A health management method for an audio integrated signal processing machine GPU computing board card

Abstract

The application discloses a health management method of an audio integrated signal processing machine GPU computing board, and belongs to the technical field of audio integrated signal processing, wherein multi-dimensional data of GPU cores, display memories and power supply modules are collected in real time through sensors, a three-dimensional temperature field is constructed by adopting an inverse distance weighted interpolation algorithm, and a thermal coupling region is accurately positioned; based on an LSTM model, thermal field evolution is predicted, high-density instructions are dynamically scheduled to NUMA cores with sufficient heat dissipation margins; for burst reading and writing of the display memory, a micro piezoelectric fan array is adopted to form vertical airflow (wind speed is greater than or equal to 4 m / s), and a dynamic voltage reduction and hierarchical heat dissipation strategy (phase change heat conduction sheet pre-cooling temperature reduction is 8-10 DEG C, and the attack angle of the centrifugal fan blade is 45 DEG) is combined. The heat source recognition accuracy of the scheme is improved by 40%, the heat dissipation delay is less than or equal to 45 mu s, the computing power loss is less than or equal to 0.8%, the display memory error rate is less than or equal to 1E-6, and the high-density computing heat management is significantly optimized.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of audio integrated signal processing, and particularly relates to a health management method for a GPU computing board of an audio integrated signal processing machine. BACKGROUND

[0002] The current audio integrated signal processing field has a continuously rising demand for the computing power of GPU computing board cards, but the heat runaway problem caused by high-density computing seriously restricts system stability. Traditional cooling schemes use a global forced cooling mode, such as a full-speed fan or liquid cooling circulation, which can temporarily suppress temperature rise but leads to a sharp increase in energy consumption and computing power loss. Taking air cooling as an example, the heat flow density limit is only 200 W / cm², and in the face of the concentrated heat of 3D packaged chips, the blind area of heat dissipation increases significantly. Although the existing monitoring system can collect basic data such as temperature and video memory occupancy, it lacks the ability to model three-dimensional thermal fields, and the discrete temperature point monitoring cannot capture the instantaneous thermal coupling effect of the stream processor array and the video memory overlap area, resulting in a lag in the heat dissipation response to the heat source migration. More seriously, task scheduling and heat dissipation control have been long separated, and the local temperature rise rate can reach 15℃ / s when the video memory suddenly reads and writes, causing a signal distortion rate of 1E-4, which is difficult to meet the requirements of high-precision audio processing.

[0003] In recent years, the industry has tried to optimize the heat dissipation efficiency through intelligent algorithms, but the key bottleneck has not been broken. Although the liquid cooling system improves the heat transfer coefficient by using micro-channel design, it increases the cost due to the complex structure and corrosion protection requirements; although the AI prediction model (such as LSTM) can predict the temperature trend, it is not deeply coupled with the hardware-level heat dissipation action, and lacks a preloading mechanism before the execution of the instruction, resulting in a heat dissipation delay of more than 200μs. In addition, the existing control strategy ignores the differences in thermal characteristics of heterogeneous units: the vertical heat dissipation of the video memory requires a micro piezoelectric fan array and dynamic voltage reduction, while the traditional scheme only relies on single wind speed adjustment, and the video memory error rate is still higher than 1E-5. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a health management method for a GPU computing board of an audio integrated signal processing machine, which dynamically maps a three-dimensional thermal map model to accurately match the computing peak and the heat dissipation response, and realizes the directional deployment of heat dissipation resources by combining a hierarchical preloading strategy, thereby providing a high-reliability, low-latency thermal management solution for audio processing scenarios.

[0005] To achieve the above purpose, the present application provides the following technical scheme:

[0006] The present application comprises the following steps:

[0007] S1, real-time collection of the operation intensity, instantaneous power consumption and temperature data of the GPU core, video memory and power supply module, and generation of a data set;

[0008] S2, constructing a temperature-load three-dimensional thermodynamic map based on the data set, and establishing a dynamic mapping model for calculating the peak and heat dissipation response based on the three-dimensional thermodynamic map;

[0009] S3, based on the dynamic mapping model, dispersing high-density computing instructions to different operation units according to a preset "hot spot prediction-heat dissipation preloading" algorithm;

[0010] S4, analyzing the data set, starting the phase change heat dissipation sheet to quickly absorb heat when the core temperature reaches a first threshold value, and when the core temperature reaches a second threshold value, the centrifugal fan is linked to speed up and adjust the air outlet angle to form a directional air duct, and when the video memory bursts read and write, the vertical heat dissipation airflow is preferentially enhanced instead of global forced cooling.

[0011] Further, the step S2 comprises:

[0012] Collecting temperature data of GPU cores, video memory and power supply module area, recording the computing load and power consumption parameters of each operation unit, mapping the hardware physical layout into a three-dimensional coordinate system, using inverse distance weighted interpolation algorithm to perform spatial interpolation on discrete temperature points based on temperature data, and forming a continuous three-dimensional temperature field;

[0013] The three-dimensional temperature field is converted into a thermodynamic map expressed in two dimensions of color and height by layer rendering technology;

[0014] Based on the simultaneously recorded computing load and power consumption parameters of each operation unit, a dynamic mapping model of computing task and thermodynamic map is established, and the operation instruction flow is converted into heat flux density function H through matrix superposition algorithm, and the coupling law of high load instruction and specific area heat flux density is determined by convolution operation with real-time temperature field.

[0015] Further, the step S2 further comprises:

[0016] By implanting a dynamic threshold correction mechanism, the interpolation accuracy of the thermodynamic field and the heat dissipation response timing are continuously optimized according to historical data, and finally the closed-loop linkage of computing task scheduling, temperature field evolution and heat dissipation resource allocation is realized;

[0017] The implementation of the dynamic threshold correction mechanism comprises:

[0018] (1) Constructing a thermodynamic gradient residual function based on historical temperature field data , wherein is the predicted temperature value, is the measured temperature value;

[0019] (2) When R exceeds 5℃ for 3 consecutive periods, automatically adjust the power parameter of the inverse distance weighted interpolation algorithm , and recalculate the three-dimensional temperature field;

[0020] (3) Slide window calibration of heat dissipation response timing: the heat dissipation action trigger delay is counted in 50ms cycles, if the standard deviation of the delay is ≥8μs, then shorten the preload trigger advance by 20% steps;

[0021] (4) Optimize the matching relationship between interpolation accuracy and heat dissipation timing through genetic algorithm, so that the thermal gradient prediction deviation is stabilized within ±2℃.

[0022] Further, the step S3 comprises:

[0023] Based on the dynamic mapping model, determine the coupling law of high-load instructions and specific area heat flux density, and use the LSTM time series model to predict the thermal field evolution trend in the next 200ms;

[0024] When it is identified that the heat flux density of a specific area will exceed the standard, the system automatically activates the instruction redistribution engine, disassembles high-density computing instructions into multiple sub-tasks, and according to the physical core mapping relationship of the NUMA architecture, combined with the real-time heat dissipation margin table of each operation unit, preferentially schedules the instruction units with thermal effect value exceeding the threshold to the idle area of the heat dissipation resource;

[0025] Trigger the target area heat dissipation preload 150μs before the actual execution of the instruction; simultaneously adjust the blade angle of the centrifugal fan to 45°, forming a directional accelerated air duct in the target area;

[0026] Through the implantation of the reinforcement learning framework, continuously optimize the instruction dispersion granularity and heat dissipation response timing parameters, and finally realize the optimal matching of computing instruction distribution and heat dissipation resources.

[0027] Further, the heat flux density function H in the dynamic mapping model is optimized in the following way:

[0028] (1) Introduce dynamic voltage frequency adjustment parameter correction coefficient , where , is the instantaneous temperature value of the target area at the current time;

[0029] (2) Input the corrected heat flux density function into the LSTM prediction model to generate a space-time correlation matrix containing temperature field gradient and heat dissipation margin;

[0030] (3) When the predicted temperature rise rate is ≥10℃ / s, trigger the phase change heat conduction sheet pre-cooling mode of the target area simultaneously, so that the base temperature drops by 8-10℃ within 120μs;

[0031] (4) Optimize the weight distribution of the DVFS parameter correction coefficient through the Q-learning algorithm, so that the heat dissipation response delay is compressed to within 45μs and the computing power loss rate is ≤0.8%.

[0032] Furthermore, the vertical heat dissipation strategy for burst read / write of video memory in step S4 includes:

[0033] (1) Calculate the vertical heat flux density in real time by using the memory bandwidth utilization rate. ,in This refers to the utilization rate of video memory bandwidth. This refers to the temperature of the memory chips.

[0034] (2) When V≥85%, the following compound action is initiated:

[0035] ① Activate the micro piezoelectric fan array 2mm above the memory module to form a local vertical airflow with a wind speed ≥4m / s;

[0036] ② Close the horizontal air duct of the adjacent power supply module and concentrate 80% of the airflow to the memory area;

[0037] ③ Implement dynamic voltage reduction for the video memory power supply circuit, with a voltage drop of [missing information]. for, Nominal voltage;

[0038] (3) By optimizing the coordination parameters of vertical airflow distribution and voltage drop using a reinforcement learning framework, the memory temperature rise rate is reduced by 40% and the bit error rate is ≤1E-6.

[0039] The beneficial effects of this invention are as follows:

[0040] This invention constructs a three-dimensional heat map and establishes a dynamic mapping model by collecting multi-dimensional data in real time. It combines hotspot prediction and heat dissipation preloading algorithms to achieve intelligent instruction scheduling and simultaneously optimizes vertical heat dissipation strategies. This effectively solves the problems of high heat dissipation latency and high memory error rate in traditional solutions. It has the advantages of accurate three-dimensional thermal field modeling and dynamic heat dissipation coordinated control, and significantly reduces memory error rate and computing power consumption.

[0041] Other advantages, objectives, and features of the invention will be set forth in the following description and will be apparent to those skilled in the art in some respects, or may be learned by practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0042] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the following figures are provided for illustration:

[0043] Figure 1 This is a flowchart illustrating the steps of the health management method of the present invention; Detailed Implementation

[0044] refer to Figure 1The application provides a health management method for an audio integrated signal processor GPU computing board, which comprises the following steps: collecting operation intensity, instantaneous power consumption and temperature data of GPU cores, display memory and power supply modules in real time and generating a data set; constructing a temperature-load three-dimensional thermal map according to the data set and establishing a dynamic mapping model for calculating wave peaks and heat dissipation responses; dispersing high-density computing instructions to different operation units according to a preset algorithm based on the dynamic mapping model; and analyzing the data set and starting phase change heat conduction sheets to absorb heat, adjusting centrifugal fan parameters and enhancing vertical heat dissipation air flow in stages.

[0045] The data set refers to a collection of synchronous collected operation intensity, power consumption and temperature parameters, which can be realized by an embedded sensor array and a power monitoring chip and is used for reflecting dynamic thermal correlation characteristics of hardware modules. The temperature-load three-dimensional thermal map refers to a heat distribution visualization model formed by mapping a physical layout into a three-dimensional coordinate system, which is generated by using an inverse distance weighted interpolation algorithm and can accurately represent the spatial thermal coupling effect of a chip packaging structure. The dynamic mapping model refers to a space-time correlation matrix of computing tasks and heat dissipation resources, which is constructed by using a matrix superposition algorithm and is used for quantifying the coupling relationship between operation instructions and heat flux density. The "hot spot prediction-heat dissipation preloading" algorithm refers to a task scheduling strategy based on thermal field prediction, which is realized by using an LSTM model and can trigger heat dissipation preloading actions before instruction execution. The hierarchical response strategy refers to starting differential heat dissipation measures according to temperature thresholds, which is realized by using phase change materials and fluid mechanics control and is used for optimizing vertical air flow distribution for display memory burst read-write.

[0046] Specifically, multi-dimensional data collection covers dynamic thermal parameters of operation units, providing multi-dimensional input for three-dimensional thermal field modeling. Discrete temperature data is converted into a continuous temperature field by using a spatial interpolation algorithm, accurately capturing the thermal coupling effect of the overlapping area of a stream processor array and display memory. The dynamic mapping model quantizes computing instructions into heat flux density functions, predicting the thermal evolution trend of a specific area. When potential heat flux density exceeds the standard, the task scheduling engine is used to disassemble high-density instructions into subtasks and distribute them to areas with sufficient heat dissipation margins. For display memory burst read-write scenarios, a micro piezoelectric fan array is used to form local vertical air flow, and the horizontal air duct is closed to concentrate heat dissipation resources. The hierarchical triggering mechanism selects phase change heat conduction or directional air duct heat dissipation according to real-time temperature data, avoiding energy waste caused by global cooling.

[0047] Compared with the prior art, the traditional scheme relies on two-dimensional temperature monitoring and full-speed fan control, and cannot accurately locate the thermal coupling area of the three-dimensional packaging structure. The scheme realizes real-time tracking of heat source migration by constructing a three-dimensional thermal field model, which improves the identification accuracy of the overlapping area of the processor array compared with the plane monitoring method. The dynamic mapping model establishes the space-time correlation between the computing task and the heat dissipation resource, overcoming the defect that the task scheduling and heat dissipation control are separated in the traditional scheme. The vertical heat dissipation strategy optimizes the air flow distribution for the GPU burst read-write scene, and is more suitable for the thermal management needs of high-bandwidth data exchange than the horizontal air duct mode. The hierarchical response mechanism effectively reduces the invalid heat dissipation energy consumption by triggering the heat dissipation action in a differentiated manner.

[0048] Through the above technical scheme, the present application realizes the accurate modeling of the three-dimensional thermal field in the high-density computing scene and effectively solves the problem of heat dissipation response lag. The dynamic task scheduling mechanism disperses the computing load to the area with sufficient heat dissipation margin, avoiding the loss of computing power caused by local overheating. The vertical heat dissipation strategy significantly reduces the temperature rise rate during GPU burst read-write, ensuring the stability of high-bandwidth data exchange. The hierarchical triggering control selects the optimal heat dissipation mode according to real-time temperature data, maintaining the safe operation of the system while reducing the overall energy consumption level.

[0049] The present application further proposes the following technical scheme: collecting temperature data of GPU cores, GPU memory and power supply module area, recording the computing load and power consumption parameters of each operation unit at the same time, mapping the hardware physical layout into a three-dimensional coordinate system, using the inverse distance weighted interpolation algorithm to perform spatial interpolation on the discrete temperature points based on the temperature data, forming a continuous three-dimensional temperature field; through the layer rendering technology, the temperature gradient in the three-dimensional temperature field is converted into a thermal map expressed in two dimensions of color and height; based on the simultaneously recorded computing load and power consumption parameters of each operation unit, a dynamic mapping model of computing task and thermal map is established, and through the matrix superposition algorithm, the operation instruction is converted into a heat flux density function H, and the real-time temperature field is convolved to determine the coupling law of high-load instruction and specific area heat flux density.

[0050] The inverse distance weighted interpolation algorithm refers to a method of interpolating discrete temperature points based on spatial position weight, and can be specifically implemented by calculating the inverse distance weighted value of each interpolation point and the known temperature point. This method can break through the spatial resolution limit of discrete monitoring points and form a temperature field that conforms to the physical heat conduction law. The layered rendering technology refers to a method of converting gradient information of a three-dimensional temperature field into visual expression, and can be specifically implemented by using a method of mapping temperature values to color depth and height parameter superimposed rendering. This method enables the dynamic change of temperature distribution to be intuitively identified. The heat flux density function H refers to a calculation model of converting an operation instruction stream into a spatial heat load distribution, and can be specifically implemented by using a matrix superposition algorithm to quantify the instantaneous power consumption and heat distribution generated by instruction execution. This function can represent the heat effect accumulation process of a computing task in a three-dimensional space.

[0051] Specifically, the scheme synchronously collects temperature data and computing load parameters, establishes a three-dimensional coordinate system of the hardware physical layout, and ensures the physical accuracy of thermal field modeling. The inverse distance weighted interpolation algorithm generates a continuous three-dimensional temperature field by spatially interpolating discrete temperature points, accurately representing the heat flow transmission path inside the chip. The layered rendering technology converts temperature gradients into two-dimensional expressions of color and height, improving the visualization of thermal distribution and facilitating the identification of the instantaneous thermal coupling effect of the stream processor array and the video memory overlap area. The heat flux density function generated based on the operation instruction stream reveals the dynamic correlation between the computing task and the local heat flux density through convolution operation with the real-time temperature field, providing a physical model basis for subsequent heat dissipation control.

[0052] Compared with the prior art, the traditional scheme only uses discrete temperature point monitoring, which cannot accurately capture the heat flow transmission process in a three-dimensional space, resulting in a lag of heat dissipation response behind heat source migration. The present scheme constructs a continuous temperature field through a spatial interpolation algorithm, and realizes visual tracking of thermal distribution by using layered rendering technology, which can reflect the temperature gradient change of the stream processor array and the video memory overlap area in real time. The dynamic mapping model breaks the fragmented relationship between the operation task and the thermal distribution in the prior art, and predicts the heat flux density change caused by high-load instructions in advance through convolution operation of the heat flux density function and the temperature field.

[0053] Through the above technical scheme, the present application solves the problem that the traditional monitoring method cannot capture the instantaneous thermal coupling effect, and realizes accurate modeling of the thermal field distribution in a three-dimensional space. The dynamic correlation analysis of the heat flux density function and the temperature field can identify the local temperature rise risk caused by high-load instructions in advance. The layered rendering technology visualizes the hidden thermal coupling area, providing spatial positioning basis for precise allocation of heat dissipation resources. The mapping relationship between the operation instruction and the thermal distribution is established, ensuring that the heat dissipation response action is synchronized with the heat source migration, and avoiding the temperature rise out of control caused by heat dissipation delay in the traditional scheme.

[0054] The application further proposes to continuously optimize the thermal field interpolation accuracy and heat dissipation response timing based on historical data by implanting a dynamic threshold correction mechanism, and finally realize the closed-loop linkage of computing task scheduling, temperature field evolution and heat dissipation resource allocation. The implementation of the dynamic threshold correction mechanism includes: constructing a thermal gradient residual function based on historical temperature field data, automatically adjusting the power parameter of the inverse distance weighted interpolation algorithm and recalculating the three-dimensional temperature field when the residual continuously exceeds the threshold, and sliding window calibration of the heat dissipation response timing, and optimizing the matching relationship of interpolation accuracy and heat dissipation timing by genetic algorithm.

[0055] Among them, the dynamic threshold correction mechanism refers to the system module that dynamically adjusts the temperature field modeling parameters and heat dissipation timing control according to historical operation data, which can be realized by embedded adaptive control algorithm, which is used to eliminate the cumulative error of the prediction model and the actual heat field distribution. The residual function refers to the cumulative amount of the absolute difference between the predicted temperature and the measured temperature, which can be realized by sliding window statistical method, which is used to quantify the deviation degree of heat field modeling accuracy. The inverse distance weighted interpolation algorithm refers to a mathematical method for calculating the temperature value of an unknown point based on the inverse distance weighting, which can be realized by variable power parameter adjustment strategy, which is used to dynamically optimize the sensitivity of temperature field spatial interpolation. The sliding window calibration refers to the optimization mechanism for delaying the heat dissipation action based on time series fragment statistics, which can be realized by standard deviation threshold judgment method, which is used to eliminate the random fluctuation of heat dissipation response timing. Genetic algorithm refers to a global optimization algorithm that simulates the biological evolution process, which can be realized by multi-objective fitness function design, which is used to search for the optimal matching combination of interpolation parameters and heat dissipation timing.

[0056] Specifically, in the temperature field modeling process, the thermal gradient residual function is continuously calculated to monitor the accuracy of the prediction model. When it is detected that the residual index continuously exceeds the set threshold, the core parameter of the inverse distance weighted interpolation algorithm is automatically adjusted, for example, the power parameter is increased from the default value to a higher value, so as to enhance the influence weight of the neighboring temperature points and improve the interpolation accuracy of the local hot spot area. In terms of heat dissipation control timing optimization, the system statistically calculates the actual trigger delay of the heat dissipation action at a fixed period, and when the timing stability decreases, the control parameter is dynamically corrected by gradually shortening the preloaded trigger time. Finally, genetic algorithm is applied to the joint optimization of the temperature field interpolation parameter set and the heat dissipation timing parameter set, and the parameter combination that can minimize the prediction deviation is selected by simulating the natural selection process, forming a closed-loop linkage mechanism of temperature field evolution and heat dissipation resource allocation.

[0057] Compared with the prior art, the traditional method adopts a fixed parameter interpolation algorithm and lacks a timing calibration mechanism, resulting in accumulated temperature field modeling errors over time, and timing misalignment between heat dissipation actions and heat source migration. The scheme realizes adaptive adjustment of modeling parameters and accurate matching of heat dissipation timing through dynamic threshold correction and genetic algorithm optimization, overcoming the defect that a static model cannot adapt to dynamic heat field changes.

[0058] Through the above technical solution, the application can effectively improve the spatial resolution of three-dimensional thermal field modeling, reduce the influence of thermal gradient prediction deviation on heat dissipation control, and at the same time ensure the accurate synchronization of the heat dissipation action triggering time and the heat source migration process. The technical solution solves the problem of local overheating caused by modeling errors and response delays in the traditional method, and realizes intelligent collaborative control of computing task scheduling and heat dissipation resource allocation.

[0059] The application further proposes a health management method for an audio integrated signal processing machine GPU computing board, comprising the steps of: determining the coupling law of high-load instructions and specific area heat flux based on a dynamic mapping model, and predicting the thermal field evolution trend in the future time period using an LSTM time series model; activating the instruction reallocation engine when it is identified that the heat flux of the specific area will exceed the standard, and decomposing the high-density computing instructions into multiple sub-tasks, and scheduling the instruction units according to the physical core mapping relationship of the NUMA architecture combined with the real-time heat dissipation margin dynamic table; triggering the target area heat dissipation preload and adjusting the centrifugal fan blade angle to form a directional acceleration air duct before the actual execution of the instructions; and optimizing the instruction dispersion granularity and heat dissipation response timing parameters through a reinforcement learning framework.

[0060] The LSTM time series model refers to a time series prediction model established by a long short-term memory neural network, which can be trained using historical temperature field data to capture the time correlation of thermal field evolution. The instruction reallocation engine refers to a logical module based on computing task decomposition and resource scheduling, which can be implemented through dynamic task queue management to disperse centralized computing load to different operation units. The physical core mapping relationship of the NUMA architecture refers to the physical connection topology of the processor core and the memory area in the non-uniform memory access architecture, which can be implemented by binding the core number and the memory controller to ensure the physical consistency of task scheduling and hardware layout. The real-time heat dissipation margin dynamic table refers to a dynamic data table recording the current heat dissipation capacity of each operation unit, which can be generated using temperature sensor and heat sink state monitoring data to guide heat dissipation resource matching during task scheduling. The reinforcement learning framework refers to a machine learning architecture that optimizes decision-making strategies based on environmental feedback, which can be implemented using the Q-learning algorithm to continuously optimize the collaborative parameters of instruction allocation and heat dissipation actions.

[0061] Specifically, this method achieves precise matching of heat dissipation resources through a collaborative mechanism of thermal field prediction and dynamic task scheduling. First, an LSTM model is used to analyze the temperature field evolution trend, such as predicting the heat flux density distribution within the next 200ms, to identify potential overheating areas in advance. When the heat flux density is detected to be about to exceed a threshold, the instruction reassignment engine decomposes high-density computing tasks into multiple subtasks, such as decomposing matrix operations into parallel vector calculations, and prioritizing the allocation of high-heat-generating tasks to computing units with sufficient heat dissipation margins based on the physical layout of the NUMA architecture. After task scheduling is completed, the heat dissipation devices in the target area are activated in advance before the actual execution of the instructions, such as activating phase change heat sinks 150μs beforehand, while adjusting the centrifugal fan blade angle to form directional airflow, effectively shortening the heat dissipation response delay. Finally, a reinforcement learning algorithm is used to continuously optimize the matching relationship between task granularity and heat dissipation timing, such as updating the scheduling strategy parameters weekly, gradually improving the system's adaptability.

[0062] Compared to existing technologies, traditional solutions employ global temperature monitoring and full-speed heat dissipation, which cannot predict the evolution trend of heat flux density, resulting in heat dissipation lagging behind heat source migration. This solution, however, utilizes a thermal field prediction model and real-time task scheduling to pre-configure heat dissipation resources before computational instructions are executed, such as pre-forming directional airflow channels, effectively preventing localized temperature runaway. In existing technologies, task scheduling and heat dissipation control are separate modules, while this solution uses a reinforcement learning framework to achieve joint optimization of both, such as synchronously adjusting task allocation granularity and fan response timing, significantly improving system coordination efficiency.

[0063] Through the above technical solution, this application effectively solves the problem of excessive local heat flux density caused by high-density computing, and avoids the risk of thermal runaway by predicting in advance and redistributing tasks. This method optimizes the collaborative efficiency of computing task scheduling and heat dissipation resource allocation, achieving precise resource matching through physical layout mapping and real-time heat dissipation margin assessment. The system has dynamic adaptive capabilities, continuously optimizing instruction distribution strategies and heat dissipation response parameters to ensure thermal balance under high-load computing.

[0064] This application further proposes to optimize the heat flux density function by introducing a dynamic voltage and frequency adjustment parameter correction coefficient. ,in , The instantaneous temperature value of the target region at the current moment; the corrected heat flux density function. An input LSTM prediction model generates a space-time correlation matrix containing temperature field gradient and heat dissipation margin; when the predicted temperature rise rate is greater than or equal to 10℃ / s, the pre-cooling mode of the phase change heat conduction sheet in the target area is triggered synchronously, so that the substrate temperature drops by 8-10℃ within 120μs; the weight distribution of the DVFS parameter correction coefficient is optimized by the Q-learning algorithm, so that the heat dissipation response delay is compressed to within 45μs and the computing power loss rate is less than or equal to 0.8%.

[0065] Wherein, the dynamic voltage frequency adjustment parameter correction coefficient γ refers to the nonlinear correction factor of the influence of dynamically adjusting voltage frequency on heat generation according to instantaneous temperature change, which can be calculated by combining the values collected by the real-time temperature sensor with the square function, and is used to quantify the actual influence of voltage frequency adjustment on heat flux density. The space-time correlation matrix refers to the prediction model input structure that integrates temperature gradient distribution and heat dissipation resource state, which can be realized by using LSTM neural network to extract space-time features of multidimensional data, and is used to establish the dynamic correlation between heat field evolution and heat dissipation capacity. The pre-cooling mode of the phase change heat conduction sheet refers to the active heat dissipation mechanism of using phase change material to quickly absorb heat at a specific temperature interval, which can be realized by using the solid-liquid phase change characteristics of gallium-based liquid metal composite, and is used to quickly reduce the substrate temperature when the temperature rise rate exceeds the standard. The weight distribution of the Q-learning algorithm optimized DVFS parameter correction coefficient refers to dynamically adjusting the matching relationship between voltage frequency parameters and heat dissipation response through the reinforcement learning framework, which can be realized by using the state-action value function iterative update method, and is used to balance the conflict between heat dissipation delay and computing power loss.

[0066] Specifically, the technical scheme realizes precise coordination of thermal field prediction and heat dissipation control by establishing a heat flux density function optimization mechanism linked with dynamic voltage frequency adjustment. The correction coefficient γ calculated based on the instantaneous temperature value is introduced into the heat flux density function, which converts the nonlinear influence of voltage frequency adjustment on heat generation into a quantifiable adjustment factor. After the corrected heat flux density function is input into the time series prediction model, a space-time correlation matrix reflecting both temperature gradient distribution and heat dissipation resource margin is generated, forming a prediction framework that takes into account both heat field evolution rules and heat dissipation resource state. When the temperature rise rate exceeds the set threshold, the pre-cooling mode of the phase change heat conduction sheet is immediately activated, using the rapid heat absorption characteristics of phase change materials to achieve a sharp drop in substrate temperature. The weight distribution of the voltage frequency parameter correction coefficient is continuously optimized by the reinforcement learning algorithm, so that the heat dissipation response delay and computing power loss are optimally balanced.

[0067] Compared with the prior art, the existing heat flow density prediction model does not consider the influence of dynamic voltage frequency adjustment on heat generation, resulting in deviation of heat dissipation preloading opportunity and heat dissipation resource allocation. The traditional scheme adopts fixed coefficient to predict heat flow density, which cannot reflect the thermal effect fluctuation caused by dynamic voltage frequency adjustment. The scheme introduces dynamic correction coefficient and reinforcement learning optimization mechanism to establish the dynamic coupling relationship between heat generation and heat dissipation control, and overcomes the defect of insufficient prediction accuracy of static model.

[0068] Through the above technical scheme, the application can reflect the influence of voltage frequency adjustment on heat generation in real time, improve the prediction accuracy of heat flow density, and optimize the accuracy and timeliness of heat dissipation resource allocation. When rapid temperature rise is detected, directional rapid cooling is realized, and at the same time, the parameter configuration is continuously optimized through intelligent algorithm, so as to effectively balance the contradiction between heat dissipation efficiency and calculation performance.

[0069] The application further proposes a vertical heat dissipation strategy for GPU burst read and write, which includes calculating the vertical heat flow density in real time through GPU bandwidth occupancy, starting the composite heat dissipation action when the vertical heat flow density reaches the threshold, and optimizing the cooperative parameters by using the reinforcement learning framework.

[0070] The vertical heat flow density is a quantitative index constructed by combining GPU bandwidth occupancy and GPU particle temperature, which can be realized by using the weighted calculation formula to dynamically reflect the urgency of GPU area heat dissipation demand. The micro piezoelectric fan array refers to the micro air flow generator arranged at a specific distance above the GPU module, which can be realized by using the piezoelectric ceramic driven fan manufactured by MEMS technology, and directional vertical air flow is generated through high frequency vibration to penetrate the dense packaging structure of the GPU module. Dynamic voltage reduction refers to adjusting the GPU power supply voltage according to the over-limit degree of vertical heat flow density, which can be realized by adjusting the PWM duty cycle through a digital power management chip to reduce power loss from the heat source end to suppress temperature rise. The reinforcement learning framework refers to a parameter optimization system based on state-action-reward mechanism, which can be realized by using deep deterministic policy gradient algorithm, and is used to dynamically coordinate the synergistic effect of air flow distribution and voltage adjustment.

[0071] Specifically, when the vertical heat flow density index is monitored in real time to exceed the set threshold, the heat dissipation control system performs multi-dimensional cooperative action: first, activate the micro piezoelectric fan array above the GPU area to form high-speed air flow in the vertical direction at the physical level; at the same time, close the transverse air duct of the adjacent power supply module to concentrate the limited heat dissipation resources on the GPU high temperature area; simultaneously start the dynamic adjustment of GPU power supply voltage to reduce heat source by reducing operating voltage. In this process, the reinforcement learning system continuously collects heat dissipation response data to ensure the dynamic matching of vertical air flow intensity and voltage drop through iterative optimization, so as to maintain the integrity of GPU signal while avoiding excessive heat dissipation loss.

[0072] Compared with the prior art, the traditional scheme relies on a fixed-direction transverse air duct for global heat dissipation, cannot effectively cover the heat dissipation blind area of the vertically stacked structure during burst read and write of the display memory, and has a response delay between the heat dissipation action and the heat source migration. The scheme breaks through the coverage limit of the traditional heat dissipation path on the high-density display memory module by constructing a display memory exclusive vertical heat dissipation system, using a local micro fan array and air duct redirection technology, simultaneously introducing dynamic adjustment of the power supply voltage to form a heat source suppression mechanism, and realizing full-link control from heat conduction to heat generation.

[0073] Through the above technical solution, the application effectively solves the problem of excessively high local temperature rise rate in the display memory burst read and write scenario, significantly reduces the transient temperature fluctuation of the display memory particles through the synergistic effect of vertical airflow directional reinforcement and heat source power suppression, and ensures the continuous adaptation of the heat dissipation control strategy to the actual load of the display memory through parameter optimization driven by reinforcement learning, finally maintains the heat dissipation efficiency while controlling the bit error rate within the system safety threshold.

[0074] Finally, it should be pointed out that the above preferred embodiments are only used to illustrate the technical solutions of the application and not to limit it, although the application has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the application.

Claims

1. A health management method for an audio integrated signal processing unit (GPU) computing board card, comprising the following steps: S1. Real-time collection of operation intensity, instantaneous power consumption and temperature data of GPU cores, video memory and power supply modules, and generation of a data set; S2. Construction of a temperature-load three-dimensional thermal map based on the data set, and establishment of a dynamic mapping model for computing wave peaks and heat dissipation response based on the three-dimensional thermal map; S3. Based on the dynamic mapping model, high-density computing instructions are dispersed to different operation units according to a preset hotspot prediction-heat dissipation preloading algorithm, which is a task scheduling strategy based on thermal field prediction and is implemented by an LSTM model, capable of triggering heat dissipation preloading actions before instruction execution; a hierarchical response strategy refers to the start of differential heat dissipation measures according to temperature thresholds, which is implemented by phase change materials and fluid mechanics control, and the vertical airflow distribution is optimized for video memory burst read and write; S4. Analysis of the data set, when the core temperature reaches a first threshold, the phase change heat dissipation sheet is started to quickly absorb heat, when the core temperature reaches a second threshold, the centrifugal fan is accelerated and the air outlet angle is adjusted to form a directional air duct, and when the video memory bursts read and write, the vertical heat dissipation airflow is preferentially enhanced instead of global forced cooling.

2. The health management method of an audio integrated signal processing unit (GPU) computing board card according to claim 1, wherein, The step S2 comprises: Collecting temperature data of GPU cores, video memory and power supply module regions, recording the calculation load and power consumption parameters of each operation unit, mapping the hardware physical layout to a three-dimensional coordinate system, using inverse distance weighted interpolation algorithm to perform spatial interpolation on discrete temperature points based on temperature data, and forming a continuous three-dimensional temperature field; The temperature gradient in the three-dimensional temperature field is converted into a thermal map expressed in two dimensions of color and height by using layer rendering technology; Based on the simultaneously recorded calculation load and power consumption parameters of each operation unit, a dynamic mapping model of computing tasks and thermal map is established, and the operation instruction flow is converted into a heat flux density function H by using matrix superposition algorithm, and the coupling law of high load instruction and specific region heat flux density is determined by convolution operation with real-time temperature field.

3. The health management method of an audio integrated signal processing unit (GPU) computing board card according to claim 2, wherein, The step S2 further comprises: Through implantation of a dynamic threshold correction mechanism, the thermal field interpolation accuracy and the heat dissipation response timing are continuously optimized according to historical data, and finally the closed-loop linkage of computing task scheduling, temperature field evolution and heat dissipation resource allocation is realized; The implementation of the dynamic threshold correction mechanism comprises: (1) Constructing a heat gradient residual function based on historical temperature field data wherein is a predicted temperature value, is a measured temperature value; (2) When R continuously exceeds 5℃ for 3 cycles, automatically adjust the power parameter of the inverse distance weighted interpolation algorithm and recalculate the three-dimensional temperature field; (3) Sliding window calibration of heat dissipation response timing: the heat dissipation action trigger delay is counted in 50 ms cycles, if the standard deviation of the delay is greater than or equal to 8 μs, the preloading trigger advance is shortened by 20% step; (4) The matching relationship between interpolation accuracy and heat dissipation timing is optimized by genetic algorithm, so that the thermal gradient prediction deviation is stabilized within ±2℃.

4. The health management method of an audio integrated signal processing unit (GPU) computing board card according to claim 1, wherein, The step S3 comprises: Based on the dynamic mapping model, the coupling law of high load instruction and specific region heat flux density is determined, and the LSTM time series model is used to predict the thermal field evolution trend in the next 200 ms; When it is detected that the heat flux density in a specific area is about to exceed the standard, the system automatically activates the instruction reallocation engine, breaks down the high-density computing instructions into multiple sub-tasks, and, based on the physical core mapping relationship of the NUMA architecture and combined with the dynamic table of real-time heat dissipation margin of each computing unit, prioritizes scheduling instruction units with heat effect values ​​exceeding the threshold to areas with idle heat dissipation resources. 150μs before the actual execution of the command, the target area is preloaded for heat dissipation; the blade angle of attack of the centrifugal fan is simultaneously adjusted to 45° to form a directional acceleration airflow in the target area. By continuously optimizing the instruction dispersion granularity and heat dissipation response timing parameters through the implantation of a reinforcement learning framework, the optimal matching between computation instruction distribution and heat dissipation resources is ultimately achieved.

5. The health management method for the GPU computing board of the audio integrated signal processing machine according to claim 4, characterized in that: The heat flux density function H in the dynamic mapping model is optimized in the following way: (1) Introducing dynamic voltage frequency adjustment parameter correction coefficient wherein , is the instantaneous temperature value of the target region at the current time; (2) the modified heat flux density function An LSTM prediction model is inputted to generate a space-time correlation matrix containing temperature field gradient and heat dissipation allowance. (3) When the predicted temperature rise rate is ≥10℃ / s, the phase change heat conduction sheet pre-cooling mode of the target area is triggered simultaneously, so that the base temperature drops by 8-10℃ within 120μs; (4) The weight allocation of the DVFS parameter correction coefficient is optimized by the Q-learning algorithm, so that the heat dissipation response delay is compressed to within 45μs and the computing power loss rate is ≤0.8%.

6. The health management method for the GPU computing board of the audio integrated signal processor according to claim 1, characterized in that: The vertical heat dissipation strategy for burst read / write of video memory in step S4 includes: (1) Real-time calculation of vertical heat flow density by video memory bandwidth occupancy wherein is the video memory bandwidth occupancy, is the video memory particle temperature; (2) When V≥85%, the following compound action is initiated: ① Activate the micro piezoelectric fan array 2mm above the memory module to form a local vertical airflow with a wind speed ≥4m / s; ② Close the horizontal air duct of the adjacent power supply module and concentrate 80% of the airflow to the memory area; ③ Dynamic voltage reduction is implemented for the video memory power supply circuit, and the voltage reduction amplitude is, is the nominal voltage; (3) By optimizing the synergistic parameters of vertical airflow distribution and voltage drop through reinforcement learning framework, the memory temperature rise rate is reduced by 40% and the bit error rate is ≤1E-6.

Images