption P (e.g., peak power consumption, average power consumption), load rate L (e.g., core utilization, %), (real-time) temperature T (chip core temperature), test duration t, etc., of different combinations of voltage V (e.g., 0.8~1.2V, step size 0.05V) and frequency F (e.g., 0.8~1.2GHz, step size 0.1GHz), as well as the corresponding historical temperature data of the temperature control module, such as initial temperature T0, target temperature H, etc. This forms a sample dataset, which can contain multiple sets of V, F, P, T, L, t, and H data. Each set of V, F, P, T, L, t, and H data corresponds to one historical test.

[0075] Then, data preprocessing and feature engineering can be performed. For the sample dataset, invalid data caused by test interruption and equipment failure are removed. Through correlation analysis, voltage V and frequency F are obtained as primary driving features, power consumption P and load rate L are as secondary intermediate features, and test duration t and initial temperature T0 of temperature control module are as environmental features. Finally, voltage V, frequency F, power consumption P, load rate L, test duration t, and initial temperature T0 of temperature control module can be determined as model input features, thus forming model training samples.

[0076] Next, model training and optimization can be performed. A two-stage prediction model can be used, where the first-stage prediction model can be the first prediction model and the second-stage prediction model can be the second prediction model.

[0077] For the first prediction model, with voltage V and frequency F as inputs, the outputs are predicted power consumption P_pred and predicted load rate L_pred. The gradient boosting tree algorithm can be used for training to ensure that the power consumption prediction error is less than or equal to 3% (|predicted load rate P_pred - actual power consumption P| / actual power consumption P×100%≤3%), and to ensure that the load rate prediction error is less than or equal to 5% (|predicted load rate L_pred - actual load rate L| / actual load rate L×100%≤5%).

[0078] For the second prediction model, the inputs are voltage V, frequency F, predicted power consumption P_pred, predicted load rate L_pred, initial temperature T0 and test duration t. The output is the predicted temperature T_pred of the artificial intelligence chip. The model can be trained by capturing the temporal correlation between VFPL and the temperature of the artificial intelligence chip, so that the temperature prediction error is less than or equal to 2℃ (|predicted temperature T_pred - actual temperature T|≤2℃).

[0079] The first and second prediction models can be integrated to form a VFPLT load-coordinated prediction model, which can achieve end-to-end prediction of "voltage V-frequency F input → predicted power consumption P_pred, predicted load rate L_pred → predicted temperature T_pred". This provides an accurate basis for adjusting the target temperature of the temperature control module (such as a silicon-based temperature controller) in advance, and solves the temperature lag problem of passive adaptation in traditional solutions.

[0080] During the testing and iterative optimization phase, automated testing of AI chips and iterative optimization of prediction models can be performed.

[0081] First, S1 test initialization can be performed. During S1 test initialization, the automated test system is started, and the VF parameter sequence to be tested is loaded. Tools can be used to collect data such as the initial power consumption P0 and initial load rate L0 of the current AI chip. Simultaneously, the initial temperature T0 of the temperature control module can be initialized (e.g., 60℃) to complete heat exchange preheating. Then, S2 collaborative prediction can be performed. In S2 collaborative prediction, the current VF parameters can be input into the first prediction model to obtain the predicted power consumption P_pred and predicted load rate L_pred. Then, the voltage V, frequency F, predicted power consumption P_pred, predicted load rate L_pred, initial temperature T0, and test duration t are input into the second prediction model to obtain the predicted temperature T_pred. Next, S3 feedforward temperature control execution can be performed. In S3 feedforward temperature control execution, the temperature adjustment amount Delta can be obtained based on the difference between the predicted temperature T_pred and the required test temperature. Based on this, a temperature adjustment command is sent to the temperature control module. This temperature adjustment command can be used to indicate the target temperature (T0 + Delta) of the temperature control module. Next, S4 testing and data recording can be performed. During S4 testing and data recording, the current VF parameter test can be executed, and real-time data such as actual power consumption P_real, actual temperature T_real of the AI ​​chip, actual load rate L_real, and actual target temperature of the temperature control module (which may also include dynamically changing temperatures during adjustment) can be collected and recorded to a MySQL database. Then, S5 model iteration can be performed. In S5 model iteration, after every 50 sets of VF parameter tests, data such as voltage V, frequency F, actual power consumption P_real, actual load rate L_real, actual temperature T_real, initial temperature T0, test time t, and target temperature H of the temperature control module can be added to the sample dataset. The first and second prediction models are then retrained using the supplemented sample dataset to ensure that the temperature prediction error for the AI ​​chip is always less than or equal to, for example, 2℃. Simultaneously, the matching relationship between the predicted temperature of the AI ​​chip and the target temperature of the temperature control module can be optimized to further improve temperature control accuracy. The above steps S1 to S5 can be executed repeatedly until all VF parameter tests are completed. A standardized report containing data such as "VF parameters, P_pred and P_real, L_pred and L_real, T_pred and T_real, target temperature adjustment record of the temperature control module, and perf / watt (Performance per Watt) value" can then be output. Furthermore, multi-dimensional data acquisition can obtain data such as L_real and the temperature of the VFPT-temperature control module, generating a load curve with time on the horizontal axis and six dimensions (voltage, frequency, power consumption, load rate, temperature of the AI ​​chip, and temperature of the temperature control module) on the vertical axis. This curve outputs load warning information for the AI ​​chip, which may include load peak warnings and high load sample data.The load peak warning could be, for example, a prediction that the load rate will rise to 90% in 5 minutes. High load sample data could include data such as the voltage, frequency, power consumption, load rate, and temperature of AI chips with a load rate greater than or equal to 60%, as well as the temperature of the temperature control module. Therefore, the load curve and load warning information can provide a basis for optimizing the target temperature benchmark of the temperature control module.

[0082] The temperature control method for testing artificial intelligence chips according to the embodiments of this application can realize automated test temperature feedforward control of the voltage-frequency-power consumption-load rate-temperature of artificial intelligence chips based on model prediction. By constructing a collaborative prediction model of voltage-frequency-power consumption-load rate-temperature, the temperature changes of artificial intelligence chips at different load stages can be predicted in advance, and the target temperature of the temperature control module can be accurately adjusted. This enables the temperature control module to actively adapt to the temperature requirements of the artificial intelligence chip, solves the problem of temperature lag adjustment in traditional technologies, ensures the accuracy of test data, improves test efficiency and reduces hardware risks. It can be applied to VF parameterized test scenarios of artificial intelligence chips such as GPUs that require precise control of temperature interference, and can use temperature control modules such as silicon-based temperature control heads to adjust their own temperature to achieve efficient temperature control.

[0083] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0084] Based on the same inventive concept, this application also provides a temperature control device for implementing the temperature control method for testing artificial intelligence chips as described above. The solution provided by this device is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the temperature control device for testing artificial intelligence chips provided below can be found in the limitations of the temperature control method for testing artificial intelligence chips described above, and will not be repeated here.

[0085] In an exemplary embodiment, as shown in FIG5, a temperature control device for testing artificial intelligence chips is provided. The temperature control device 500 for testing artificial intelligence chips may include:

[0086] The acquisition module 501 is used to acquire the current voltage and current frequency for the current test of the artificial intelligence chip;

[0087] The first prediction module 502 is used to input the current voltage and current frequency into the first prediction model; the first prediction model is used to output the current predicted power consumption and current predicted load rate of the artificial intelligence chip based on the current voltage and current frequency.

[0088] The second prediction module 503 is used to input the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature of the temperature control module, and current test duration for the current test of the artificial intelligence chip into the second prediction model; the second prediction model is used to output the current predicted temperature of the artificial intelligence chip based on the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature, and current test duration.

[0089] The sending module 504 is used to send a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module before performing the current test, based on the current predicted temperature, the test requirement temperature and the current initial temperature; the test requirement temperature is the required temperature for the current test of the artificial intelligence chip.

[0090] In an exemplary embodiment, the sending module 504 is configured to obtain a temperature adjustment amount for the current test of the artificial intelligence chip based on the current predicted temperature and the test requirement temperature; and send a temperature adjustment instruction for the current test of the artificial intelligence chip to the temperature control module based on the temperature adjustment amount and the current initial temperature.

[0091] In an exemplary embodiment, the temperature control device 500 for testing artificial intelligence chips may further include: a training module, configured to obtain model training samples based on a sample dataset; the sample dataset includes historical test data of artificial intelligence chip samples and historical temperature data of corresponding temperature control module samples; the model training samples include voltage samples, frequency samples, power consumption samples, load rate samples, test duration samples, and temperature samples of the artificial intelligence chip samples, as well as initial temperature samples of the corresponding temperature control module samples; the first prediction model is trained based on the voltage samples, frequency samples, power consumption samples, and load rate samples of the artificial intelligence chip samples; the second prediction model is trained based on the voltage samples, frequency samples, predicted power consumption samples and predicted load rate samples output by the first prediction model, and the initial temperature samples and test duration samples.

[0092] In an exemplary embodiment, the training module is configured to input voltage and frequency samples of the AI ​​chip sample into the first prediction model to obtain predicted power consumption samples and predicted load rate samples output by the first prediction model; and train the first prediction model based on the error between the predicted power consumption sample and the power consumption sample, and the error between the predicted load rate sample and the load rate sample; and / or, the training module is configured to input voltage and frequency samples of the AI ​​chip sample, predicted power consumption samples and predicted load rate samples output by the first prediction model, as well as the initial temperature sample and test duration sample, into the second prediction model to obtain predicted temperature samples output by the second prediction model; and train the second prediction model based on the error between the predicted temperature sample and the temperature sample.

[0093] In an exemplary embodiment, the training module is further configured to record the test data of the current test; if the number of tests meets the target condition, the test data recorded each time is added to the sample dataset, and the first prediction model and the second prediction model are retrained based on the sample dataset.

[0094] In an exemplary embodiment, the temperature control device 500 for testing the artificial intelligence chip may further include: an information processing module, configured to collect the voltage, frequency, power consumption, load rate, and temperature of the artificial intelligence chip and the temperature of the temperature control module at sampling time points; generate a load curve based on the sampling time points and the voltage, frequency, power consumption, load rate, and temperature of the artificial intelligence chip and the temperature of the temperature control module; and output load warning information for the artificial intelligence chip.

[0095] The modules in the temperature control device for the aforementioned artificial intelligence chip test can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0096] In an exemplary embodiment, a computer device is provided, which can be a terminal or a server, and its internal structure diagram is shown in Figure 6. The computer device includes a processor, memory, input / output interfaces, and a communication interface. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface is connected to the system bus via the input / output interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for wired or wireless communication with external devices; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a temperature control method for testing an artificial intelligence chip.

[0097] Those skilled in the art will understand that the structure shown in Figure 6 is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0098] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0099] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0100] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0101] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0102] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0103] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0104] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A temperature control method for testing artificial intelligence chips, characterized in that, The method includes: acquiring the current voltage and current frequency for the current test of the artificial intelligence chip; inputting the current voltage and current frequency into a first prediction model; the first prediction model is used to output the current predicted power consumption and current predicted load rate of the artificial intelligence chip based on the current voltage and current frequency; inputting the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature of the temperature control module, and current test duration for the current test of the artificial intelligence chip into a second prediction model; the second prediction model is used to output the current predicted temperature of the artificial intelligence chip based on the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature, and current test duration; before executing the current test, sending a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module based on the current predicted temperature, the test requirement temperature, and the current initial temperature; the test requirement temperature is the required temperature for the current test of the artificial intelligence chip.

2. The method according to claim 1, characterized in that, The step of sending a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module based on the current predicted temperature, the required test temperature, and the current initial temperature includes: obtaining a temperature adjustment amount for the current test of the artificial intelligence chip based on the current predicted temperature and the required test temperature; and sending the temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module based on the temperature adjustment amount and the current initial temperature.

3. The method according to claim 1, characterized in that, The method further includes: obtaining model training samples based on a sample dataset; the sample dataset includes historical test data of artificial intelligence chip samples and historical temperature data of corresponding temperature control module samples; the model training samples include voltage samples, frequency samples, power consumption samples, load rate samples, test duration samples, and temperature samples of the artificial intelligence chip samples, as well as initial temperature samples of the corresponding temperature control module samples; training the first prediction model based on the voltage samples, frequency samples, power consumption samples, and load rate samples of the artificial intelligence chip samples; and training the second prediction model based on the voltage samples, frequency samples, predicted power consumption samples and predicted load rate samples output by the first prediction model, as well as the initial temperature samples and test duration samples.

4. The method according to claim 3, characterized in that, The step of training the first prediction model based on the voltage, frequency, power consumption, and load rate samples of the AI ​​chip sample includes: inputting the voltage and frequency samples of the AI ​​chip sample into the first prediction model to obtain the predicted power consumption and predicted load rate samples output by the first prediction model; training the first prediction model based on the error between the predicted power consumption sample and the power consumption sample, and the error between the predicted load rate sample and the load rate sample; and / or, the step of training the second prediction model based on the voltage and frequency samples of the AI ​​chip sample, the predicted power consumption and predicted load rate samples output by the first prediction model, and the initial temperature and test duration samples includes: inputting the voltage and frequency samples of the AI ​​chip sample, the predicted power consumption and predicted load rate samples output by the first prediction model, and the initial temperature and test duration samples into the second prediction model to obtain the predicted temperature sample output by the second prediction model; training the second prediction model based on the error between the predicted temperature sample and the temperature sample.

5. The method according to claim 3, characterized in that, After sending a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module based on the current predicted temperature, the test requirement temperature, and the current initial temperature, the method further includes: recording the test data of the current test; if the number of tests meets the target condition, adding the test data recorded each time to the sample dataset, and retraining the first prediction model and the second prediction model based on the sample dataset.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: collecting the voltage, frequency, power consumption, load rate, temperature of the artificial intelligence chip and the temperature of the temperature control module according to the sampling time point; generating a load curve based on the sampling time point and the voltage, frequency, power consumption, load rate, temperature of the artificial intelligence chip and the temperature of the temperature control module; and outputting load warning information of the artificial intelligence chip.

7. A temperature control device for testing artificial intelligence chips, characterized in that, The device includes: an acquisition module for acquiring the current voltage and current frequency for the current test of the artificial intelligence chip; a first prediction module for inputting the current voltage and current frequency into a first prediction model; the first prediction model for outputting the current predicted power consumption and current predicted load rate of the artificial intelligence chip based on the current voltage and current frequency; a second prediction module for inputting the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature of the temperature control module, and current test duration for the current test of the artificial intelligence chip into the second prediction model; the second prediction model for outputting the current predicted temperature of the artificial intelligence chip based on the current voltage, current frequency, current predicted power consumption, current predicted load rate, current initial temperature, and current test duration; and a sending module for sending a temperature adjustment command for the current test of the artificial intelligence chip to the temperature control module before executing the current test, based on the current predicted temperature, the required test temperature, and the current initial temperature; the required test temperature is the temperature required for the current test of the artificial intelligence chip.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images