GPU power consumption control method, system, computer device and storage medium

Abstract

The present invention provides a GPU power consumption control method, system, computer device, and storage medium, wherein the method includes obtaining historical load information of a GPU within a first target time period; training a machine learning model based on the historical load information of the GPU within the first target time period, and using the trained machine learning model as a load prediction model; predicting the load information of the GPU within a future second time period based on the load prediction model; adjusting the operating voltage and operating frequency of the GPU within the future second time period based on the load information of the GPU within the future second time period; using the load information corresponding to the current operating voltage and operating frequency of the GPU as historical load information, and re-obtaining the historical load information of the GPU within the first target time period. The present invention achieves a balance between GPU performance and energy consumption, can regularly collect new data, and continuously update and optimize the prediction model to adapt to load changes in different application scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of electronic digital data processing, and in particular relates to a GPU power consumption control method, system, computer equipment and storage medium. Background Art

[0002] With the rapid development of computationally demanding applications such as deep learning, real-time rendering, and scientific computing, GPU performance requirements continue to increase, leading to a sharp increase in power consumption. Traditional static power management methods struggle to achieve an optimal balance between performance and energy consumption, especially in scenarios where multitasking workloads change dynamically.

[0003] Existing technologies primarily use a preset fixed policy to adjust GPU voltage and frequency (DVFS). However, this approach cannot adapt to complex changes in computing load in real time, easily leading to resource waste or performance degradation. Therefore, a management solution that can dynamically adjust GPU power consumption based on real-time workloads is urgently needed. Summary of the Invention

[0004] The present invention provides a GPU power consumption control method, system, computer equipment and storage medium to dynamically adjust GPU power consumption according to real-time workload.

[0005] In a first aspect, the present invention provides a GPU power consumption control method, comprising:

[0006] S1, obtaining historical load information of the GPU within a first target time period; wherein the historical load information includes historical core usage;

[0007] S2, training a machine learning model based on historical GPU load information during the first target time period, and using the trained machine learning model as a load prediction model;

[0008] S3, predicting GPU load information in a second future time period according to the load prediction model;

[0009] S4, adjusting the operating voltage and operating frequency of the GPU in the second time period in the future according to the load information of the GPU in the second time period in the future;

[0010] S5: Use the load information corresponding to the current operating voltage and operating frequency of the GPU as historical load information, and return to execute the operation of S1.

[0011] Optionally, the step of training a machine learning model according to historical load information of the GPU within the first target time period and using the trained machine learning model as a load prediction model includes:

[0012] Construct the first loss function L1:

[0013]

[0014] Where M is the total number of historical load information; The value of the mth load information predicted by the machine learning model; n m is the actual value of the mth load information;

[0015] And / or, construct a second loss function L2:

[0016]

[0017] Where M' is the total number of categories of load information; h m' is the true label of the m'th category load information; The probability that the load information predicted by the machine learning model belongs to the m'th type of load information;

[0018] Determine the total loss of the machine learning model according to the first loss function L1 and / or the second loss function L2;

[0019] The machine learning model is trained according to the total loss of the machine learning model, and the trained machine learning model is used as a load prediction model.

[0020] Optionally, determining the total loss of the machine learning model according to the first loss function L1 and / or the second loss function L2 includes:

[0021] The total loss L of the machine learning model is calculated according to the following formula f :

[0022] L f =αL1+βL2;

[0023] Among them, α is the preset weight of the first loss function; β is the preset weight of the second loss function.

[0024] Optionally, adjusting the operating voltage and operating frequency of the GPU in the future second time period according to the load information of the GPU in the future second time period includes:

[0025] Obtain the load forecast model's predicted value of core utilization at a future target time;

[0026] When the predicted value of the core usage rate is less than the first target ratio, the first preset voltage and the first preset frequency are used as the operating voltage and the operating frequency of the GPU;

[0027] When the predicted value of the core usage rate is greater than or equal to the first target ratio and less than or equal to the second target ratio, the second preset voltage and the second preset frequency are used as the operating voltage and operating frequency of the GPU;

[0028] When the predicted value of the core usage rate is greater than the second target ratio, the third preset voltage and the second preset frequency are used as the operating voltage and the operating frequency of the GPU.

[0029] In a second aspect, the present invention provides a GPU power consumption control system, comprising:

[0030] An acquisition module, configured to acquire historical load information of the GPU within a first target time period; wherein the historical load information includes historical core usage;

[0031] A training module, configured to train a machine learning model based on historical GPU load information within a first target time period, and use the trained machine learning model as a load prediction model;

[0032] A prediction module, configured to predict GPU load information in a second future time period according to a load prediction model;

[0033] An adjustment module, configured to adjust an operating voltage and an operating frequency of the GPU in the second time period in the future according to load information of the GPU in the second time period in the future;

[0034] The determination module is used to use the load information corresponding to the current operating voltage and operating frequency of the GPU as the historical load information, and return to execute the operation of the acquisition module.

[0035] Optionally, the training module includes:

[0036] The first construction unit is used to construct the first loss function L1:

[0037]

[0038] Where M is the total number of historical load information; The value of the mth load information predicted by the machine learning model; n m is the actual value of the mth load information;

[0039] And / or, a second construction unit, for constructing a second loss function L2:

[0040]

[0041] Where M' is the total number of categories of load information; h m' is the true label of the m'th category load information; The probability that the load information predicted by the machine learning model belongs to the m'th type of load information;

[0042] A first determining unit, configured to determine a total loss of the machine learning model according to the first loss function L1 and / or the second loss function L2;

[0043] The training unit is used to train the machine learning model according to the total loss of the machine learning model, and use the trained machine learning model as a load prediction model.

[0044] Optionally, the first determining unit includes:

[0045] A computing device for calculating the total loss L of the machine learning model according to the following formula f :

[0046] L f =αL1+βL2;

[0047] Among them, α is the preset weight of the first loss function; β is the preset weight of the second loss function.

[0048] Optionally, the adjustment module includes:

[0049] An acquisition unit, used to obtain a predicted value of the core usage rate at a future target moment by a load prediction model;

[0050] a second determining unit, configured to use a first preset voltage and a first preset frequency as the operating voltage and operating frequency of the GPU when the predicted value of the core usage rate is less than the first target ratio;

[0051] a third determining unit, configured to use the second preset voltage and the second preset frequency as the operating voltage and operating frequency of the GPU when the predicted value of the core usage rate is greater than or equal to the first target ratio and less than or equal to the second target ratio;

[0052] The fourth determining unit is configured to use the third preset voltage and the second preset frequency as the operating voltage and operating frequency of the GPU when the predicted value of the core usage rate is greater than the second target ratio.

[0053] In a third aspect, the present invention provides a computer device comprising a processor and a memory; wherein, when the processor executes a computer program stored in the memory, the steps of the GPU power consumption control method described in the first aspect are implemented.

[0054] In a fourth aspect, the present invention provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, the steps of the GPU power consumption control method described in the first aspect are implemented.

[0055] The present invention provides a GPU power consumption control method, system, computer device and storage medium. The method trains a prediction model based on historical data to accurately predict load changes, and combines it with a dynamic power consumption adjustment strategy to achieve a balance between GPU performance and energy consumption. The self-learning mechanism can regularly collect new data and continuously update and optimize the prediction model to adapt to load changes in different application scenarios, providing strong support for the application of GPUs in multiple fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0057] Figure 1 A flowchart of a GPU power consumption control method provided by an embodiment of the present invention;

[0058] Figure 2 A schematic diagram of the structure of a GPU power consumption control system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0059] The following will provide a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0060] Example 1

[0061] like Figure 1 As shown, this embodiment provides a GPU power consumption control method, including:

[0062] S1. Obtain historical load information of a GPU within a first target time period; wherein the historical load information includes historical core usage.

[0063] Historical data is the basis for training prediction models. It is mainly generated through GPU runtime logs and data recorded by the monitoring module, including but not limited to the following:

[0064] Task queue length: This records the number of pending tasks and their processing time, reflecting the current workload. This can be obtained through registers in the GPU hardware scheduler. For example, GPU hardware APIs (such as CUDA or ROCm) provide interfaces that allow querying the task queue status.

[0065] Core Utilization: The utilization rate of GPU cores, which is used to measure the intensity of computing resource usage.

[0066] Video memory usage: allocation and usage of video memory, reflecting video memory pressure.

[0067] Temperature curve: includes the time series changes of GPU core temperature and memory temperature, used to evaluate the impact of thermal load on performance.

[0068] Current and voltage: record the actual status of the power supply system and analyze power consumption trends.

[0069] Fan speed: As auxiliary information, it reflects the operating status of the cooling system.

[0070] Frame rate (FPS): Records frame rate changes in real-time rendering tasks to measure the real-time requirements of the tasks.

[0071] Task category label: records the specific type of task (such as deep learning reasoning, image rendering, scientific computing, etc.) to facilitate classification and processing.

[0072] This data is collected through hardware sensors and the driver layer and periodically saved in a historical database as training data.

[0073] S2: Train a machine learning model based on historical load information of the GPU in the first target time period, and use the trained machine learning model as a load prediction model.

[0074] In this step, illustratively, the first loss function L1 is constructed:

[0075]

[0076] Where M is the total number of historical load information; The value of the mth load information predicted by the machine learning model; n m is the actual value of the mth load information.

[0077] And / or, construct a second loss function L2:

[0078]

[0079] Where M' is the total number of categories of load information; h m' is the true label of the m'th category load information; The probability that the load information belongs to the m'th category of load information is predicted by the machine learning model.

[0080] The total loss of the machine learning model is determined based on the first loss function L1 and / or the second loss function L2.

[0081] The machine learning model is trained according to the total loss of the machine learning model, and the trained machine learning model is used as a load prediction model.

[0082] Among them, the total loss L of the machine learning model is calculated according to the following formula f :

[0083] L f =αL1+βL2.

[0084] Among them, α is the preset weight of the first loss function; β is the preset weight of the second loss function.

[0085] Adjust the loss function weight according to the scenario to continuously optimize the load forecasting model and better adapt to different scenarios.

[0086] S3, predicting GPU load information in a second future time period according to the load prediction model.

[0087] S4: Adjust the operating voltage and operating frequency of the GPU in the second time period in the future according to the load information of the GPU in the second time period in the future.

[0088] Exemplarily, this step includes:

[0089] Get the load forecast model's predicted value of core utilization at a target time in the future.

[0090] When the predicted value of the core usage is less than the first target ratio (eg, 30%), the first preset voltage and the first preset frequency are used as the operating voltage and the operating frequency of the GPU (ie, the operating frequency of the GPU core clock).

[0091] When the predicted value of the core usage is greater than or equal to the first target ratio and less than or equal to the second target ratio (eg, 70%), the second preset voltage and the second preset frequency are used as the operating voltage and operating frequency of the GPU.

[0092] When the predicted value of the core usage rate is greater than the second target ratio, the third preset voltage and the second preset frequency are used as the operating voltage and the operating frequency of the GPU.

[0093] Adjust the GPU voltage and frequency according to different load conditions. During the adjustment process, ensure that voltage changes take precedence over frequency changes to prevent insufficient power supply due to excessive frequency. In low-load scenarios (predicted core utilization is less than 30%), reduce voltage and frequency to avoid resource waste and effectively save energy. Under high load (predicted core utilization is greater than 70%), quickly increase voltage and frequency to meet the real-time requirements of tasks while accurately matching computing resources with load, significantly improving energy utilization efficiency and greatly improving energy efficiency compared to traditional fixed strategies.

[0094] S5: Use the load information corresponding to the current operating voltage and operating frequency of the GPU as historical load information, and return to execute the operation of S1.

[0095] During operation, the self-learning mechanism regularly collects the latest operational data, expands the training dataset, and uses incremental learning to update the parameters of the load prediction model online. Simultaneously, it dynamically adjusts the weights of the loss function based on the application scenario, such as increasing the error penalty for real-time tasks. This enables the load prediction model to continuously optimize, better adapt to the dynamic changes in GPU load in different application scenarios, and maintain good control performance over the long term.

[0096] In summary, this embodiment provides a GPU power consumption control method that trains a prediction model based on historical data to accurately predict load changes, and combines it with a dynamic power consumption adjustment strategy to achieve a balance between GPU performance and energy consumption; the self-learning mechanism can regularly collect new data and continuously update and optimize the prediction model to adapt to load changes in different application scenarios, providing strong support for the application of GPUs in multiple fields.

[0097] Based on the same inventive concept as Example 1, this embodiment provides a GPU power consumption control system. Since the principle of solving the problem by this system is similar to that of the aforementioned GPU power consumption control method, the implementation of this system can refer to the implementation of the GPU power consumption control method.

[0098] like Figure 2 As shown, the GPU power consumption control system includes:

[0099] The acquisition module 10 is configured to acquire historical load information of the GPU within a first target time period; wherein the historical load information includes historical core usage.

[0100] The training module 20 is used to train a machine learning model based on the historical load information of the GPU in the first target time period, and use the trained machine learning model as a load prediction model.

[0101] The prediction module 30 is configured to predict the load information of the GPU in a second future time period according to the load prediction model.

[0102] The adjustment module 40 is configured to adjust the operating voltage and the operating frequency of the GPU in the second future time period according to the load information of the GPU in the second future time period.

[0103] The determination module 50 is configured to use the load information corresponding to the current operating voltage and operating frequency of the GPU as the historical load information, and return to execute the operation of the acquisition module.

[0104] Exemplarily, the training module includes:

[0105] The first construction unit is used to construct the first loss function L1:

[0106]

[0107] Where M is the total number of historical load information; The value of the mth load information predicted by the machine learning model; n m is the actual value of the mth load information;

[0108] And / or, a second construction unit, for constructing a second loss function L2:

[0109]

[0110] Where M' is the total number of categories of load information; h m' is the true label of the m'th category load information; The probability that the load information predicted by the machine learning model belongs to the m'th type of load information;

[0111] A first determining unit, configured to determine a total loss of the machine learning model according to the first loss function L1 and / or the second loss function L2;

[0112] The training unit is used to train the machine learning model according to the total loss of the machine learning model, and use the trained machine learning model as a load prediction model.

[0113] Exemplarily, the first determining unit includes:

[0114] A computing device for calculating the total loss L of the machine learning model according to the following formula f :

[0115] L f =αL1+βL2;

[0116] Among them, α is the preset weight of the first loss function; β is the preset weight of the second loss function.

[0117] Exemplarily, the adjustment module includes:

[0118] An acquisition unit, used to obtain a predicted value of the core usage rate at a future target moment by a load prediction model;

[0119] a second determining unit, configured to use a first preset voltage and a first preset frequency as the operating voltage and operating frequency of the GPU when the predicted value of the core usage rate is less than the first target ratio;

[0120] a third determining unit, configured to use the second preset voltage and the second preset frequency as the operating voltage and operating frequency of the GPU when the predicted value of the core usage rate is greater than or equal to the first target ratio and less than or equal to the second target ratio;

[0121] The fourth determining unit is configured to use the third preset voltage and the second preset frequency as the operating voltage and operating frequency of the GPU when the predicted value of the core usage rate is greater than the second target ratio.

[0122] For more specific working processes of the above modules, please refer to the corresponding content disclosed in Example 1, which will not be repeated here.

[0123] Example 3

[0124] This embodiment provides a computer device, including a processor and a memory; wherein the processor implements the steps of the GPU power consumption control method described in Example 1 when executing a computer program stored in the memory.

[0125] For more specific details about the above method, please refer to the corresponding content disclosed in Example 1, which will not be repeated here.

[0126] Example 4

[0127] This embodiment provides a computer-readable storage medium for storing a computer program; when the computer program is executed by a processor, the steps of the GPU power consumption control method described in Example 1 are implemented.

[0128] For more specific details about the above method, please refer to the corresponding content disclosed in Example 1, which will not be repeated here.

[0129] Example 5

[0130] This embodiment provides a computer program product, including computer-executable instructions or a computer program. When the computer-executable instructions or the computer program are executed by a processor, the steps of the GPU power consumption control method described in Example 1 are implemented.

[0131] For more specific details about the above method, please refer to the corresponding content disclosed in Example 1, which will not be repeated here.

[0132] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from the other embodiments. References to the same or similar parts between the various embodiments will be sufficient. The systems, devices, storage media, and computer program products disclosed in the embodiments correspond to the methods disclosed in the embodiments, so their descriptions are relatively simplified. For relevant details, refer to the method descriptions.

[0133] Those skilled in the art will clearly understand that the techniques in the embodiments of the present invention can be implemented using software plus a necessary general-purpose hardware platform. Based on this understanding, the technical solutions in the embodiments of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium such as ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments of the present invention or certain portions of the embodiments.

[0134] In some embodiments, computer-executable instructions may be in the form of a program, software, software module, script, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as a stand-alone program or as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0135] As an example, computer-executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file that stores other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinating files (e.g., files storing one or more modules, subroutines, or code portions).

[0136] By way of example, computer-executable instructions may be deployed to be executed on one electronic device, or on multiple electronic devices located at one site, or on multiple electronic devices distributed across multiple sites and interconnected by a communication network.

[0137] The present invention has been described in detail above with reference to specific embodiments and exemplary examples. However, these descriptions should not be construed as limiting the present invention. Those skilled in the art will appreciate that various equivalent substitutions, modifications, or improvements may be made to the technical solutions and implementations of the present invention without departing from the spirit and scope of the present invention, all of which fall within the scope of the present invention. The scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A GPU power consumption control method, characterized in that: include: S1, obtaining historical GPU load information within a first target time period; The historical load information includes historical core usage; S2: training a machine learning model based on historical GPU load information within the first target time period, and using the trained machine learning model as a load prediction model; S2 includes: Construct the first loss function L1: ； Where M is the total number of historical load information; The value of the mth load information predicted by the machine learning model; n m is the actual value of the mth load information; And / or, construct a second loss function L2: ； Where M' is the total number of categories of load information; h m' is the true label of the m'th category load information; The probability that the load information predicted by the machine learning model belongs to the m'th type of load information; Determine the total loss of the machine learning model according to the first loss function L1 and / or the second loss function L2; Train the machine learning model according to its total loss, and use the trained machine learning model as a load prediction model; S3, predicting GPU load information in a second future time period according to the load prediction model; S4, adjusting the operating voltage and operating frequency of the GPU in the second time period in the future according to the load information of the GPU in the second time period in the future; S5: Use the load information corresponding to the current operating voltage and operating frequency of the GPU as historical load information, and return to execute the operation of S1.

2. The GPU power consumption control method according to claim 1, wherein: Determining the total loss of the machine learning model according to the first loss function L1 and / or the second loss function L2 includes: The total loss L of the machine learning model is calculated according to the following formula f : L f =αL1+βL2; Among them, α is the preset weight of the first loss function; β is the preset weight of the second loss function.

3. The GPU power consumption control method according to claim 1, wherein: The adjusting the operating voltage and the operating frequency of the GPU in the second time period in the future according to the load information of the GPU in the second time period in the future includes: Obtain the load forecast model's predicted value of core utilization at a future target time; When the predicted value of the core usage rate is less than the first target ratio, the first preset voltage and the first preset frequency are used as the operating voltage and the operating frequency of the GPU; When the predicted value of the core usage rate is greater than or equal to the first target ratio and less than or equal to the second target ratio, the second preset voltage and the second preset frequency are used as the operating voltage and operating frequency of the GPU; When the predicted value of the core usage rate is greater than the second target ratio, the third preset voltage and the second preset frequency are used as the operating voltage and the operating frequency of the GPU.

4. A GPU power consumption control system, characterized in that: include: An acquisition module, configured to acquire historical load information of the GPU within a first target time period; The historical load information includes historical core usage; A training module is configured to train a machine learning model based on historical GPU load information within a first target time period and use the trained machine learning model as a load prediction model; the training module includes: The first construction unit is used to construct the first loss function L1: ； Where M is the total number of historical load information; The value of the mth load information predicted by the machine learning model; n m is the actual value of the mth load information; And / or, a second construction unit, for constructing a second loss function L2: ； Where M' is the total number of categories of load information; h m' is the true label of the m'th category load information; The probability that the load information predicted by the machine learning model belongs to the m'th type of load information; A first determining unit, configured to determine a total loss of the machine learning model according to the first loss function L1 and / or the second loss function L2; A training unit is used to train the machine learning model according to the total loss of the machine learning model, and use the trained machine learning model as a load prediction model; A prediction module, configured to predict GPU load information in a second future time period according to a load prediction model; An adjustment module, configured to adjust an operating voltage and an operating frequency of the GPU in the second time period in the future according to load information of the GPU in the second time period in the future; The determination module is used to use the load information corresponding to the current operating voltage and operating frequency of the GPU as the historical load information, and return to execute the operation of the acquisition module.

5. The GPU power consumption control system according to claim 4, wherein: The first determining unit includes: A computing device for calculating the total loss L of the machine learning model according to the following formula f : L f =αL1+βL2; Among them, α is the preset weight of the first loss function; β is the preset weight of the second loss function.

6. The GPU power consumption control system according to claim 4, wherein: The adjustment module includes: An acquisition unit, used to obtain a predicted value of the core usage rate at a future target moment by a load prediction model; a second determining unit, configured to use a first preset voltage and a first preset frequency as the operating voltage and operating frequency of the GPU when the predicted value of the core usage rate is less than the first target ratio; a third determining unit, configured to use the second preset voltage and the second preset frequency as the operating voltage and operating frequency of the GPU when the predicted value of the core usage rate is greater than or equal to the first target ratio and less than or equal to the second target ratio; The fourth determining unit is configured to use the third preset voltage and the second preset frequency as the operating voltage and operating frequency of the GPU when the predicted value of the core usage rate is greater than the second target ratio.

7. A computer device, characterized in that: The method comprises a processor and a memory; wherein, when the processor executes the computer program stored in the memory, the steps of the GPU power consumption control method according to any one of claims 1 to 3 are implemented.

8. A computer-readable storage medium, characterized in that Used to store a computer program; when the computer program is executed by a processor, the steps of the GPU power consumption control method according to any one of claims 1 to 3 are implemented.

Images