Method, device and equipment for dynamically adjusting power consumption of video card component

Abstract

The invention relates to the technical field of graphics card power consumption adjustment, and discloses a method, a device and equipment for dynamically adjusting the power consumption of a graphics card component. The method comprises the following steps: carrying out multi-channel data acquisition and working load characteristic analysis on operation parameters of a graphics card GPU core, a video memory and a power management unit to obtain a component temperature state evaluation matrix; executing component-level power consumption distribution based on multi-objective optimization to obtain a power consumption distribution scheme of each component; inputting the power consumption distribution scheme of each component into a power consumption-frequency-voltage mapping model, and calculating an optimal frequency voltage configuration parameter through a binary search algorithm to obtain a frequency voltage control instruction; performing multi-step prediction to obtain a module power consumption demand prediction value; according to the method, the frequency voltage combination with the highest performance efficiency can be found under the condition that the power consumption constraint is met, and accurate power consumption control is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of graphics card power consumption regulation, and particularly to a method, device and equipment for dynamically regulating the power consumption of graphics card components. Background Art

[0002] With the rapid development of computer graphics processing and artificial intelligence applications, the graphics card, as a core hardware component, undertakes increasingly complex computing tasks, and its power consumption management problem has become increasingly prominent. Traditional graphics card power consumption management technologies mainly adopt a global power consumption limit strategy with a fixed threshold, which cannot perform fine-grained regulation according to the load characteristics of different components, resulting in problems such as hotspot concentration and energy efficiency imbalance in high-load scenarios. This coarse-grained power consumption management method is difficult to adapt to the increasingly diverse application requirements, especially in applications with different load characteristics such as game rendering, AI inference, and video encoding and decoding, showing obvious limitations.

[0003] Although the dynamic power consumption management technology commonly used in the industry currently introduces a real-time monitoring mechanism, it still has inherent defects. The existing technology lacks the flexibility to perform configurable on-demand power consumption optimization under different workloads, and cannot effectively handle the mutual influence and spatio-temporal correlation of power consumption among internal components of the graphics card. Especially in a working environment with large temperature fluctuations, due to the lack of component-level fine control ability and forward-looking prediction mechanism, it often leads to the system overly conservatively reducing the overall performance, or overly aggressively increasing the power consumption, triggering the thermal protection mechanism, resulting in serious fluctuations in user experience and a reduction in the service life of the graphics card. Summary of the Invention

[0004] The present invention provides a method, device and equipment for dynamically regulating the power consumption of graphics card components. The present invention can find the frequency-voltage combination with the highest performance efficiency under the condition of meeting the power consumption constraint, realizes precise power consumption control, avoids the blindness and inefficiency of frequency-voltage regulation in traditional methods, and improves the energy utilization efficiency.

[0005] In a first aspect, the present invention provides a method for dynamically regulating the power consumption of graphics card components, and the method for dynamically regulating the power consumption of graphics card components includes: Performing multi-channel data acquisition and workload characteristic analysis on the operating parameters of the graphics card GPU core, video memory and power management unit to obtain a component temperature state evaluation matrix; According to the component temperature state evaluation matrix, performing component-level power consumption allocation based on multi-objective optimization to obtain each component power consumption allocation scheme; Inputting each component power consumption allocation scheme into a power consumption-frequency-voltage mapping model, and calculating optimal frequency-voltage configuration parameters through a binary search algorithm to obtain a frequency-voltage control instruction; Input the standardized data set into the adaptive multi-frequency adjustment learning network for multi-step prediction to obtain the predicted value of the component power consumption demand; According to the frequency-voltage control instruction and the predicted value of the component power consumption demand, perform dynamic adjustment of the graphics card component power consumption to obtain the feedback result of the dynamic power consumption adjustment.

[0006] In a second aspect, the present invention provides a device for dynamically adjusting the power consumption of a graphics card component. The device for dynamically adjusting the power consumption of a graphics card component includes: A feature analysis module, configured to perform multi-channel data acquisition and workload feature analysis on the operating parameters of the graphics card GPU core, video memory, and power management unit to obtain a component temperature state evaluation matrix; A power consumption allocation module, configured to perform component-level power consumption allocation based on multi-objective optimization according to the component temperature state evaluation matrix to obtain a power consumption allocation scheme for each component; A calculation module, configured to input the power consumption allocation schemes for each component into a power consumption-frequency-voltage mapping model, and calculate optimal frequency-voltage configuration parameters through a binary search algorithm to obtain a frequency-voltage control instruction; A multi-step prediction module, configured to input the standardized data set into the adaptive multi-frequency adjustment learning network for multi-step prediction to obtain the predicted value of the component power consumption demand; A dynamic adjustment module, configured to perform dynamic adjustment of the graphics card component power consumption according to the frequency-voltage control instruction and the predicted value of the component power consumption demand to obtain the feedback result of the dynamic power consumption adjustment.

[0007] In a third aspect of the present invention, a device for dynamically adjusting the power consumption of a graphics card component is provided, including: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the device for dynamically adjusting the power consumption of a graphics card component executes the above-mentioned method for dynamically adjusting the power consumption of a graphics card component.

[0008] In the technical solution provided by the present invention, the present invention realizes the accurate monitoring of the operating parameters at the component level by setting up a dedicated sensor network in the GPU core, video memory and power management unit, and combines multi-dimensional feature extraction technologies such as Fourier transform, wavelet transform and autocorrelation analysis to accurately identify the workload type, providing a precise data basis for subsequent power consumption regulation, and effectively solving the problems of rough data collection and inaccurate workload feature extraction in traditional methods. The multi-level temperature threshold mechanism set according to the characteristics of different components realizes the refined hierarchical management of the temperature state. By dynamically adjusting the threshold to respond to different workload characteristics, it avoids the over-conservative or over-aggressive problems of the traditional fixed threshold scheme in different application scenarios, and improves the accuracy and flexibility of temperature control. The present invention establishes a multi-objective optimization function that simultaneously considers performance maximization and temperature risk control, and combines the Lagrange multiplier method to solve the optimal power consumption allocation scheme. On the premise of ensuring the overall power consumption budget constraint, it realizes the reasonable allocation of component-level power consumption resources, and effectively balances the relationship between performance output and temperature control. By constructing a power consumption-frequency-voltage mapping model and applying the binary search algorithm, the present invention can find the frequency-voltage combination with the highest performance efficiency under the condition of meeting the power consumption constraint, realizing accurate power consumption control, avoiding the blindness and inefficiency of frequency-voltage regulation in traditional methods, and improving the energy utilization efficiency. The adaptive multi-frequency regulation learning network is used to perform multi-step prediction on the power consumption demand. By capturing spatio-temporal feature correlations through multi-channel convolution and self-attention mechanisms, it realizes accurate prediction of future power consumption demand, overcomes the limitations of traditional methods that only rely on historical data for simple prediction, and provides a reliable guarantee for predictive power consumption regulation. The present invention integrates the dual mechanisms of reactive regulation and predictive regulation. Through a multi-level regulation control framework and a priority sorting strategy, it prepares in advance for workload changes while dealing with temperature anomalies, realizing the combination of timeliness and forward-looking of power consumption regulation, and improving the stability and response speed of the system. By analyzing the mutual influence between components through the power consumption influence matrix between components, the state changes of other components are considered collaboratively when adjusting the power consumption of a certain component, avoiding the resource imbalance problem caused by independent regulation between components in traditional methods, and improving the coordination and efficiency of the overall system. Description of the Drawings

[0009] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0010] Figure 1 It is a schematic flow chart of the method for dynamically adjusting the power consumption of the graphics card components provided by the embodiments of the present application; Figure 2Schematic block diagram of the device for dynamically adjusting the power consumption of a graphics card component provided by an embodiment of the present application; Figure 3 Schematic block diagram of the device for dynamically adjusting the power consumption of a graphics card component provided by an embodiment of the present application. Detailed implementation manners

[0011] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0012] The flowchart shown in the accompanying drawings is only an example illustration, and does not necessarily include all contents and operations / steps, nor does it necessarily need to be executed in the described order. For example, some operations / steps can also be decomposed, combined, or partially merged. Therefore, the actual execution order may change based on the actual situation.

[0013] It should also be understood that the terms used in the specification of the present application are only for the purpose of describing specific embodiments and are not intended to limit the present application. As used in the specification of the present application and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0014] It should be further understood that the term " / and" as used in the specification of the present application and the appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0015] Next, some embodiments of the present application will be described in detail in conjunction with the accompanying drawings. Without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0016] Please refer to Figure 1 , Figure 1 which is a schematic flowchart of the method for dynamically adjusting the power consumption of a graphics card component provided by an embodiment of the present application. As Figure 1 shown, the method for dynamically adjusting the power consumption of a graphics card component provided by an embodiment of the present application includes steps S100 to S500.

[0017] Step S100: Perform multi-channel data acquisition and workload characteristic analysis on the operating parameters of the graphics card GPU core, video memory, and power management unit to obtain a component temperature status evaluation matrix; It can be understood that the execution subject of the present invention can be a graphics card component power consumption dynamic adjustment device, or a terminal or a server, and specific limitations are not made here. In the embodiments of the present invention, the server is taken as an execution subject for illustration.

[0018] Specifically, temperature sensors are arranged in the GPU core area to monitor the spatial distribution of its core temperature. Voltage sensors are installed in the video memory module to monitor the power supply status of the video memory in real time. At the same time, current sensors are configured at the power management unit to track the power output. The acquisition frequency of these sensors is set to once every 50 milliseconds to ensure that the data has high timeliness, and all the collected raw parameter data is transmitted to the data processing unit through a dedicated data bus. In the data processing unit, preprocessing operations are performed on the raw data, including denoising, normalization, and outlier filtering. The purpose of denoising is to eliminate high-frequency noise caused by environmental interference or sensor errors to improve the reliability of the data; normalization is used to eliminate the scale differences between different parameters so that all data can be processed within the same numerical range to improve the accuracy of subsequent feature analysis; outlier filtering identifies and removes outliers through methods such as setting thresholds or statistical analysis to prevent sudden error data from affecting the accuracy of the overall analysis. After preprocessing, all the data is standardized and stored as a data set containing timestamps and component parameter vectors, where the timestamps are used to identify the timing characteristics of data acquisition, and the component parameter vectors contain the temperature data of the GPU core, the voltage data of the video memory, and the current data of the power management unit. The standardized data set is input into the signal processing module to extract more representative features. Fourier transform is performed on the data to analyze the periodic characteristics of the power consumption data from the frequency domain perspective to identify the power consumption patterns of the graphics card under different workloads. Wavelet transform is used to analyze the non-linear characteristics of the data to obtain short-term power consumption changes and local load fluctuations, and effectively capture sudden load change patterns, such as random surges in video memory access or transient fluctuations in current demand. At the same time, through the autocorrelation analysis method, the temporal dependence of the data is analyzed to identify the short-term and long-term correlations of load changes, so as to predict future power consumption trends. These analysis steps finally construct a workload feature matrix, where each row represents the key power consumption features within a time period, and the column vectors contain the feature information extracted by different signal analysis methods. The workload feature matrix is input into a pre-trained decision tree classifier to achieve automatic workload type identification. The decision tree classifier is trained based on historical data. By comparing the feature distributions under different load patterns, it quickly determines what type the current workload belongs to, such as graphics rendering, AI inference, or video encoding and decoding, etc. When the input data highly matches a certain historical pattern, the classifier outputs the corresponding load type identifier. Based on the standardized data set and the workload type identifier, multi-level temperature thresholds are set for the GPU core, video memory, and power management unit respectively to achieve dynamic temperature management for different workloads.For the GPU core, four levels of temperature thresholds are set, namely 65°C, 75°C, 85°C, and 95°C, corresponding to low load, medium load, high load, and critical state respectively; for the video memory, the temperature thresholds are set to 60°C, 70°C, 80°C, and 90°C to reflect the relationship between the video memory power supply state and the workload; while for the power management unit, its temperature thresholds are set to 55°C, 65°C, 75°C, and 85°C to ensure that the current management system operates within a reasonable range. Based on these settings, a component temperature status evaluation matrix is constructed, which is used to identify the current temperature status of each component and adopts a five-level classification method, namely low temperature, normal, warning, high temperature, and critical state.

[0019] Extract the historical temperature data of each component from the standardized dataset to obtain the temperature change patterns of each component during long-term operation, reflecting the temperature distribution of the GPU core, video memory, and power management unit under different workload conditions. Through statistical analysis methods, such as calculating the temperature mean, variance, and temperature distribution density of each component under different load conditions, to accurately describe its historical temperature distribution. According to the component historical temperature distribution and workload type identification, set four-level temperature thresholds for the GPU core, video memory, and power management unit respectively to form a component temperature threshold set. During the process of setting the temperature threshold of the GPU core, combined with its temperature characteristics under different workload conditions, divide its temperature into low load, medium load, high load, and critical states, and correspond to the temperature thresholds of 65°C, 75°C, 85°C, and 95°C respectively, to ensure stable operation without triggering overheat protection when the load is high. For the video memory, its temperature thresholds are divided into 60°C, 70°C, 80°C, and 90°C to adapt to the video memory power consumption requirements under different workloads and avoid power supply instability or data transmission errors caused by too high temperature. As for the power management unit, set the corresponding temperature thresholds according to its current load characteristics, and limit its temperature to 55°C, 65°C, 75°C, and 85°C to ensure that it can work within a reasonable temperature range and avoid the stability decline of the power module caused by too high temperature. Since different workload types have different effects on power consumption, when setting the temperature threshold, make fine-tuning according to the load type. For example, for compute-intensive tasks such as AI inference, appropriately increase the temperature threshold by 2°C to improve the performance space, while for memory-intensive tasks such as video encoding and decoding, appropriately reduce the video memory temperature threshold by 1°C to reduce the impact of high temperature on the video memory data stability. Organize the temperature threshold set of all components into a component temperature threshold matrix for real-time temperature status evaluation. In this matrix, each row corresponds to a component, such as the GPU core, video memory, and power management unit, and each column corresponds to different temperature threshold levels, such as low load, medium load, high load, and critical state. On this basis, extract the real-time temperature data of each component from the standardized dataset and compare it with the corresponding thresholds in the component temperature threshold matrix to determine the state level of the current temperature. The system traverses the real-time temperature data of the GPU core, video memory, and power management unit one by one and checks whether its temperature is in the low temperature, normal, warning, high temperature, or critical state in turn. For example, when the real-time temperature of the GPU core falls between 75°C and 85°C, its state will be marked as high load, and when the real-time temperature of the video memory exceeds 90°C, its state will be marked as critical. Similarly, if the temperature of the power management unit is higher than 85°C, it will be determined to be in a critical state and trigger a power consumption adjustment mechanism to reduce the temperature. Through real-time temperature monitoring and threshold comparison, quickly determine the current temperature state of each component and construct a component temperature state evaluation matrix accordingly.

[0020] Step S200: According to the component temperature state evaluation matrix, perform component-level power consumption allocation based on multi-objective optimization to obtain the power consumption allocation scheme for each component; Specifically, set the overall power consumption budget according to the graphics card model, that is, the thermal design power (TDP) value. At the same time, determine the maximum allowable power consumption values of the GPU core, video memory, and power management unit to ensure that the entire system operates within a reasonable power consumption range. The TDP value is set according to the hardware design and heat dissipation capacity of the graphics card. At the same time, the maximum allowable power consumption values of the GPU core, video memory, and power management unit also need to be set according to their actual power consumption characteristics to build a reasonable overall power consumption budget constraint to meet the basic requirements of thermal management and power supply stability. On this basis, use the status values in the component temperature status evaluation matrix to calculate the power consumption reduction coefficient of each component, so as to appropriately limit the power consumption of high-temperature components during power consumption allocation. The calculation of the power consumption reduction coefficient is based on the current temperature status of the component. When the component is in a low-temperature or normal state, the power consumption reduction coefficient is zero, that is, the component is allowed to operate at the maximum allowable power consumption. When the component temperature enters the warning or high-temperature state, the power consumption reduction coefficient increases with the increase in temperature, thereby reducing its power consumption allocation to avoid further temperature rise. After applying these power consumption reduction coefficients to the maximum allowable power consumption value, the power consumption upper limit constraint affected by temperature is obtained. At the same time, based on the workload type identifier, extract the mapping coefficient between component power consumption and performance from the pre-established performance mapping database and construct a performance objective function. Since different workload types have significant differences in the requirements for the GPU, video memory, and power management unit, the performance mapping database stores the power consumption and performance relationship under different workload conditions and establishes a mapping model based on historical operation data and regression analysis. The calculation of the performance objective is based on the power consumption contribution of the component. Among them, the performance of the GPU core has a greater impact, so its requirements will be given priority when allocating power consumption, while the power consumption requirements of the video memory and power management unit are adjusted according to specific tasks. For example, video encoding and decoding tasks require more video memory power consumption, while the power consumption of the power management unit needs to ensure stable power supply under high load to maintain the reliability of the entire system. At the same time, to ensure that the system does not cause excessive temperature risks while optimizing performance, build a temperature risk assessment function based on the component temperature status evaluation matrix and set weight coefficients according to the workload type to form a multi-objective optimization goal. The method of temperature risk assessment is to combine the temperature status value of each component with the power consumption allocation amount. For example, if the temperature status of a certain component is at a high temperature level and its power consumption ratio is too high, the system needs to reduce its power consumption allocation to reduce the possibility of further temperature rise. The setting of the weight coefficient is adjusted according to the current workload type. For example, for performance-sensitive loads, the weight of the performance objective is set higher, while the weight of the temperature risk is set lower. For temperature-sensitive loads, the adjustment is reversed, making the priority of temperature control higher. Find the optimal balance between performance and temperature risk on the premise of ensuring that the overall power consumption budget does not exceed the standard.The overall power consumption budget constraint, temperature-sensitive power consumption upper limit constraint, performance objective function, and multi-objective optimization function are combined into a complete optimization problem, and the Lagrange multiplier method is used to solve it to obtain the optimal power consumption allocation scheme. The Lagrange multiplier method transforms the constraint conditions into part of the optimization objective by introducing additional variables, so that the optimal power consumption allocation value can be solved on the premise of meeting the constraints. During the solution process, the power consumption budget limit, the power consumption upper limit of each component, and the temperature constraint are considered simultaneously, and an iterative optimization method is used to find the power consumption allocation scheme that can maximize the performance and meet the constraint conditions. The optimal power consumption allocation scheme is verified to ensure that it meets all constraint conditions, especially the overall power consumption budget constraint. The verification process includes calculating the sum of the power consumption of all components and checking whether it exceeds the TDP limit, while ensuring that the power consumption allocation value of each component does not exceed its temperature-affected power consumption upper limit. At the same time, check whether the power consumption allocation of any component is too low, resulting in a decrease in system performance. If this situation is found, the optimization weights need to be adjusted and the optimization process needs to be re-executed. The verified power consumption allocation scheme is determined as the final scheme.

[0021] Step S300: Input the power consumption allocation scheme of each component into the power consumption-frequency-voltage mapping model, calculate the optimal frequency-voltage configuration parameters through the binary search algorithm, and obtain the frequency-voltage control instruction; Specifically, power consumption-frequency-voltage mapping data tables are constructed for the GPU core, video memory, and power management unit respectively to record the corresponding relationships between frequency and voltage of each component under different power consumption conditions. The establishment of this mapping database depends on a large amount of experimental data and historical operation records. Among them, the mapping data of the GPU core needs to cover different working states from low power consumption to high power consumption, and record the corresponding frequency and voltage values under different power consumption input conditions; the mapping data of the video memory needs to consider the relationship between bandwidth requirements and power consumption to ensure a stable data transfer rate at different power consumption levels; the mapping data of the power management unit needs to consider the relationship between input current, output voltage, and overall power conversion efficiency, so as to construct a component mapping database, enabling the system to dynamically adjust frequency and voltage under different workloads to optimize power consumption management. Input the power consumption values in the power consumption allocation scheme of each component into the corresponding component mapping database respectively, and query the candidate frequency-voltage combinations to form an initial frequency-voltage candidate set. According to the power consumption allocation values of the GPU core, video memory, and power management unit provided by the power consumption allocation scheme, find the matching frequency-voltage pairs in the database and filter out the candidate combinations that meet the requirements. These candidate combinations include multiple pairs of frequency-voltage values. The system will preferentially select the combinations that meet the power consumption constraints and sort them according to different operating efficiencies for subsequent optimization to select the optimal configuration scheme. After obtaining the initial frequency-voltage candidate set, set the upper and lower bounds of the frequency for each component to determine a reasonable search range. The upper and lower bounds of the GPU core frequency are set by the graphics card manufacturer; the frequency range of the video memory is determined according to its model; the voltage range of the power management unit is relatively fixed, but it still needs to ensure that it fluctuates within a safe range. After determining the search range, the system performs a binary search within this range to improve the efficiency of the optimization calculation. The process of binary search first selects the middle frequency point of the search range as the test point, calculates the voltage value corresponding to this frequency point according to the mapping database, and then judges whether the calculated voltage meets the constraint conditions of the current power consumption allocation scheme. If the voltage corresponding to this frequency point causes the power consumption to exceed the allocated power consumption value, the search range needs to be narrowed and a lower frequency is selected for continued testing; on the contrary, if the voltage corresponding to this frequency point causes the power consumption to be lower than the allocated power consumption value, a higher frequency is tried to improve the performance of the system. By continuously narrowing the search range, finally, a frequency-voltage combination that meets the power consumption constraints is found. Calculate the performance efficiency index for the frequency-voltage values that meet the power consumption constraints to select the combination with the highest performance efficiency as the final configuration scheme. The calculation of performance efficiency involves the computing power per unit power consumption, that is, under the same power consumption conditions, the configuration that can provide higher computing throughput or graphics rendering ability will be preferentially selected.Since different frequency-voltage combinations can affect the overall power consumption curve of the system, when calculating the performance efficiency, the mutual influence between components is considered, such as whether the increase in the frequency of the GPU core will lead to an increase in the load of the video memory, or whether the voltage change of the power management unit will affect the overall stability of the system. By evaluating the performance efficiency of all combinations that meet the power consumption constraints, the optimal frequency-voltage parameters are selected. The optimal frequency-voltage parameters and the execution time information are encapsulated in the format of a control instruction to generate a frequency-voltage control instruction. This control instruction contains clear component identifiers to ensure that each component can correctly receive and execute the adjustment command; it also contains the target frequency and target voltage to ensure that the adjusted operating state can conform to the optimization plan; at the same time, the control instruction contains the execution time information to ensure that the adjustment operation can be carried out within an appropriate time window to avoid sudden frequency-voltage changes from affecting the stability of the system. The frequency-voltage control instruction is sent to the GPU core, the video memory, and the power management unit to make them operate according to the optimal frequency-voltage parameters.

[0022] Step S400: Input the standardized data set into the adaptive multi-frequency adjustment learning network for multi-step prediction to obtain the predicted value of the component power consumption demand; Specifically, a time series is constructed for the standardized dataset to capture the power consumption change trends of different components over a historical time period. The standardized dataset includes the temperature, voltage, and power consumption information of the GPU core, the power supply status and bandwidth utilization of the video memory, as well as the current output and power consumption data of the power management unit. These data are segmented according to a time window of fixed length to form a time series data structure, ensuring that the subsequent model can fully learn the power consumption change patterns of each component at different time steps. After segmentation by the time window, each time segment contains power consumption-related data for multiple time steps, thereby constructing time series data of component parameters, enabling the system to use the power consumption change information over a past period of time to predict future power consumption requirements. The time series data of component parameters is fed into the data preprocessing layer of the adaptive multi-frequency regulation learning network to ensure that the data meets the requirements of neural network modeling. Z-score standardization is performed on the data to eliminate the dimensional differences between different data sources, enabling parameters such as power consumption, temperature, voltage, and current of all components to be calculated and compared within the same scale range. The process of Z-score standardization normalizes the original data by calculating the mean and standard deviation of the data to ensure the balance of the input data distribution and improve the stability of model training. The data is reduced in dimension through principal component analysis to reduce redundant information and extract the most representative feature variables, thereby reducing the computational complexity and improving the generalization ability of the model. In this way, the denoised feature representation more accurately reflects the main operating characteristics of each component. The denoised feature representation is input into the adaptive multi-sampling module of the adaptive multi-frequency regulation learning network. This module automatically determines the optimal sampling frequency according to the power consumption fluctuation characteristics of different components and constructs multiple power consumption monitoring channels to ensure that the system tracks power consumption changes at different time scales. Since the power consumption of the GPU core usually fluctuates relatively quickly, while the power consumption of the video memory changes relatively slowly, and the power consumption of the power management unit is affected by multiple factors, the adaptive multi-sampling module analyzes the power consumption fluctuation patterns of each component and dynamically adjusts the sampling rates of different channels, enabling the monitoring channels of the GPU core to collect data at a higher frequency, while the monitoring channels of the video memory and the power management unit use a lower sampling rate to reduce data redundancy and improve computational efficiency. The multi-frequency monitoring sequence output by this module can cover power consumption changes at different time scales. The multi-frequency monitoring sequence is input into the multi-channel convolutional layer of the adaptive multi-frequency regulation learning network to extract the time feature patterns of each component by parallel processing convolutional kernels of different scales, obtaining component time feature maps. In this process, the network uses multiple convolutional kernels of different sizes to perform parallel calculations on the data. Smaller convolutional kernels are used to capture short-term power consumption changes, such as the transient increase in power consumption of the GPU core under high-load tasks, while larger convolutional kernels are used to identify long-term trends, such as the slow change in video memory bandwidth utilization.Through the calculation of the multi-channel convolutional layer, the characteristic information of different components in the time dimension is effectively extracted, enabling the network to more accurately model the variation law of power consumption over time. After extracting the time feature map, the self-attention module of the adaptive multi-frequency adjustment learning network is used to calculate the correlation weight matrix between different components to selectively focus on key power consumption change regions and obtain context features with spatial dependence. Since the power consumption changes of the GPU core, video memory, and power management unit are interrelated, the self-attention module dynamically allocates attention weights by calculating the dependence relationship between components, making the network more focused on the key regions affecting power consumption changes. Through the attention mechanism, the prediction ability of the model under complex load conditions is effectively improved, and the accuracy of power consumption demand prediction is increased. The context features are input into the gated recurrent unit of the adaptive multi-frequency adjustment learning network to predict the component power consumption demand values of the GPU core, video memory, and power management unit at multiple future time steps through the time series memory mechanism. The gated recurrent unit can retain important information in long time series data while filtering out irrelevant data, thus effectively modeling the long-term dependence of the power consumption of each component. During the prediction process, the network combines historical power consumption data, the current system state, and the interaction relationship between components to generate the power consumption demand curve for the future time period. The prediction model provides a multi-step prediction result, enabling the system to adjust the power management strategy in advance to ensure that the graphics card can maintain the best power consumption performance balance when the future load changes.

[0023] Step S500: According to the frequency-voltage control instruction and the predicted value of component power consumption demand, perform dynamic adjustment of the graphics card component power consumption to obtain the dynamic adjustment feedback result of power consumption.

[0024] Specifically, the frequency-voltage control instruction and the predicted value of component power consumption demand are input into the multi-level adjustment control framework. The multi-level adjustment control framework consists of a real-time monitoring layer, a policy decision layer, and an execution control layer. The real-time monitoring layer continuously collects the operating states of various components of the graphics card, including data such as the power consumption and temperature of the GPU core, the voltage and bandwidth usage of the video memory, and the power output of the power management unit. At the same time, it monitors the current workload type to ensure that the adjustment strategy adapts to the actual operating requirements of the system. These real-time data are transmitted to the policy decision layer, which parses the frequency-voltage control instruction and generates an initial adjustment control strategy in combination with the predicted value of power consumption demand. In this process, the current power consumption state of each component and the power consumption demand in a future period of time are evaluated to ensure that the adjusted power management strategy can not only meet the performance requirements but also effectively control temperature and power consumption fluctuations. The policy decision layer transmits the initially calculated control strategy to the execution control layer, which converts the adjustment instruction into specific hardware control signals to perform corresponding frequency, voltage, and power adjustments. The initial adjustment control strategy is prioritized, and the temperature status exception response is processed first because too high a temperature of the graphics card can cause hardware damage or trigger an emergency downclocking mechanism. Therefore, when the temperature status of a certain component reaches a high or critical level, its power consumption is reduced to ensure that its temperature remains within a safe range. Secondly, the power consumption overlimit response is processed, that is, when the power consumption of a certain component exceeds its set allocated power, its frequency and voltage are adjusted to prevent the overall power consumption from exceeding the TDP budget and affecting the overall stability of the graphics card. Perform predictive adjustment operations, which are based on the predicted value of power consumption demand and adjust the power consumption of each component in advance to adapt to future load changes. Through the prioritization method, ensure that the graphics card can always remain stable during operation and reasonably adjust power consumption in different scenarios. At the same time, calculate the adjustment period according to the current workload characteristics and component temperature status to achieve an adaptive adjustment time interval. The calculation of the adjustment period needs to comprehensively consider multiple factors, including the current temperature status, the power consumption change rate, and the load fluctuation situation. For example, when the system temperature is close to the critical value or the load fluctuates greatly, the adjustment period needs to be shortened to improve the response speed, while when the system runs stably and the temperature fluctuates little, the adjustment period is appropriately extended to reduce the system overhead caused by frequent adjustments. In addition, analyze the historical operation data to dynamically adjust the adjustment time interval so that it can not only adapt to short-term load changes but also maintain the stability of long-term power management. The calculated adaptive adjustment time interval is used to control the execution frequency of the adjustment operation to ensure that the system can perform power consumption adjustment at the most appropriate time. After determining the adjustment period, analyze the cross-impacts of power consumption adjustment of each component based on the pre-established power consumption impact matrix between components and perform coordinated and optimized power consumption adjustments if necessary.When power consumption limitation is required for a certain component, the system not only needs to adjust the power consumption of this component, but also analyze its dependency relationships with other components, and accordingly adjust the power consumption allocation of related components to obtain a co-optimized power control instruction. Through the co-optimization strategy, the problem of resource imbalance caused by individually adjusting the power consumption of a certain component is avoided, and the overall operating efficiency of the graphics card is improved. The co-optimized power control instruction is sent to the controllers of the GPU core, video memory, and power management unit through the graphics card driver layer interface to perform specific frequency adjustment, voltage setting, and power consumption limitation operations. The graphics card driver layer interface converts high-level adjustment instructions into commands executable by hardware and ensures that these commands are correctly executed on each component of the graphics card. The power control instruction of the GPU core adjusts its frequency and voltage to ensure that it operates at the optimal power consumption level; the control instruction of the video memory adjusts its operating voltage and access bandwidth to adapt to the current load requirements; the control instruction of the power management unit optimizes the current output to improve the overall power consumption conversion efficiency. All these adjustment operations strictly comply with the power consumption allocation scheme to ensure that the total power consumption of the graphics card always remains within a safe range. After the power consumption adjustment is completed, the actual operating status of each component is monitored for performance indicators, and a power consumption dynamic adjustment feedback result is generated based on the collected data. This monitoring process includes tracking the temperature change, power consumption fluctuation, frequency adjustment situation, and overall performance of each component to evaluate the effect of the adjustment operation. And record the historical data of the adjustment operation, and improve the future power management strategy through data analysis to improve the overall adjustment accuracy and efficiency. The power consumption dynamic adjustment feedback result is used to optimize the next round of adjustment decisions, enabling the graphics card to achieve the best power consumption and performance balance under different load scenarios.

[0025] In a specific embodiment, the process of executing step S100 may specifically include the following steps: Set a temperature sensor in the GPU core of the graphics card, a voltage sensor in the video memory, and a current sensor in the power management unit, and collect the original operating parameter data of the components; Perform denoising, normalization, and outlier filtering on the original operating parameter data of the components to obtain a standardized data set including timestamps and component parameter vectors; Perform Fourier transform, wavelet transform, and autocorrelation analysis on the standardized data set, extract frequency domain features, non-linear features, and time dependencies to obtain a workload feature matrix; Input the workload feature matrix into a pre-trained decision tree classifier, classify by comparing historical data patterns, and obtain a workload type identifier; Based on the standardized data set and the workload type identifier, set multi-level temperature thresholds for the GPU core, video memory, and power management unit respectively to obtain a component temperature status evaluation matrix.

[0026] Specifically, temperature sensors are set in the GPU core area to monitor the real-time changes in its operating temperature and ensure timely identification of the temperature rising trend under high-load conditions. At the same time, voltage sensors are installed in the video memory area to record the real-time changes in the power supply status of the video memory, thereby providing a basis for voltage regulation and avoiding affecting the stability of data transmission due to voltage fluctuations. Current sensors are configured in the power management unit to monitor the load conditions of the overall power supply system in real time and ensure that the power output always remains within a reasonable range. The acquisition frequency of these sensors is set to one measurement within the time interval, and is transmitted to the data processing unit through a high-precision data bus to ensure the timeliness and accuracy of the data. For the time sampled data is represented as a vector: Among them, represents the temperature data of the GPU core, is the voltage data of the video memory, is the current data of the power management unit. The collected data is denoised, normalized, and outlier filtered. The denoising process uses the wavelet threshold denoising method, that is, the data is decomposed by wavelet transform, and an adaptive threshold is set for the high-frequency noise part to weaken it, so that the data retains the main signal characteristics while reducing noise interference. Assuming that the wavelet decomposition coefficient of the signal is , then its denoised data is represented as: Among them, is the inverse wavelet transform, is the adaptive threshold, is the indicator function, which is used to retain the valid signal greater than the threshold. The normalization process uses mean-standard deviation normalization, that is, by calculating the mean and standard deviation of each parameter, the data is normalized to the form of zero mean and unit variance: to keep the data of different physical quantities within the same numerical range for subsequent analysis. At the same time, in order to improve the quality of the data, outlier detection is carried out. Assuming that the normal range of the data follows a Gaussian distribution, outliers are detected by setting a threshold multiples of the standard deviation, that is, when: it is considered that the data point is an outlier, and linear interpolation is used for correction to ensure data continuity. After the above preprocessing, the data set is converted into a standardized data set , which contains timestamps and the corresponding component parameter vectors: After data preprocessing is completed, key features of the data are extracted for workload analysis. The frequency-domain features of the data are extracted using Fourier transform to identify the periodic power consumption fluctuations under different load patterns. Let the power consumption signal be , and its Fourier transform is expressed as: where is the frequency component, represents the contribution of the power consumption data at different frequencies. By analyzing the results of the Fourier transform, the main power consumption patterns of the graphics card within a specific frequency range are identified. The wavelet transform is used to extract the non-linear features of the data. The wavelet transform analyzes the signal simultaneously in the time-frequency domain to capture the power consumption change patterns within a short period. Let the wavelet basis function be , then the wavelet transform of the power consumption signal is expressed as: where is the scale factor, is the translation factor, represents the power consumption signal response at different scales and time offsets. By analyzing the changes in the wavelet coefficients, the power consumption change patterns of the graphics card under different load scenarios are discovered, and the power consumption adjustment strategy is optimized. Autocorrelation analysis is used to identify time dependence. The definition of the autocorrelation function is as follows: where represents the correlation of the data at the time interval . If the autocorrelation value has a high peak within a specific interval, it indicates that there are periodic fluctuations in the power consumption data, and this information is used for subsequent load prediction and adjustment. The feature matrix is input into a pre-trained decision tree classifier to identify the current workload type. Assuming the workload type set is , then the classification process is expressed as: where represents the decision tree classifier, is the identified workload type, such as game rendering, AI inference, or video codec, etc. Based on the standardized dataset and the workload type identification, multi-level temperature thresholds are set for the GPU core, video memory, and power management unit respectively to optimize the operating stability of the graphics card. Let the temperature threshold set be: Among them, the temperature threshold of each component is dynamically adjusted according to its workload type. For example, for high-load tasks such as AI inference, the temperature threshold of the GPU core is appropriately increased, while for memory-intensive tasks such as video encoding and decoding, the temperature threshold of the video memory is reduced to reduce data transfer latency. All temperature thresholds are organized into a component temperature status evaluation matrix.

[0027] In a specific embodiment, the process of setting multi-level temperature thresholds for the GPU core, video memory, and power management unit respectively based on the standardized dataset and workload type identifier to obtain the component temperature status evaluation matrix may specifically include the following steps: Extract the historical temperature data of the GPU core, video memory, and power management unit from the standardized dataset to obtain the component historical temperature distribution; According to the component historical temperature distribution and workload type identifier, set four-level temperature thresholds for the GPU core, four-level temperature thresholds for the video memory, and four-level temperature thresholds for the power management unit to obtain a set of component temperature thresholds; Organize the set of component temperature thresholds into a component temperature threshold matrix; Extract the real-time temperature data of each component from the standardized dataset, compare it with the corresponding threshold in the component temperature threshold matrix, and obtain the temperature status value of each component; Construct a component temperature status evaluation matrix according to the temperature status values of each component.

[0028] Specifically, extract the historical temperature data of the GPU core, video memory, and power management unit from the standardized dataset, set adaptive temperature thresholds based on the historical temperature distribution and workload type, and construct a temperature status evaluation matrix at the same time. Let the temperature dataset in the standardized dataset be , which contains the measured values of the GPU core, video memory, and power management unit at different times , expressed as: Among them, represents the GPU core temperature at time , represents the video memory temperature, represents the power management unit temperature, is the total number of samples. To analyze the historical temperature distribution, calculate the temperature probability density distribution of each component under different load conditions. Assume that the set of workload types of the graphics card is , then classify the historical temperature data according to the load type and construct a temperature distribution function: Among them, represents the temperature probability density under the load type , is the total number of samples under the load type , is an indicator function, which takes 1 if the condition is satisfied and 0 otherwise. By calculating these probability density distributions, the temperature statistical characteristics of each component under different load types are obtained. After obtaining the historical temperature distribution of the components, four-level temperature thresholds are set according to their statistical characteristics. Assuming that the temperature threshold of each component needs to adapt to different load types, four key temperature levels, namely low load, medium load, high load, and critical load, are set through quantile analysis. Let represent the -th quantile of the temperature, then the temperature thresholds of each component are set as follows: , , , . Similarly, the temperature thresholds of the video memory and the power management unit are calculated in the same way: , , where the selection of is adjusted according to system requirements. For example, if the video memory is sensitive to high temperatures, the quantile of the critical threshold is reduced. These temperature thresholds are organized into a component temperature threshold matrix so that the temperature status of each component can be quickly accessed in subsequent calculations. Let the temperature threshold matrix be: Among them, each row corresponds to a component, and each column corresponds to the temperature thresholds of different load levels. After constructing the temperature threshold matrix, the real-time temperature data at the current moment is extracted from the standardized dataset and compared with the temperature threshold matrix to determine the current temperature status value of each component. Let the real-time temperature vector be: Then the temperature status value of each component is calculated by comparing its temperature with the threshold matrix: Among them, represent the temperature status levels of the GPU core, video memory, and power management unit respectively. The numerical range is between 1 and 4, corresponding to low temperature, normal, warning, and high temperature states respectively. The temperature status values of all components are combined into a component temperature status evaluation matrix : This matrix is used to monitor the operating status of the graphics card in real time and provide a basis for power consumption regulation and heat dissipation optimization.

[0029] In a specific embodiment, the process of executing step S200 may specifically include the following steps: Set the total power consumption budget TDP value and the maximum allowable power consumption values of the GPU core, video memory, and power management unit according to the graphics card model to obtain the total power consumption budget constraint; Based on the status values in the component temperature status evaluation matrix, calculate the power consumption reduction coefficient for each component, and apply the power consumption reduction coefficient to the maximum allowable power consumption to obtain the temperature-sensitive power consumption upper limit constraint; According to the workload type identifier, extract the mapping coefficient between component power consumption and performance from the pre-established performance mapping database to construct the performance objective function; Based on the component temperature status evaluation matrix, construct the temperature risk assessment function, and set the weight coefficient according to the workload type identifier to obtain the multi-objective optimization function; Combine the total power consumption budget constraint, the temperature-sensitive power consumption upper limit constraint, the performance objective function, and the multi-objective optimization function into a complete optimization problem, and solve it by the Lagrange multiplier method to obtain the optimal power consumption allocation scheme; Verify the optimal power consumption allocation scheme to obtain the power consumption allocation scheme for each component that satisfies the total power consumption budget constraint.

[0030] Specifically, set the total power consumption budget, that is, the thermal design power TDP value, according to the graphics card model, and determine the maximum allowable power consumption values of the GPU core, video memory, and power management unit to ensure that the power consumption allocation does not exceed the power supply capacity and heat dissipation capacity of the system. Let TDP be , representing the maximum power consumption budget of the entire graphics card. At the same time, let the maximum allowable power consumptions of the GPU core, video memory, and power management unit be , and respectively. Then the basic power consumption constraint is expressed as: Among them, , and respectively represent the actual values of the current power consumption allocated to the GPU core, video memory, and power management unit. Since the temperature of the graphics card changes with the load, calculate the power consumption reduction coefficient based on the status values in the component temperature status evaluation matrix during the power consumption allocation process to dynamically adjust the power consumption upper limit of each component. Let the temperature status evaluation matrix define the temperature status of each component, where , and respectively represent the temperature status levels of the GPU core, video memory, and power management unit, with a value range of 1 to 4, corresponding to low temperature, normal, warning, and high temperature respectively. To control the power consumption in the high-temperature state, a power consumption reduction coefficient is defined : Among them, is the reduction coefficient, which determines the influence degree of the temperature status on the power consumption. Applying the power consumption reduction coefficient to the maximum allowable power consumption, the temperature-sensitive power consumption upper limit is obtained: Thus, it is ensured that the power consumption of each component will not exceed the reasonable range due to excessive temperature. According to the workload type identifier, the mapping coefficient between the component power consumption and performance is extracted from the pre-established performance mapping database, and the performance objective function is constructed. Assume that the performance outputs of the GPU core, video memory, and power management unit are respectively represented as , and , then the total performance is expressed as: Among them, is the weight coefficient of the component, which is dynamically adjusted according to the workload type. While optimizing the power consumption allocation, a temperature risk assessment function is constructed to ensure that the temperature does not exceed the safe range. Let the temperature risk function be: Among them, is the risk adjustment coefficient, and its magnitude depends on the influence of the temperature on the system stability. The component with a higher temperature corresponds to a greater risk weight. Combining the performance objective function and the temperature risk assessment function, and setting the weight coefficient, a multi-objective optimization function is constructed: Among them, and are respectively the balance coefficients of performance optimization and temperature risk, which are adjusted according to the load type. For example, in the case of high-performance tasks, has a larger value, while in the case of high-temperature operating conditions, has a larger value. To solve the optimal power consumption allocation scheme, the overall power consumption budget constraint, the temperature-sensitive power consumption upper limit constraint, the performance objective function, and the multi-objective optimization function are combined into a complete optimization problem, and it is solved by the Lagrange multiplier method. Define the Lagrangian function: By taking the partial derivative and setting it equal to zero, the optimal power consumption allocation scheme is obtained: By taking the partial derivative and setting it equal to zero, the optimal power consumption allocation scheme is obtained: The solution is . It is necessary to verify the optimal power consumption allocation scheme to ensure that it meets the total power consumption budget constraint. If: and all do not exceed the power consumption upper limit sensitive to temperature, then the scheme is valid; otherwise, it is necessary to adjust the optimization parameters and recalculate the optimal solution. This method ensures that the graphics card can achieve the best power consumption and performance balance under different loads, while keeping the temperature within a reasonable range and improving the overall energy efficiency.

[0031] In a specific embodiment, the process of executing step S300 may specifically include the following steps: Construct power consumption-frequency-voltage mapping data tables for the GPU core, video memory, and power management unit respectively, record the corresponding relationship between frequency and voltage of each component under different power consumption conditions, and obtain the component mapping database; Input the power consumption values in the power consumption allocation scheme of each component into the corresponding component mapping database respectively, query the candidate frequency-voltage combinations, and obtain the initial frequency-voltage candidate set; Set the upper and lower bounds of the frequency for the initial frequency-voltage candidate set to obtain the search interval; Perform a binary search within the search interval. First, take the middle frequency point as the test point, calculate the corresponding voltage value, and determine whether the power consumption meets the requirements of the allocation scheme. Narrow the search interval according to the judgment result to obtain the frequency-voltage values that meet the power consumption constraint; Calculate the performance efficiency index for the frequency-voltage values that meet the power consumption constraint, select the combination with the highest performance efficiency as the final configuration, and obtain the optimal frequency-voltage parameters of each component; Package the optimal frequency-voltage parameters and execution time information into the control instruction format to obtain the frequency-voltage control instruction. The frequency-voltage control instruction includes component identification, target frequency, target voltage, and execution time information.

[0032] Specifically, power consumption-frequency-voltage mapping data tables are constructed for the GPU core, video memory, and power management unit respectively to record the corresponding relationship between frequency and voltage under different power consumption conditions. This mapping database is constructed based on experimental measurements and historical operation data, and the working frequency and voltage parameters of each component under specific power consumption are calculated by empirical formulas or fitting models. Let the power consumption-frequency-voltage mapping relationship of the GPU core be , the mapping relationship of the video memory be , and the mapping relationship of the power management unit be , the complete mapping database is represented as: Among them, represents the power consumption of the component, represents the corresponding operating frequency, represents the corresponding operating voltage. Input the power consumption values in the power consumption allocation scheme into this database to query the candidate frequency-voltage combinations that meet the requirements and obtain the initial frequency-voltage candidate set. Assume that the current power consumption allocation scheme is , then by searching for the records in the database that are closest to these power consumption values, the preliminary frequency-voltage pairs are obtained: The initially obtained frequency-voltage candidate set contains multiple possible frequency-voltage combinations. To screen out the optimal frequency-voltage configuration, a search range is set for the candidate set. Assume that the minimum frequency and maximum frequency of the GPU core, video memory, and power management unit are and , respectively, then the search range is represented as: And set the initial search range to ensure that the selected frequency does not exceed the operating range of the component. After determining the search range, perform a binary search to efficiently find the best frequency-voltage combination that meets the power consumption constraint. Take the middle frequency point in the search range as the test point. Assume that the currently searched frequency is , then calculate its corresponding voltage: Among them, is the power-frequency-voltage mapping function obtained by polynomial fitting. Calculate the actual power consumption under this frequency-voltage combination: Among them, is the power consumption calculation model of the component. If exceeds the maximum allowable value in the power consumption allocation scheme, then narrow the search range and reduce the frequency; if is lower than the target value, then increase the frequency to improve the performance. Iterate in this way until the search converges to the optimal frequency-voltage pair that meets the power consumption requirements. After finding the frequency-voltage combination that meets the power consumption constraint, calculate its performance efficiency index to select the configuration with the highest efficiency as the final scheme. Assume that the performance efficiency index is , defined as the computing power per unit power consumption: Among them, is a performance efficiency metric, which is the computing performance per unit power consumption. A larger metric indicates higher efficiency; is the th component's (such as GPU core, video memory, or power management unit) operating frequency during testing; is the th component's corresponding operating voltage during testing. This formula represents the power consumption utilization efficiency of the component under specific frequency and voltage conditions. The performance efficiency metric calculated by this formula is used to guide the selection of frequency-voltage combinations to ensure that the graphics card components operate in the optimal energy efficiency state while meeting the power consumption constraint conditions. Among all frequency-voltage combinations that satisfy the constraints, select the configuration with the highest value as the optimal solution: where, and are the optimal frequency and optimal voltage respectively. Package the optimal frequency-voltage parameters and execution time information into the control instruction format to ensure that the instructions are correctly sent to the hardware for execution. Let the control instruction be , then it contains the component identifier , the target frequency , the target voltage , and the execution timestamp : where, is used to ensure that the frequency-voltage adjustment is executed within a suitable time window to avoid system instability caused by sudden changes.

[0033] In a specific embodiment, the process of executing step S400 may specifically include the following steps: Split the standardized dataset according to a fixed-length time window to construct the component parameter time series data including GPU core, video memory, and power management unit; Input the component parameter time series data into the data preprocessing layer of the adaptive multi-frequency adjustment learning network, perform Z-score standardization and principal component analysis dimensionality reduction on the data to obtain the denoised feature representation; Input the denoised feature representation into the adaptive multi-sampling module of the adaptive multi-frequency adjustment learning network, automatically determine the optimal sampling frequency according to the power consumption fluctuation characteristics of different components, construct multiple power consumption monitoring channels, and obtain the multi-frequency monitoring sequence; Input the multi-frequency monitoring sequence into the multi-channel convolutional layer of the adaptive multi-frequency adjustment learning network, extract the time feature patterns of each component by parallel processing convolutional kernels of different scales, and obtain the component time feature map; Map the component time feature to the self-attention module of the adaptive multi-frequency adjustment learning network, calculate the correlation weight matrix between different components, selectively focus on the key power consumption change areas, and obtain the context features with spatial dependence; Input the context features into the gated recurrent unit of the adaptive multi-frequency adjustment learning network, and predict the predicted values of the component power consumption requirements of the GPU core, video memory, and power management unit for the next m time steps through the temporal memory mechanism.

[0034] Specifically, the standardized dataset contains temperature, frequency, voltage, and power consumption data of multiple components (GPU cores, video memory, and power management units). The standardized dataset is segmented according to a time window of a fixed length. The selection of the time window is based on the system's response speed and the data sampling frequency. Through this step, time series data containing the parameters of GPU cores, video memory, and power management units is obtained. The time series data of component parameters is fed into the data preprocessing layer of the adaptive multi-frequency adjustment learning network. At this layer, Z-score standardization is performed on the data to eliminate the dimensional differences between different components, enabling the power consumption, temperature, frequency, voltage, and other data of each component to be processed on the same scale. By standardizing the data of each component, it is ensured that the mean of all input features is 0 and the standard deviation is 1, thereby improving the training efficiency and stability of the learning network. Principal component analysis dimensionality reduction is performed on the standardized data to map the data from a high-dimensional space to a low-dimensional space, while retaining the main features in the data and removing redundant information, obtaining a denoised feature representation. The dimensionality-reduced data is input into the adaptive multi-sampling module of the adaptive multi-frequency adjustment learning network. This module determines the optimal sampling frequency for each component according to the power consumption fluctuation characteristics of different components. Since the power consumption fluctuation characteristics of different components are different, the power consumption of GPU cores usually fluctuates greatly, so a higher sampling frequency is required to capture the rapidly changing power consumption signal; while the power consumption change of the video memory is relatively stable, and the sampling frequency is appropriately reduced. The power consumption change of the power management unit is affected by multiple factors, so its sampling frequency needs to be considered comprehensively in multiple aspects. The adaptive multi-sampling module will allocate appropriate sampling frequencies for each component based on these characteristics, and then construct multiple power consumption monitoring channels. These monitoring channels can track the power consumption changes of each component in real time and generate multi-frequency monitoring sequences. The multiple frequency monitoring sequences are input into the multi-channel convolutional layer of the adaptive multi-frequency adjustment learning network. The convolutional layer extracts the temporal feature patterns of the component power consumption data by parallel processing convolutional kernels of different scales. Each convolutional kernel extracts features from different time scales. Smaller-scale convolutional kernels are suitable for capturing rapidly changing power consumption patterns, while larger-scale convolutional kernels extract trends over longer time periods. Through the parallel processing of the multi-channel convolutional layer, the power consumption fluctuations of components are analyzed from multiple time scales simultaneously, generating a component temporal feature map that provides the power consumption change patterns of each component within different time windows. The component temporal feature map is input into the self-attention module of the adaptive multi-frequency adjustment learning network. The role of the self-attention module is to calculate the correlation weight matrix between different components. Through these weight matrices, the system selectively focuses on the components that have a greater impact on power consumption changes. Through the self-attention mechanism, effective associations are established between multiple components, and key power consumption change regions are selectively focused on, obtaining context features with spatial dependence. The context features are input into the gated recurrent unit of the adaptive multi-frequency adjustment learning network.The gated recurrent unit is a type of recurrent neural network that, by introducing a gating mechanism, effectively processes long-term temporal dependencies and avoids the problem of vanishing gradients. At this layer, the power consumption requirements of each component within the next m time steps are predicted through a temporal memory mechanism. This process is based on the current power consumption data and temporal dependencies, that is, the system predicts future requirements based on historical power consumption changes. For example, by observing the power consumption fluctuation pattern of the GPU core and the usage of video memory, the power consumption requirements of the GPU core and video memory within the next few steps are predicted, so as to adjust the power consumption allocation in advance to ensure the stability and efficiency of power management.

[0035] In a specific embodiment, the process of executing step S500 may specifically include the following steps: Input the frequency-voltage control instruction and the predicted value of component power consumption requirements into the multi-level adjustment control framework, and through the hierarchical processing of the real-time monitoring layer, policy decision layer, and execution control layer, obtain the initial adjustment control strategy; Perform priority sorting on the initial adjustment control strategy, classify the temperature state abnormal response, power consumption overlimit response, and predictive adjustment operation according to the priority level, and obtain the hierarchical adjustment instruction set; Based on the current workload characteristics and component temperature state, calculate the adjustment period to obtain the adaptive adjustment time interval; According to the pre-established power consumption impact matrix between components, analyze the cross-impact of power consumption adjustment of each component, and when any component needs power consumption limitation, cooperate to adjust the power consumption allocation of related components to obtain the power consumption control instruction optimized through cooperation; Send the power consumption control instruction optimized through cooperation to the controllers of the GPU core, video memory, and power management unit through the graphics card driver layer interface, and perform frequency adjustment, voltage setting, and power consumption limitation operations to obtain the actual operating state of the components; Monitor the performance indicators of the actual operating state of the components to obtain the feedback result of power consumption dynamic adjustment.

[0036] Specifically, a multi-level adjustment and control framework is designed. This framework consists of three levels: the real-time monitoring layer, the policy decision-making layer, and the execution control layer. The real-time monitoring layer collects data from various components of the graphics card, including the temperature, voltage, frequency of the GPU core, the usage of video memory, and the power consumption of the power management unit, etc. Through these data, the real-time monitoring layer effectively detects the real-time status of each component of the graphics card, and timely discovers potential problems such as abnormal temperature and excessive power consumption. The real-time data, along with the predicted power consumption values of the components and the frequency-voltage control instructions, are passed to the policy decision-making layer. In the policy decision-making layer, the current state of the graphics card is evaluated, and based on the input frequency-voltage control instructions and the predicted power consumption values, a preliminary adjustment and control strategy is generated. This strategy decides which components' frequencies and voltages need to be adjusted based on the current temperature state, power consumption distribution, and predicted load changes. For example, if the temperature of the GPU core is approaching the threshold, the policy decision-making layer will recommend reducing its frequency to reduce power consumption, or making appropriate power consumption limits according to the usage of video memory and the power management unit. After the control strategies are initially generated, these strategies are prioritized to ensure that the most important adjustment tasks can be processed first. Specifically, the abnormal temperature state response is regarded as the highest-priority task because too high temperature can cause hardware damage or performance degradation and must be dealt with immediately. Secondly, the excessive power consumption response is a task that needs to be responded to in a timely manner because exceeding the power consumption limit will not only trigger overheat protection but also affect the overall system stability. The predictive adjustment operation has a relatively low priority. This type of operation is used to adjust power consumption in advance to cope with upcoming workload changes and avoid sudden increases in power consumption or sudden rises in temperature. By prioritizing these tasks, precise power consumption adjustment is carried out at the most needed moment to ensure that the graphics card can maintain the best performance and stability at any time. Calculate the adjustment period based on the current workload characteristics and the component temperature state. The calculation of the adjustment period depends on the workload of the graphics card and the temperature conditions of each component. Workload characteristics include different tasks such as graphics rendering, AI inference, video encoding, etc. These tasks have very different power consumption requirements, so the adjustment period is adjusted according to the task type. The temperature state is also an important factor determining the length of the adjustment period. When the temperature is low and the system load is not high, the adjustment period is extended to reduce the computational overhead brought by adjustment operations; while when the temperature is high or the load is heavy, the adjustment period is shortened to perform power consumption adjustment more frequently to ensure that the graphics card will not malfunction due to excessive load or overheating. Based on these factors, an adaptive adjustment time interval is calculated, and the adjustment period is adjusted accordingly to achieve the best power management. When dealing with power consumption adjustment, consider the cross-influence between components, especially between the GPU core, video memory, and power management unit. Since these components are interrelated, a change in the power consumption of one component often affects the power consumption requirements of other components. For example, an increase in the power consumption of the GPU core will lead to an increase in the video memory bandwidth requirement, and the burden on the power management unit will also increase accordingly.When the system performs power consumption limitation, it considers the interaction effects among these components. To this end, a power consumption impact matrix among components is established, which records the power consumption dependency relationships among different components. When the power consumption of a component exceeds the limit or approaches the maximum power consumption, this matrix is used to coordinately adjust the power consumption allocation of other relevant components. In this way, on the premise of ensuring that the overall power consumption of the graphics card does not exceed the standard, the balanced adjustment of each component is realized, and the overall efficiency of the system is improved. After the collaborative optimization of the power consumption adjustment control instructions is completed, these instructions are sent to the controllers of the GPU core, video memory, and power management unit through the graphics card driver layer interface, instructing them to perform operations such as frequency adjustment, voltage setting, and power consumption limitation. The graphics card driver layer is an important bridge connecting the hardware and the control system, which can ensure that the adjustment instructions are accurately transmitted to the hardware and there is no delay or error during execution. When the controller receives the adjustment instructions, it performs the actual frequency adjustment, voltage setting, and power consumption limitation operations. The GPU core, video memory, and power management unit will adjust their working states according to the new control instructions, so as to achieve the expected power consumption adjustment target. After the adjustment operation is completed, the actual operating states of each component are monitored, and relevant performance index data are collected. These data include the temperature change of the GPU core, the bandwidth utilization rate of the video memory, the power consumption of the power management unit, etc. By monitoring these performance indexes, the effect of the adjustment operation is evaluated. For example, if the temperature of the GPU core remains within the safe range and the performance indexes do not decrease, it indicates that the adjustment strategy is effective; if the temperature is still too high or the performance drops too much, the adjustment strategy needs to be re-evaluated and optimized. The feedback result of the power consumption dynamic adjustment will be provided to the system for guiding subsequent adjustment operations.

[0037] Please refer to Figure 2 , Figure 2 which is a schematic structural block diagram of the graphics card component power consumption dynamic adjustment device 200 provided by an embodiment of the present application. As Figure 2 shown, the graphics card component power consumption dynamic adjustment device 200 includes: A feature analysis module 210, configured to perform multi-channel data acquisition and workload feature analysis on the operating parameters of the GPU core, video memory, and power management unit of the graphics card, and obtain a component temperature state evaluation matrix; A power consumption allocation module 220, configured to perform component-level power consumption allocation based on multi-objective optimization according to the component temperature state evaluation matrix, and obtain a power consumption allocation scheme for each component; A calculation module 230, configured to input the power consumption allocation scheme for each component into a power consumption-frequency-voltage mapping model, and calculate optimal frequency-voltage configuration parameters through a binary search algorithm to obtain frequency-voltage control instructions; A multi-step prediction module 240, configured to input a standardized data set into an adaptive multiple frequency adjustment learning network for multi-step prediction, and obtain a predicted value of the component power consumption demand; The dynamic adjustment module 250 is configured to perform dynamic adjustment of the power consumption of the graphics card components according to the frequency-voltage control instruction and the predicted value of the component power consumption requirement, and obtain a feedback result of the dynamic power consumption adjustment.

[0038] Through the collaborative cooperation of the above-mentioned various components, the present invention realizes the accurate monitoring of the operating parameters at the component level by setting a dedicated sensor network in the GPU core, video memory and power management unit, and combines multi-dimensional feature extraction technologies such as Fourier transform, wavelet transform and autocorrelation analysis to accurately identify the type of workload, providing a precise data basis for subsequent power consumption adjustment, and effectively solving the problems of rough data collection and inaccurate workload feature extraction in traditional methods. The multi-level temperature threshold mechanism set according to the characteristics of different components realizes the refined hierarchical management of the temperature state, and by dynamically adjusting the threshold to respond to different workload characteristics, it avoids the over-conservative or over-aggressive problems of the traditional fixed threshold scheme in different application scenarios, and improves the accuracy and flexibility of temperature control. The present invention establishes a multi-objective optimization function that simultaneously considers performance maximization and temperature risk control, and combines the Lagrange multiplier method to solve the optimal power consumption allocation scheme. On the premise of ensuring the overall power consumption budget constraint, it realizes the reasonable allocation of component-level power consumption resources, effectively balancing the relationship between performance output and temperature control. By constructing a power consumption-frequency-voltage mapping model and applying the binary search algorithm, the present invention can find the frequency-voltage combination with the highest performance efficiency under the condition of meeting the power consumption constraint, realizing precise power consumption control, avoiding the blindness and inefficiency of frequency-voltage regulation in traditional methods, and improving the energy utilization efficiency. The adaptive multi-frequency regulation learning network is used to perform multi-step prediction of the power consumption demand, and the spatio-temporal feature correlation is captured through multi-channel convolution and self-attention mechanism, realizing the accurate prediction of the future power consumption demand, overcoming the limitation of traditional methods that only rely on historical data for simple prediction, and providing a reliable guarantee for predictive power consumption regulation. The present invention integrates the dual mechanisms of reactive regulation and predictive regulation, and through the multi-level regulation control framework and priority sorting strategy, while dealing with temperature anomalies, it prepares in advance for workload changes, realizing the combination of timeliness and forward-looking of power consumption regulation, and improving the stability and response speed of the system. By analyzing the mutual influence between components through the power consumption influence matrix between components, the state changes of other components are considered collaboratively when adjusting the power consumption of a certain component, avoiding the resource imbalance problem caused by independent regulation between components in traditional methods, and improving the coordination and efficiency of the overall system.

[0039] Please refer to Figure 3 , Figure 3Schematic block diagram of the graphics card component power consumption dynamic adjustment device 300 provided by the embodiments of the present application. The graphics card component power consumption dynamic adjustment device 300 includes a processor 301 and a memory 302. The processor 301 and the memory 302 are connected through a device bus 303. Among them, the memory 302 may include a non-volatile storage medium and an internal memory.

[0040] The non-volatile storage medium can store a computer program. The computer program includes program instructions. When the program instructions are executed by the processor 301, the processor 301 can be made to execute any of the above-mentioned graphics card component power consumption dynamic adjustment methods.

[0041] The processor 301 is used to provide computing and control capabilities to support the operation of the entire graphics card component power consumption dynamic adjustment device 300.

[0042] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor 301, the processor 301 can be made to execute any of the above-mentioned graphics card component power consumption dynamic adjustment methods.

[0043] Those skilled in the art can understand that Figure 3 the structure shown in is only a block diagram of some structures related to the solution of the present application, and does not constitute a limitation on the graphics card component power consumption dynamic adjustment device 300 involved in the solution of the present application. The specific graphics card component power consumption dynamic adjustment device 300 may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0044] It should be understood that the processor 301 may be a central processing unit (CPU), and the processor 301 may also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among them, the general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0045] It should be noted that those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working process of the above-described graphics card component power consumption dynamic adjustment device 300 can refer to the corresponding process of the foregoing graphics card component power consumption dynamic adjustment method, and will not be elaborated here.

[0046] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the systems, systems, and units described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein.

[0047] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The foregoing storage medium includes: various media that can store program codes, such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs.

[0048] As described above, the above embodiments are only used to illustrate the technical solutions of this application, rather than to limit them; although this application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of this application.

Claims

1. A method for dynamically adjusting power consumption of a graphics card component, characterized in that: include: Perform multi-channel data collection and workload feature analysis on the operating parameters of the graphics card GPU core, video memory, and power management unit to obtain a component temperature status assessment matrix; According to the component temperature state evaluation matrix, performing component-level power consumption allocation based on multi-objective optimization to obtain a power consumption allocation scheme for each component; Input the power consumption allocation scheme of each component into the power consumption-frequency-voltage mapping model, calculate the optimal frequency and voltage configuration parameters through a binary search algorithm, and obtain a frequency and voltage control instruction; The standardized data set is input into the adaptive multiple frequency regulation learning network for multi-step prediction to obtain the component power consumption demand prediction value; According to the frequency and voltage control instructions and the predicted value of the component power consumption demand, the graphics card component power consumption is dynamically adjusted to obtain a dynamic power consumption adjustment feedback result.

2. The method for dynamically adjusting the power consumption of a graphics card component according to claim 1, characterized in that: The multi-channel data collection and workload feature analysis of the operating parameters of the graphics card GPU core, video memory and power management unit are performed to obtain a component temperature state evaluation matrix, including: Set a temperature sensor in the GPU core of the graphics card, a voltage sensor in the video memory, and a current sensor in the power management unit, and collect the original operating parameter data of the components; De-noising, normalizing and outlier filtering are performed on the original operating parameter data of the component to obtain a standardized data set including a timestamp and a component parameter vector; Performing Fourier transform, wavelet transform and autocorrelation analysis on the standardized data set to extract frequency domain features, nonlinear features and time dependency to obtain a workload feature matrix; Input the workload feature matrix into a pre-trained decision tree classifier, classify by comparing historical data patterns, and obtain a workload type identifier; Based on the standardized data set and the workload type identifier, multi-level temperature thresholds are set for the GPU core, the video memory, and the power management unit, respectively, to obtain a component temperature state evaluation matrix.

3. The method for dynamically adjusting the power consumption of a graphics card component according to claim 2, characterized in that: The method of setting a multi-level temperature threshold for each of the GPU core, the video memory, and the power management unit based on the standardized data set and the workload type identifier to obtain a component temperature state evaluation matrix includes: Extracting historical temperature data of the GPU core, the video memory, and the power management unit from the standardized data set to obtain component historical temperature distribution; According to the component historical temperature distribution and the workload type identifier, four temperature thresholds are set for the GPU core, four temperature thresholds are set for the video memory, and four temperature thresholds are set for the power management unit to obtain a component temperature threshold set; organizing the set of component temperature thresholds into a component temperature threshold matrix; Extracting real-time temperature data of each component from the standardized data set, and comparing it with the corresponding threshold in the component temperature threshold matrix to obtain the temperature state value of each component; A component temperature state evaluation matrix is ​​constructed according to the temperature state values ​​of the components.

4. The method for dynamically adjusting the power consumption of a graphics card component according to claim 3, characterized in that: The method of performing component-level power consumption allocation based on multi-objective optimization according to the component temperature state evaluation matrix to obtain a power consumption allocation scheme for each component includes: According to the graphics card model, the overall power consumption budget TDP value and the maximum allowable power consumption values ​​of the GPU core, video memory, and power management unit are set to obtain the overall power consumption budget constraint; Calculating a power consumption reduction factor of each component based on the state value in the component temperature state evaluation matrix, and applying the power consumption reduction factor to the maximum allowable power consumption to obtain a temperature-sensitive power consumption upper limit constraint; According to the workload type identifier, extracting mapping coefficients of component power consumption and performance from a pre-established performance mapping database to construct a performance objective function; Based on the component temperature state assessment matrix, a temperature risk assessment function is constructed, and a weight coefficient is set according to the workload type identifier to obtain a multi-objective optimization function; The overall power consumption budget constraint, the temperature-sensitive power consumption upper limit constraint, the performance objective function and the multi-objective optimization function are combined into a complete optimization problem, and solved by Lagrange multiplier method to obtain an optimal power consumption allocation scheme; The optimal power consumption allocation scheme is verified to obtain a power consumption allocation scheme for each component that meets the total power consumption budget constraint.

5. The method for dynamically adjusting the power consumption of a graphics card component according to claim 4, characterized in that: The step of inputting the power consumption allocation scheme of each component into a power consumption-frequency-voltage mapping model, calculating the optimal frequency and voltage configuration parameters by a binary search algorithm, and obtaining a frequency and voltage control instruction includes: Construct power consumption-frequency-voltage mapping data tables for GPU core, video memory and power management unit respectively, record the corresponding relationship between frequency and voltage of each component under different power consumption conditions, and obtain component mapping database; Input the power consumption values ​​in the power consumption allocation scheme of each component into the corresponding component mapping database, query the candidate frequency-voltage combination, and obtain the initial frequency-voltage candidate set; Setting upper and lower frequency limits for the initial frequency and voltage candidate set to obtain a search interval; Perform a binary search in the search interval, first take the middle frequency point as the test point, calculate the corresponding voltage value, and judge whether the power consumption meets the allocation plan requirements, narrow the search interval according to the judgment result, and obtain the frequency and voltage value that meets the power consumption constraint; Calculate the performance efficiency index for the frequency and voltage values ​​that meet the power consumption constraints, select the combination with the highest performance efficiency as the final configuration, and obtain the optimal frequency and voltage parameters of each component; The optimal frequency and voltage parameters and the execution time information are packaged into a control instruction format to obtain a frequency and voltage control instruction, wherein the frequency and voltage control instruction includes a component identifier, a target frequency, a target voltage and execution time information.

6. The method for dynamically adjusting the power consumption of a graphics card component according to claim 5, characterized in that: The step of inputting the standardized data set into the adaptive multiple frequency adjustment learning network for multi-step prediction to obtain the component power consumption demand prediction value includes: The standardized data set is divided into fixed-length time windows to construct component parameter timing data including GPU core, video memory and power management unit; The component parameter time series data is passed into the data preprocessing layer of the adaptive multiple frequency modulation learning network, and Z-score standardization and principal component analysis dimension reduction are performed on the data to obtain a feature representation after noise reduction; Input the de-noised feature representation into the adaptive multi-sampling module of the adaptive multi-frequency adjustment learning network, automatically determine the optimal sampling frequency according to the power consumption fluctuation characteristics of different components, construct multiple power consumption monitoring channels, and obtain a multi-frequency monitoring sequence; Inputting the multiple frequency monitoring sequence into the multi-channel convolution layer of the adaptive multiple frequency adjustment learning network, extracting the time feature pattern of each component by parallel processing convolution kernels of different scales, and obtaining a component time feature map; Inputting the component time feature map into the self-attention module of the adaptive multiple frequency regulation learning network, calculating the association weight matrix between different components, selectively focusing on the key power consumption change area, and obtaining context features with spatial dependence; The context features are input into the gated recurrent unit of the adaptive multiple frequency regulation learning network, and the component power consumption demand prediction values ​​of the GPU core, video memory and power management unit in the next m time steps are predicted through the temporal memory mechanism.

7. The method for dynamically adjusting the power consumption of a graphics card component according to claim 6, characterized in that: The step of performing dynamic power consumption adjustment of the graphics card component according to the frequency and voltage control instruction and the predicted value of the power consumption requirement of the component to obtain a dynamic power consumption adjustment feedback result includes: Input the frequency and voltage control instructions and the predicted value of the power consumption demand of the component into a multi-level regulation control framework, and obtain an initial regulation control strategy through hierarchical processing of a real-time monitoring layer, a strategy decision layer, and an execution control layer; Prioritizing the initial adjustment control strategies, grading the temperature state abnormality response, power consumption overlimit response and predictive adjustment operation according to the priority, and obtaining a graded adjustment instruction set; Based on the current workload characteristics and component temperature status, the adjustment period is calculated to obtain an adaptive adjustment time interval; According to the pre-established power consumption impact matrix between components, the cross-impact of power consumption regulation of each component is analyzed. When any component needs power consumption restriction, the power consumption allocation of related components is adjusted in a coordinated manner to obtain a coordinated optimized power consumption control instruction. The collaboratively optimized power consumption control instruction is sent to the controllers of the GPU core, video memory and power management unit through the graphics card driver layer interface to perform frequency adjustment, voltage setting and power consumption limit operations to obtain the actual operating status of the components; The actual operating status of the components is monitored for performance indicators to obtain dynamic power consumption adjustment feedback results.

8. A device for dynamically adjusting power consumption of a graphics card component, characterized in that: Used to execute the method for dynamically adjusting the power consumption of a graphics card component according to any one of claims 1 to 7, the device for dynamically adjusting the power consumption of a graphics card component comprising: The feature analysis module is used to perform multi-channel data collection and workload feature analysis on the operating parameters of the graphics card GPU core, video memory and power management unit to obtain the component temperature status evaluation matrix; A power consumption allocation module, used to perform component-level power consumption allocation based on multi-objective optimization according to the component temperature state evaluation matrix, and obtain a power consumption allocation plan for each component; A calculation module, used to input the power consumption allocation scheme of each component into a power consumption-frequency-voltage mapping model, calculate the optimal frequency and voltage configuration parameters through a binary search algorithm, and obtain a frequency and voltage control instruction; A multi-step prediction module is used to input the standardized data set into the adaptive multiple frequency adjustment learning network for multi-step prediction to obtain the component power consumption demand prediction value; The dynamic adjustment module is used to perform dynamic adjustment of the power consumption of the graphics card component according to the frequency and voltage control instructions and the predicted value of the power consumption demand of the component, and obtain the dynamic adjustment feedback result of the power consumption.

9. A device for dynamically adjusting power consumption of a graphics card component, characterized in that: The graphics card component power consumption dynamic adjustment device comprises: a memory and at least one processor, wherein the memory stores instructions; The at least one processor calls the instruction in the memory so that the graphics card component power consumption dynamic adjustment device executes the graphics card component power consumption dynamic adjustment method according to any one of claims 1-7.

Images