CPU idle power distribution system for PC based on AI

Abstract

The invention relates to the technical field of electric power distribution, and provides an AI-based CPU idle electric power distribution system for a PC, which can flexibly optimize CPU electric power distribution according to real-time data of different scenes by virtue of an AI algorithm, meets diversified workload requirements, has the advantages of energy conservation, performance and adaptability, brings more efficient and stable use experience for PC users, and has better application prospects in the aspect of energy conservation. The system can accurately sense load changes and adjust the power supply voltage and frequency of a CPU in real time, in office, game and video editing scenes, power consumption is reduced compared with a traditional system, electric power waste is effectively reduced, the use cost is reduced, in terms of performance, the system greatly improves user experience, software and a browser run smoothly during office, the frame rate is stable during game playing, and the service life of the system is prolonged. Video editing and rendering are fast, and preview is not blocked.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution, and in particular to an AI-based PC CPU idle power distribution system. Background Art

[0002] In today's digital age, PCs have become essential tools for work, entertainment, and creativity. The power distribution efficiency of their CPUs plays a key role in overall performance and energy consumption. Traditional PC CPU power distribution systems often use a fixed distribution model, which presents significant shortcomings.

[0003] On the one hand, as the diversity and complexity of programs running on PCs increase, the CPU resource requirements of different programs vary greatly, and fixed power allocation methods cannot adapt to these changes. For example, in office scenarios, when running simple office software and browsers, the actual CPU utilization rate is low, but traditional systems still maintain high power supply voltage and frequency, resulting in a large amount of power waste. This not only increases energy costs but can also cause CPU temperatures to overheat, shortening the hardware lifespan.

[0004] On the other hand, in scenarios like gaming and video editing, which demand high CPU performance and experience frequent load fluctuations, traditional systems cannot dynamically adjust power based on real-time load. When switching between complex scenes in games, fixed power allocations can't improve performance in a timely manner, leading to frame rate fluctuations and dropped frames. In video editing, rendering tasks vary in intensity, but traditional systems can't flexibly adapt, resulting in extended rendering times and preview lags, severely impacting user experience and work efficiency.

[0005] Therefore, in order to solve the problems of low power distribution efficiency, resource waste and inability to meet diverse load requirements in traditional systems, an AI-based PC CPU idle power distribution system is proposed to achieve more efficient and accurate power distribution and improve PC performance. Summary of the Invention

[0006] Technical problem to be solved: In view of the shortcomings of the existing technology, the present invention provides an AI-based PC CPU idle power distribution system.

[0007] Technical Solution: To achieve the above-mentioned solution, the present invention provides the following technical solution: an AI-based PC CPU idle power distribution system, comprising the following modules: S1. Data acquisition module: uses motherboard sensors, operating system performance monitoring tools and process monitoring interfaces to collect real-time hardware status data and resource demand information of currently running programs; S2, algorithm processing module: S2.1, Multimodal data fusion submodule: Input CPU temperature data into CNN, according to the convolution operation formula Extract features and get , the time series data of CPU usage and frequency are input into LSTM, and processed by LSTM to obtain , and finally merge the two into ;in, is the feature extracted by CNN for temperature data, The features extracted by LSTM from the time series data of CPU usage and frequency are: for and The fusion characteristics of S2.2, attention mechanism submodule: In the LSTM prediction model and decision model, according to the formula ,in, Calculate the attention weight and perform weighted processing on the data; among them, is the LSTM hidden state sequence The hidden states, For the current moment The hidden state of is the LSTM hidden state sequence, ; S2.3, Adaptive reinforcement learning submodule: The adaptive reinforcement learning module uses DQN as the framework and and action space , calculate the Q value , where the state space The action space is composed of the current CPU hardware status and program resource requirements. is a combination of different supply voltages and frequencies, based on the dynamic reward function Evaluate the benefits of different actions, here and is the weight that is dynamically adjusted according to the system status. Indicates the energy efficiency after taking action in the current state, Indicates performance loss; are the network parameters of the deep Q network; S3, power control module: Based on the decision results of the algorithm processing module, the CPU power supply voltage and frequency are adjusted through the motherboard BIOS or hardware control chip to achieve accurate distribution of idle power. At the same time, the feedback data after power adjustment is sent back to the data acquisition module to form a closed-loop control.

[0008] Preferably, in the multimodal data fusion submodule, the temperature data is , input convolutional neural network to extract features, the CNN convolution operation formula is: ,in It is Convolutional layer The output of the position, It is Layer convolution kernel weights, It is layer Position input, It is Layer bias, and Is the size of the convolution kernel, and the temperature data feature representation is obtained through the convolution operation .

[0009] Preferably, in the multimodal data fusion submodule, the time series composed of CPU usage and frequency is , input long short-term memory network, LSTM core calculation formula: Input Gate: ; Forget Gate: ; Output gate: ; Memory unit: ; Hidden state: ; in is the sigmoid function, represents element-wise multiplication, is the weight matrix, is the bias vector, is the current input, is the hidden state at the previous moment, It is the memory unit of the previous moment, and the time series data feature representation is obtained through LSTM .

[0010] Preferably, in the adaptive reinforcement learning submodule, the Q value is updated by the Bellman equation: ,in is the learning rate, It’s an instant reward. is the discount factor, S' is the next state, and a' is the next action.

[0011] Preferably, at the initial stage of operation, the system uses historical operation data to perform initial training on the LSTM prediction model and the adaptive reinforcement learning model to determine the initial parameters. During the operation of the system, incremental training is performed every half an hour using newly collected real-time data.

[0012] Preferably, during the operation of the system, a transfer learning method is adopted to retain the common features that the model has learned and only update the parameters related to the new data.

[0013] Beneficial effects: Compared with the prior art, the present invention provides an AI-based PC CPU idle power distribution system with the following beneficial effects: 1. This AI-based PC CPU idle power distribution system, with its AI algorithm, can flexibly optimize CPU power distribution based on real-time data from different scenarios to meet diverse workload requirements. Its advantages in energy saving, performance, and adaptability bring PC users a more efficient and stable user experience.

[0014] 2. This AI-based PC CPU idle power distribution system can accurately sense load changes and adjust the CPU power supply voltage and frequency in real time in terms of energy saving. In office, gaming, and video editing scenarios, power consumption is lower than that of traditional systems, effectively reducing power waste and lowering usage costs. In terms of performance, the system greatly improves the user experience. Software and browsers run smoothly during office work, the frame rate is stable and there is no frame drop during gaming, and video editing renders quickly and previews without lag. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 Schematic diagram of the system framework of the present invention; Figure 2 Schematic diagram of the system operation flow of the present invention. DETAILED DESCRIPTION

[0016] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0017] See also Figures 1 and 2 The present invention proposes an AI-based PC CPU idle power distribution system, the contents of which are as follows: 1. System Module S1, data acquisition module The data acquisition module is the cornerstone of the entire system's operation, collecting critical data through multiple channels. Motherboard sensors accurately capture low-level hardware status data, such as CPU temperature and voltage. Based on physical principles, they directly perceive the real-time hardware status and offer extremely high reliability. Operating system performance monitoring tools, operating at a macro level, collect information such as CPU usage and operating frequency. This data reflects the overall CPU status under system process scheduling. The process monitoring interface, penetrating the program level, collects information on each running program's CPU resource requirements, such as the number of threads and the complexity of the computational task.

[0018] The collected data suffers from formatting discrepancies and noise, necessitating preliminary organization. During this organization process, data cleaning algorithms are used to remove outliers, such as unreasonable data where CPU usage momentarily exceeds 100%. Furthermore, timestamps and data source identifiers are added to the data. The timestamp accurately records the moment of data collection, facilitating subsequent analysis of data trends over time. The data source identifier clearly identifies the source, whether it is an onboard sensor, operating system tool, or process interface, facilitating traceability and classification, ensuring efficient operation of subsequent modules.

[0019] S2, algorithm processing module: S2.1, Multimodal data fusion submodule: In the multimodal data fusion submodule, specific models are used to process different types of data characteristics. The temperature data is input into the convolutional neural network (CNN). Assume that the temperature data is , input into the convolutional neural network (CNN) to extract features. The CNN convolution operation formula is: ,in It is Convolutional layer The output of the position, It is Layer convolution kernel weights, It is layer Position input, It is Layer bias, and Is the convolution kernel size. After the convolution operation, the temperature data feature representation is obtained .

[0020] For time series data consisting of CPU usage and frequency, the Long Short-Term Memory (LSTM) network comes into play. LSTM has the ability to process long-term and short-term dependencies in time series data, and its core calculation formula consists of multiple parts. Suppose the time series consisting of CPU usage and frequency is , input the long short-term memory network (LSTM). The core calculation formula of LSTM is: Input Gate: ; Forget Gate: ; Output gate: ; Memory unit: ; Hidden state: ; in is the sigmoid function, represents element-wise multiplication, is the weight matrix, is the bias vector, is the current input, is the hidden state at the previous moment, It is the memory unit of the previous moment, and the time series data feature representation is obtained through LSTM The input gate formula controls the extent to which the current input information enters the memory unit through parameters such as the sigmoid function and the weight matrix; the forget gate formula determines whether to retain or discard the information of the previous memory unit; the output gate formula determines the output content based on the current memory unit and input information; the memory unit formula combines the results of the forget gate and the input gate to update the memory content; and the hidden state formula generates the final output based on the memory unit and the output gate.

[0021] Finally and Fusion, get fusion features .

[0022] S2.2, attention mechanism submodule: Introduce the attention mechanism into the LSTM prediction model. Assume that the LSTM hidden state sequence is , the current time is , calculate the attention weight ,in , 、 、 are learnable parameters. These parameters are continuously optimized during the training process, allowing the model to automatically focus on key data. , which is used for subsequent predictions. When predicting future CPU load changes, the model can focus on the hidden state information during the recent period of large load fluctuations to improve prediction accuracy.

[0023] In the decision model, the fusion features Calculate attention weights to determine the importance of different features to decision-making. Because different features have varying degrees of importance in different operating scenarios, calculating attention weights can clarify the importance of each feature and adjust decision-making priorities based on actual conditions, making decisions more aligned with the system's real-time needs and improving decision accuracy and adaptability.

[0024] S2.3, Adaptive reinforcement learning submodule: The adaptive reinforcement learning submodule is built on a deep Q-network (DQN) framework. The state space, composed of the current CPU hardware status (such as temperature, usage, and frequency) and program resource requirements (number of processes, resource priority, etc.), comprehensively describes the system's current operating state. The action space encompasses different power supply voltage and frequency combinations, representing the power allocation actions the system can take.

[0025] Based on the Deep Q Network (DQN) framework. State space The action space is composed of the current CPU hardware status and program resource requirements. is a combination of different supply voltages and frequencies. The Q value function is defined as ,in are network parameters.

[0026] Dynamic reward function Adjust according to the real-time operating status of the system, such as ,in and is the weight that is dynamically adjusted according to the system status. Indicates the energy efficiency after taking action in the current state, Indicates performance loss.

[0027] Update the Q value through the Bellman equation: ,in The learning rate controls the step size of each update to avoid updating too quickly and missing the optimal solution, or updating too slowly and affecting training efficiency. It’s an instant reward. is a discount factor reflecting the immediate benefit after taking the current action; S' is the next state, and a' is the next action. As the system continuously interacts with the environment, it adjusts its decisions based on reward feedback. Through multiple iterative learning cycles, it gradually finds the optimal power allocation strategy for the current state, achieving a balance between energy efficiency and system performance.

[0028] S3, power control module The power control module is the execution unit of system decisions. It regulates CPU power supply based on the decisions made by the algorithm processing module. This module adjusts the CPU supply voltage and frequency through the motherboard BIOS or hardware control chip. When the algorithm determines that the CPU load is low, the power control module reduces the supply voltage and frequency to reduce energy consumption. When the CPU load is high, the power control module increases the supply voltage and frequency to ensure system performance.

[0029] After adjusting power parameters, the power control module transmits feedback data, including actual voltage and frequency values, as well as CPU status changes (such as temperature fluctuations and task completion status), back to the data acquisition module, forming a closed-loop control. This closed-loop control enables the system to make timely optimization decisions based on the actual adjustment results. For example, if CPU temperatures become too high or task execution becomes stagnant after reducing voltage and frequency, the system will appropriately increase power supply parameters based on this feedback data to ensure stable system operation.

[0030] 2. Operation Process S1. Data collection stage The data acquisition module continuously collects data at carefully set intervals. Considering the frequent changes in CPU usage, data is collected every 500 milliseconds to capture its dynamic changes. Temperature and frequency change relatively slowly, so data is collected every 1 second. This ensures real-time performance while reducing unnecessary data collection overhead.

[0031] After initial organization and labeling, the collected raw data is transferred to the algorithm processing module. This initial organization includes data cleaning and normalization. Cleaning removes outliers, while normalization maps data from different ranges to a unified interval to facilitate subsequent model processing. Adding timestamps and data source identifiers clarifies the data's time series characteristics and source information, providing a higher-quality data foundation for algorithm processing.

[0032] S2, algorithm processing stage S2.1 Multimodal Data Fusion After the algorithm processing module receives the data, the multimodal data fusion module starts working. The temperature data is input into CNN and the convolution operation formula is used. Extract features and get ; Input the time series data of CPU usage and frequency into LSTM and obtain , and finally merge the two into The goal is to fully tap into the potential connections between different types of data, allowing subsequent models to obtain more comprehensive information and provide richer data support for decision-making. This fusion method fully taps into the potential connections between different types of data. For example, rising temperatures may be related to high CPU load (high usage and high frequency). The fused data can more comprehensively present the CPU's operating status, providing a rich basis for subsequent decision-making.

[0033] S2.2 Application of attention mechanism: In the LSTM prediction model and decision model, the attention mechanism module is based on the formula ( ) calculates attention weights and applies weighted processing to the data. When processing time series data, the attention mechanism allows the model to automatically focus on historical data points that have a significant impact on the current decision, such as periods of recent significant load fluctuations or data highly relevant to the current program's operating status, rather than treating all historical data equally. This improves the decision model's ability to capture key information. In the decision model, the attention mechanism dynamically adjusts the weights of different features based on different operating scenarios, improving decision accuracy and adaptability.

[0034] S2.3. Adaptive Reinforcement Learning Decision The adaptive reinforcement learning module uses DQN as the framework and and action space , calculate the Q value . The state space The action space is composed of the current CPU hardware status and program resource requirements. is a combination of different power supply voltages and frequencies. According to the dynamic reward function Evaluate the benefits of different actions, here and is the weight that is dynamically adjusted according to the system status. Indicates the energy efficiency after taking action in the current state, Represents performance loss. The system continuously interacts with the environment and adjusts its decision based on reward feedback. For example, when the energy efficiency of the system is improved after taking a certain action and the performance loss is within an acceptable range, the Q value corresponding to the action will be increased. And through the Bellman equation Update the Q value, where is the learning rate, It’s an instant reward. is the discount factor, S' is the next state, and a' is the next action. After multiple iterations of learning, the system gradually finds the optimal power allocation strategy, achieving a balance between energy and performance.

[0035] S3, power control stage Based on the power allocation strategy determined by the algorithm processing module, the power control module adjusts the CPU's supply voltage and frequency through the motherboard BIOS or hardware control chip. For example, when the algorithm determines the CPU load is low, the supply voltage and frequency are reduced; when the load is high, the supply voltage and frequency are increased. Simultaneously, feedback data from the power adjustments, such as actual voltage and frequency values, as well as CPU status changes, is transmitted back to the data acquisition module, forming a closed-loop control system that continuously optimizes the power allocation strategy.

[0036] S4, Model Training and Update Phase: During the initial system operation, the LSTM prediction model and adaptive reinforcement learning model are initially trained using a large amount of historical operating data to determine the model's initial parameters. This historical data contains rich information about CPU operating status and power distribution, helping the model to initially learn the relationship between the two. During system operation, incremental training is performed every half hour using newly collected real-time data. Using transfer learning methods, the model retains the common features already learned and only updates parameters relevant to the new data. This allows the model to continuously adapt to changes in the PC operating environment and user habits, ensuring the system's long-term stability and efficient allocation of idle CPU power.

[0037] 3. Experimental Comparison To more clearly demonstrate the advantages of the AI-based PC CPU idle power distribution system, the following compares the differences in key parameters, power consumption, and performance between the system of the present invention and traditional fixed power distribution systems in three common scenarios: office work, gaming, and video editing.

[0038]

[0039] From the comparison of the above tables, it is obvious that: In office scenarios, the CPU utilization of both the AI-based and traditional systems fluctuated between 23.5% and 38.2%. However, other metrics differed significantly. The AI-based system was able to control the CPU temperature to 41.3-44.7°C, dynamically adjusting the CPU frequency between 2.56-2.94GHz based on load, dropping to 2.63GHz under low load. The supply voltage also fluctuated between 1.18-1.22V, dropping to 1.12V under low load. Thanks to this precise dynamic adjustment, its power consumption was reduced by approximately 11.6% compared to the traditional system. The traditional system, on the other hand, had a CPU temperature of 46.1-49.3°C, a constant frequency of 3.02GHz, and a fixed supply voltage of 1.23V. This high power consumption not only wasted energy but also caused brief pauses under high load, impacting office productivity. For example, there was noticeable delay when opening large documents or multitasking.

[0040] Gaming scenarios demand high CPU performance and experience frequent load fluctuations. In this scenario, CPU utilization for both systems fluctuated between 62.4% and 78.5%. The AI-based system performed exceptionally well in temperature control, maintaining a constant temperature of 51.2-58.9°C. The CPU frequency flexibly adjusted between 3.05-3.48 GHz based on load, increasing during high loads to maintain performance. The supply voltage fluctuated between 1.21 and 1.29 V. This dynamic adjustment reduced system power consumption by approximately 8.4% compared to traditional systems, while ensuring stable frame rates and no noticeable frame drops, allowing players to enjoy a smooth gaming experience even in complex scenarios. The traditional system, with its constant 3.52 GHz frequency and fixed 1.31 V supply voltage, maintained reasonable performance in simple scenarios. However, in complex gaming scenarios, CPU temperatures reached as high as 61.5-68.7°C. The high temperatures and fixed parameters resulted in significant frame rate fluctuations and frequent, noticeable frame drops, severely impacting the gaming experience and overall immersion.

[0041] Video editing involves a large amount of data processing and rendering work, which places strict demands on CPU performance. In this scenario, the CPU utilization rate of the AI-based system is between 52.6% and 68.3%, the temperature is maintained at 46.8-53.5°C, the CPU frequency dynamically changes between 2.83-3.27GHz according to the intensity of tasks such as rendering, and the power supply voltage is adjusted between 1.22-1.24V, and the power consumption is reduced by about 9.8% compared to traditional systems. This makes video rendering fast and preview smooth, greatly improving the efficiency of video editing. Under the same CPU utilization rate, the temperature of the traditional system reaches 56.2-64.1°C, the frequency is always 3.31GHz, and the power supply voltage is fixed at 1.26V. However, the high power consumption does not bring better performance. The video rendering time is long, and there will be freezes during preview, which increases the time cost of video editing.

[0042] Comprehensive experimental data from three scenarios shows that the AI-based PC CPU idle power distribution system comprehensively outperforms traditional systems in terms of energy conservation and performance optimization. It intelligently adjusts the CPU frequency and supply voltage based on the real-time load in different scenarios, reducing power consumption while ensuring that the PC maintains excellent performance under various workloads. Traditional fixed power distribution systems, however, lack the flexibility to adapt to load changes and suffer from significant shortcomings in both energy consumption and performance.

[0043] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An AI-based PC CPU idle power distribution system, characterized by: Includes the following modules: S1. Data acquisition module: uses motherboard sensors, operating system performance monitoring tools and process monitoring interfaces to collect real-time hardware status data and resource demand information of currently running programs; S2, algorithm processing module: S2.1, Multimodal data fusion submodule: Input CPU temperature data into CNN, according to the convolution operation formula Extract features and get , the time series data of CPU usage and frequency are input into LSTM, and processed by LSTM to obtain , and finally merge the two into ;in, is the feature extracted by CNN for temperature data, The features extracted by LSTM from the time series data of CPU usage and frequency are: for and The fusion characteristics of S2.2, attention mechanism submodule: In the LSTM prediction model and decision model, according to the formula ,in, Calculate the attention weight and perform weighted processing on the data; among them, is the LSTM hidden state sequence The hidden states, For the current moment The hidden state of is the LSTM hidden state sequence, ; S2.3, Adaptive reinforcement learning submodule: The adaptive reinforcement learning module uses DQN as the framework and and action space , calculate the Q value , where the state space The action space is composed of the current CPU hardware status and program resource requirements. is a combination of different supply voltages and frequencies, based on the dynamic reward function Evaluate the benefits of different actions, here and is the weight that is dynamically adjusted according to the system status. Indicates the energy efficiency after taking action in the current state, Indicates performance loss; are the network parameters of the deep Q network; S3, power control module: Based on the decision results of the algorithm processing module, the CPU power supply voltage and frequency are adjusted through the motherboard BIOS or hardware control chip to achieve accurate distribution of idle power. At the same time, the feedback data after power adjustment is sent back to the data acquisition module to form a closed-loop control.

2. The AI-based PC CPU idle power distribution system according to claim 1, characterized in that: In the multimodal data fusion submodule, the temperature data is assumed to be , input convolutional neural network to extract features, the CNN convolution operation formula is: ,in It is Convolutional layer The output of the position, It is Layer convolution kernel weights, It is layer Position input, It is Layer bias, and Is the size of the convolution kernel, and the temperature data feature representation is obtained through the convolution operation .

3. The AI-based PC CPU idle power distribution system according to claim 1, characterized in that: In the multimodal data fusion submodule, the time series composed of CPU usage and frequency is assumed to be , input long short-term memory network, LSTM core calculation formula: Input Gate: ; Forget Gate: ; Output gate: ; Memory unit: ; Hidden state: ; in is the sigmoid function, represents element-wise multiplication, is the weight matrix, is the bias vector, is the current input, is the hidden state at the previous moment, It is the memory unit of the previous moment, and the time series data feature representation is obtained through LSTM .

4. The AI-based PC CPU idle power distribution system according to claim 1, characterized in that: In the adaptive reinforcement learning submodule, the Q value is updated by the Bellman equation: ,in is the learning rate, It’s an instant reward. is the discount factor, S' is the next state, and a' is the next action.

5. The AI-based PC CPU idle power distribution system according to claim 1, characterized in that: In the initial stage of operation, the system uses historical operation data to perform initial training on the LSTM prediction model and the adaptive reinforcement learning model to determine the initial parameters. During the operation of the system, incremental training is performed every half an hour using newly collected real-time data.

6. The AI-based PC CPU idle power distribution system according to claim 5, characterized in that: During the operation of the system, a transfer learning method is adopted to retain the common features that the model has learned and only update the parameters related to the new data.

Images