Server hardware acceleration control system based on noise temperature coupling

Abstract

The application relates to the technical field of server hardware control, and discloses a server hardware acceleration control system based on noise temperature coupling. The system comprises a multi-source noise collector, a temperature control accelerator, a noise temperature coupling analyzer and an acceleration decision master control unit. The multi-source noise collector is arranged on the surface and internal key nodes of the acceleration unit, and can capture electromagnetic radiation noise spectrum and mechanical vibration noise waveform in real time; the temperature control accelerator integrates a thermoelectric refrigeration array and a voltage frequency modulation module, receives temperature sensing data flow, generates pulse type refrigeration instructions and clock frequency correction amount according to a dynamic strategy; the coupling analyzer decomposes multi-modal noise signals, extracts energy feature vectors and constructs a coupling coefficient matrix; the acceleration decision unit receives the matrix operation, outputs a frequency scaling factor and a voltage bias to the temperature control accelerator, feeds back a failure frequency band marker result update coefficient, and realizes dynamic balance of hardware acceleration performance and stability by fusing noise and temperature coupling characteristics.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of server hardware control, in particular to a server hardware acceleration control system based on noise temperature coupling. BACKGROUND

[0002] With the rapid development of cloud computing, big data and artificial intelligence technology, the operation intensity of server hardware acceleration units continues to rise, and their stability and energy efficiency performance in high-frequency operation state have become the focus of the industry. Currently, the server hardware acceleration system mostly adopts a single-dimensional control strategy, such as adjusting the frequency only according to temperature sensing data, or filtering electromagnetic noise alone. This isolated control method cannot cope with the multi-physical field coupling problem under complex working conditions.

[0003] In actual operation, the electromagnetic radiation noise of the server hardware acceleration unit will show nonlinear characteristics with the change of operation load. When the noise energy accumulates to a certain extent, it will be coupled to the adjacent circuit nodes through electromagnetic induction, causing the signal transmission error rate to rise. At the same time, the joule heat generated by high-frequency operation will cause the temperature in the core area to rise sharply, and the uneven distribution of the temperature field will further aggravate the thermal noise of the semiconductor device, forming a "noise-temperature" vicious cycle. The existing temperature control system usually adopts a fixed threshold feedback adjustment method, which is prone to over-regulation or response lag in scenes with large temperature fluctuations, and cannot be optimized in coordination with noise suppression.

[0004] The traditional hardware acceleration control method lacks systematic analysis of multi-source noise, often only focusing on a single type of noise signal, ignoring the energy transmission relationship between mechanical vibration noise and electromagnetic noise. During the operation of the server, the vibration generated by the rotation of the fan, the mechanical movement of the hard disk reading and writing, etc. will be conducted to the acceleration unit through the structure, causing the micro-displacement of the circuit elements, changing the parasitic parameters, and then affecting the noise spectrum characteristics. The coupling effect of this multi-modal noise makes it difficult to fundamentally improve the stability of the system by using a single noise suppression method, restricting the full play of the performance of the hardware acceleration unit.

[0005] With the evolution of server hardware acceleration technology towards higher computing power and lower latency, the limitations of the existing control system in multi-physical field coordinated regulation are increasingly prominent, and an intelligent control framework that can integrate noise and temperature coupling characteristics needs to be built to achieve the dynamic balance between hardware acceleration performance and system stability. SUMMARY

[0006] The purpose of the present application is to provide a server hardware acceleration control system based on noise temperature coupling to solve the problems raised in the background art.

[0007] To achieve the above object, the application provides a server hardware acceleration control system based on noise temperature coupling, which comprises:

[0008] A multi-source noise collector is arranged on the surface and internal key circuit nodes of the server hardware acceleration unit to capture electromagnetic radiation noise spectrum features and mechanical vibration noise waveforms in real time and generate multi-modal noise original signal streams;

[0009] A temperature-controlled accelerator is integrated with a thermoelectric refrigeration array and a voltage frequency modulation module, receives temperature sensing data streams of the core area of the server hardware acceleration unit, generates pulse-type refrigeration instructions and clock frequency correction amounts according to a dynamic frequency adjustment strategy;

[0010] A noise temperature coupling analyzer is connected with the multi-source noise collector and the temperature-controlled accelerator respectively, is used for time-frequency domain decomposition of the multi-modal noise original signal streams, extraction of noise energy distribution feature vectors, synchronous fusion of the temperature sensing data streams and the clock frequency correction amounts, and construction of a noise temperature coupling coefficient matrix;

[0011] An acceleration decision master unit is bidirectionally connected with the noise temperature coupling analyzer, receives the noise temperature coupling coefficient matrix, performs convolution kernel remapping operation, outputs frequency scaling factors and voltage bias amounts of the hardware acceleration unit to the temperature-controlled accelerator, and feeds back failure frequency band mark results to the noise temperature coupling analyzer for coefficient updating.

[0012] Preferably, the multi-source noise collector comprises a multi-source signal standardization unit, and the processing flow thereof is as follows:

[0013] High-frequency carrier components and low-frequency modulation components in the electromagnetic radiation noise spectrum features are separated, wavelet packet decomposition is performed on the mechanical vibration noise waveforms, acceleration harmonic features and resonance frequency point amplitudes are extracted, and noise energy distribution feature vectors are generated;

[0014] Band pass filtering and spectrum shifting operations are performed on the high-frequency carrier components, base band reconstruction is performed on the low-frequency modulation components, and the acceleration harmonic features are converted into envelope phase sequences through Hilbert transform;

[0015] A time scale alignment channel of the electromagnetic radiation noise and the mechanical vibration noise is established, a sliding time window is used to check time sequence deviation, and a standardized noise feature tensor is generated;

[0016] Clock cycle mark points of the server hardware acceleration unit are associated, the standardized noise feature tensor is cut into noise frame sequences according to operation periods, and the noise frame sequences are output to the noise temperature coupling analyzer.

[0017] Preferably, the temperature-controlled accelerator comprises a multi-dimensional entropy value calculation module, and the operation steps thereof are as follows:

[0018] Receiving temperature gradient data stream of three thermal sensitive areas in the server hardware acceleration unit, calculating sliding variance value of temperature change rate of each area, and generating thermodynamic instability coefficient;

[0019] Extracting frequency switching record of the voltage frequency modulation module, counting frequency jump number and amplitude difference value in unit time, and generating frequency disturbance entropy value;

[0020] Fusing the thermodynamic instability coefficient and the frequency disturbance entropy value, and generating temperature control acceleration strategy index code through nonlinear weighted superposition;

[0021] According to the temperature control acceleration strategy index code, calling the current pulse width library of the thermoelectric refrigeration array, and generating a corresponding relationship table of refrigeration instruction duration and voltage bias amount.

[0022] Preferably, the noise temperature coupling analyzer comprises a delay verification rule, which is specifically:

[0023] Rule one: when the amplitude of a specific frequency band in the noise energy distribution feature vector continuously exceeds the first threshold value and the change rate of the corresponding area temperature gradient data stream is lower than the second threshold value, a frequency scaling factor down-regulation instruction is triggered;

[0024] Rule two: when the peak value of the temperature sensing data stream continuously exceeds the third threshold value and the noise frequency band energy distribution standard deviation is lower than the fourth threshold value, a voltage bias amount compensation mechanism is activated;

[0025] Rule three: when rule one and rule two are satisfied at the same time, a multi-level dynamic buffer queue is started to store the abnormal noise temperature coupling coefficient matrix, and the running state of the current server hardware acceleration unit is marked as an overload protection mode.

[0026] Preferably, the acceleration decision master control unit comprises a core feature sequence screening module, which is implemented in the following manner:

[0027] Extracting a noise temperature coupling feature subset with the highest frequency domain correlation weight from the noise temperature coupling coefficient matrix, and calculating the information entropy decay rate of each feature subset on the time axis;

[0028] According to the thermodynamic instability coefficient, the feature subsets are weighted and sorted, and the first K feature subsets are selected to construct a noise temperature core feature sequence;

[0029] The noise temperature core feature sequence is input into the convolution kernel remapping operation unit to generate a frequency scaling factor correction coefficient and a voltage bias amount compensation vector;

[0030] According to the frequency disturbance entropy value, the iteration step of the convolution kernel remapping operation is adjusted, and the optimized hardware acceleration unit control parameter is output.

[0031] Preferably, the step of the convolution kernel remapping operation comprises:

[0032] A double-channel residual learning network is constructed, a first channel inputs a noise energy distribution feature vector, and a second channel inputs a temperature gradient data stream;

[0033] The double-channel features are fused in a convolution kernel remapping operation layer, and a frequency scaling factor prediction value and a voltage bias prediction value are output;

[0034] The weighted residual accumulation value of the frequency scaling factor prediction value and the actual frequency scaling factor is calculated, and the parameters of the double-channel residual learning network are updated in reverse;

[0035] When the weighted residual accumulation value continuously decreases for three times with a rate lower than a set threshold, the convolution kernel remapping operation layer weight is frozen.

[0036] Preferably, the system further comprises a distributed topology updating module, and the operation mechanism thereof is:

[0037] A physical position topology graph of the calculation core, the cache unit and the power module in the server hardware acceleration unit is established, and the service time of an element is taken as a node attribute;

[0038] The connection weight between nodes is dynamically updated according to the temperature gradient data stream, and a thermodynamic coupling adjacency matrix is generated;

[0039] The key path propagation delay is calculated based on the thermodynamic coupling adjacency matrix, and a set of thermal noise sensitive paths is marked;

[0040] The set of thermal noise sensitive paths is mapped to the frequency domain space of the noise temperature coupling coefficient matrix, and a spatial distribution priority list of the frequency scaling factor is generated.

[0041] Preferably, the multi-source noise collector comprises an acoustic signal conversion unit, and the configuration parameters of the acoustic signal conversion unit include:

[0042] Eight groups of orthogonal mixers are arranged in the electromagnetic radiation noise collection channel, and the center frequency covers a range from megahertz to gigahertz;

[0043] A three-axis accelerometer array is arranged in the mechanical vibration noise collection channel, and the sampling rate is set to be an integer multiple of the base clock frequency of the server hardware acceleration unit;

[0044] The acoustic signal conversion unit and the thermoelectric refrigeration array share a clock synchronization signal, and noise sampling calibration is performed in the refrigeration pulse gap;

[0045] An adaptive gain controller is arranged at the output end of the acoustic signal conversion unit, and the signal amplification multiple is dynamically adjusted according to the frequency scaling factor.

[0046] Preferably, the system further comprises a cross-modal verification module, and the interaction logic thereof is:

[0047] Receive the noise temperature core feature sequence and the thermal noise sensitive path set, calculate the correlation coefficient of the path propagation delay and the noise frequency band energy attenuation;

[0048] When the correlation coefficient exceeds the dynamic threshold, trigger the partition frequency reduction instruction of the server hardware acceleration unit;

[0049] Adjust the partition frequency reduction amplitude according to the frequency scaling factor correction coefficient, and generate the phase compensation amount of the voltage bias amount;

[0050] Inject the phase compensation amount into the voltage frequency modulation module of the temperature control accelerator, and update the frequency domain distribution of the noise temperature coupling coefficient matrix synchronously.

[0051] Preferably, the system further comprises a fault prediction module, and the operation logic of the fault prediction module is:

[0052] Receive the historical change sequence of the noise temperature coupling coefficient matrix and the real-time load data stream of the server hardware acceleration unit, calculate the joint transition probability of the energy mutation of each noise frequency band and the temperature gradient anomaly through the hidden Markov model;

[0053] Based on the dynamic update record of the thermodynamic coupling adjacency matrix, extract the periodic fluctuation characteristics of the key path propagation delay;

[0054] Generate a fault probability vector by fusing the joint transition probability and the periodic fluctuation characteristics, and when any element in the fault probability vector exceeds the dynamic threshold, send a pre-order reduction instruction of the frequency scaling factor to the acceleration decision master unit;

[0055] Synchronously update the pulse width library of the thermoelectric refrigeration array of the temperature control accelerator, and increase the heat exchange intensity of the frequency band corresponding to the pre-order reduction instruction.

[0056] Compared with the prior art, the beneficial effects of the present application are:

[0057] Through the multi-source noise collector, the electromagnetic radiation noise spectrum characteristics and mechanical vibration noise waveform of the surface and internal key circuit nodes of the server hardware acceleration unit are captured in real time to generate a multi-modal noise original signal stream, breaking the limitation of traditional single noise monitoring, and can fully reflect the noise field distribution characteristics in the system running process. This multi-dimensional noise perception method enables the system to more accurately identify the source and propagation path of different types of noise, providing rich raw data support for subsequent collaborative control.

[0058] The temperature control accelerator integrates a thermoelectric refrigeration array and a voltage frequency modulation module. After receiving the temperature sensing data stream, the temperature control accelerator generates pulse-type refrigeration instructions and clock frequency correction amounts according to a dynamic frequency adjustment strategy. Compared with the traditional fixed threshold temperature control method, the temperature control accelerator can be finely adjusted according to the real-time temperature change. The pulse-type refrigeration instructions can realize dynamic allocation of refrigeration capacity, avoiding energy waste caused by continuous refrigeration. The clock frequency correction amount can adjust the operation rhythm in time when the temperature fluctuates, reducing the performance fluctuation caused by sudden temperature changes. The dynamic adjustment mechanism makes the temperature control process more adaptive, which can meet the heat dissipation demand while considering the operation efficiency.

[0059] The noise temperature coupling analyzer performs time-frequency domain decomposition on the multi-modal noise original signal stream, extracts a noise energy distribution feature vector, and synchronously fuses the temperature sensing data stream and the clock frequency correction amount to construct a coupling coefficient matrix, realizing quantitative description of the “noise-temperature” interaction relationship. Through the coupling analysis, the system can clearly master the dynamic correlation between noise energy and the temperature field, such as identifying the surge mode of noise energy in a specific temperature interval or finding the strong correlation between a certain noise frequency band and the temperature gradient, thereby providing a scientific analysis basis for the formulation of the control strategy.

[0060] The acceleration decision master control unit is bidirectionally connected with the noise temperature coupling analyzer, performs convolution kernel remapping operation after receiving the coupling coefficient matrix, outputs the frequency scaling factor and the voltage bias to the temperature control accelerator, and feeds back the failure frequency band marking result to the analyzer for coefficient updating, forming a closed-loop cooperative control mechanism. The convolution kernel remapping operation can dynamically adjust the running parameters of the hardware acceleration unit according to the change of the coupling coefficient matrix, so that the adjustment of the frequency and the voltage is more suitable for the coupling state of the noise and the temperature, realizing the dynamic balance of the performance and the stability. The feedback updating mechanism of the failure frequency band makes the coupling coefficient matrix constantly adapt to the change of the system running state, ensuring the continuous effectiveness of the control strategy.

[0061] Through the whole-process cooperation of “noise perception-temperature regulation-coupling analysis-decision feedback”, the system can effectively break the “noise-temperature” vicious cycle, suppress the multi-modal noise coupling effect, and realize efficient heat dissipation and frequency optimization of the hardware acceleration unit. The cooperative regulation of multi-source noise and the temperature field improves the stability of the server hardware acceleration unit when running at high frequency, reduces the operation interruption or performance decline caused by noise interference or high temperature, and also reduces unnecessary energy consumption, improving the hardware acceleration performance while considering the energy efficiency performance of the system. BRIEF DESCRIPTION OF DRAWINGS

[0062] Figure 1 A working principle diagram of the server hardware acceleration control system based on noise temperature coupling described in the present application;

[0063] Figure 2 Flow chart for multi-source signal standardization unit

[0064] Figure 3 Flow chart for multi-dimensional entropy value calculation module

[0065] Figure 4 Flow chart for core feature sequence screening module

[0066] Figure 5 Flow chart for distributed topology updating module DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0068] Please refer to Figure 1 The present application provides a server hardware acceleration control system based on noise temperature coupling. The system collects and analyzes multi-source noise signals and temperature data in the running process of the server hardware acceleration unit in real time, establishes a coupling relationship model of noise and temperature parameters, and dynamically adjusts the working frequency and voltage bias of the hardware acceleration unit based on the model, to realize the optimization control and reliability guarantee of the hardware acceleration performance. The specific implementation is as follows:

[0069] The server hardware acceleration control system mainly consists of a multi-source noise collector, a temperature control accelerator, a noise temperature coupling analyzer, and an acceleration decision master control unit.

[0070] The multi-source noise collector is physically deployed on the outer surface and internal key circuit node position of the server hardware acceleration unit. Its core function is to capture the electromagnetic radiation noise spectrum characteristics and mechanical vibration noise waveform generated in the running of the hardware unit in real time, and convert the captured original signals into standardized multi-modal noise original signal stream.

[0071] The temperature control accelerator is internally integrated with a thermoelectric refrigeration array and a voltage frequency modulation module. The unit receives real-time temperature sensing data stream from the core area of the server hardware acceleration unit, generates pulse-type refrigeration instructions for accurately controlling the operation of the thermoelectric refrigeration array according to the preset dynamic frequency adjustment strategy, and calculates the clock frequency correction amount needed to be applied to the hardware acceleration unit.

[0072] The noise temperature coupling analyzer is connected with the multi-source noise collector and the temperature control accelerator respectively, and its responsibility is to perform time and frequency domain joint decomposition processing on the received multi-modal noise original signal stream, and extract the feature vector representing the distribution of noise energy in the frequency band.

[0073] The analyzer synchronously fuses the temperature sensing data stream and the clock frequency correction amount information from the temperature control accelerator, and constructs a noise temperature coupling coefficient matrix reflecting the dynamic correlation between noise characteristics and temperature parameters through a specific coupling algorithm. The accelerated decision-making master control unit is bidirectionally connected with the noise temperature coupling analyzer, and the master control unit receives the noise temperature coupling coefficient matrix output by the analyzer and performs convolution kernel remapping operation on the matrix.

[0074] The operation result is output as two key control parameters: the frequency scaling factor of the hardware acceleration unit and the voltage bias, which are sent to the temperature control accelerator for execution in real time.

[0075] The failure frequency band marking result identified by the master control unit during the operation process is fed back to the noise temperature coupling analyzer for periodic updating of the coefficient matrix.

[0076] Embodiment 1: refer to Figure 2 The multi-source signal standardization unit inside the multi-source noise collector is responsible for normalizing the original collected signals. The unit includes an electromagnetic radiation noise processing channel and a mechanical vibration noise processing channel. The electromagnetic radiation noise processing channel is equipped with eight sets of orthogonal mixer circuit arrays, each set of mixer working at different center frequencies, covering the megahertz to gigahertz frequency range. When the hardware acceleration unit is running, the channel captures electromagnetic radiation signals in real time, enhances the signals through a radio frequency front-end amplifier, and then inputs them into a band-pass filter bank for frequency band division. The filtered signals enter the demodulator array to perform carrier separation operation, and the high-frequency carrier component and the low-frequency modulation component are decoupled and output. The high-frequency carrier component is shifted to the baseband through down-mixing processing, and the low-frequency modulation component is restored to the original envelope form through a baseband reconstruction filter.

[0077] The mechanical vibration noise processing channel deploys six sets of three-axis piezoelectric acceleration sensors, and the sensor array is arranged at key positions of the hardware acceleration unit shell according to a thermodynamic distribution model. The vibration waveform collected by the sensor is input into a wavelet packet decomposition processor, which extracts 25 feature sub-bands using a five-layer decomposition structure. The resonant frequency scanning algorithm is performed on a specific frequency band to identify the mechanical structure resonant frequency point and the corresponding amplitude spectrum peak. At the same time, the vibration harmonic characteristics are converted into an envelope phase sequence through a Hilbert transformer, which contains vibration amplitude change rate and phase shift information. After the above processing, a vibration feature matrix containing acceleration harmonic characteristics and resonant amplitude is generated.

[0078] The unit built-in high-precision clock synchronization system completes the cross-signal source time alignment operation. The system takes the hardware acceleration unit main clock as the reference source to generate two-way synchronous trigger signals to control the sampling time sequence of the electromagnetic radiation collection channel and the vibration collection channel. When the two groups of signals enter the time mark alignment data channel, the embedded processor automatically detects the time sequence deviation of the electromagnetic noise sampling point and the vibration data collection point. The sliding time window compensation algorithm is used to dynamically correct the maximum 2.1 microsecond deviation, and the window width is fixed at 7 milliseconds. The data after time sequence calibration is reconstructed into a four-dimensional standardized noise feature tensor, which contains four parallel data streams of electromagnetic spectrum components, modulation envelope, vibration harmonic characteristics and resonance amplitude.

[0079] After the feature tensor is reconstructed, it is bound with the operation period of the hardware acceleration unit. The special period marker detection circuit captures the processor instruction period edge in real time, and triggers the tensor divider to generate a periodic noise frame every time a complete clock period is detected. The noise frame structure includes three parts: header mark area, feature data area and check code area. The header mark area records the period start timestamp and period index number; the feature data area encapsulates the four-dimensional standardized data in the current period; the check code area includes cyclic redundancy check code and hamming error correction code. When the standardization process is completed, the unit outputs the noise frame sequence with time continuity and periodic granularity through the gigabit Ethernet interface, and transmits it to the noise temperature coupling analyzer for subsequent processing.

[0080] In the process of mechanical vibration noise processing, the sensor array is configured with a temperature drift compensation module. This module continuously monitors the change of sensor environmental temperature, and automatically activates the compensation algorithm when the temperature change exceeds 4.3 degrees Celsius. The compensation algorithm uses a polynomial regression model and dynamically corrects it by combining the pre-stored temperature-sensitivity coefficient table to ensure that the vibration feature extraction accuracy is not affected by environmental factors. The electromagnetic processing channel is provided with an anti-aliasing protection mechanism, and an elliptical filter with a stopband attenuation of 80 dB is configured to prevent high-frequency noise from aliasing down.

[0081] Multiple quality monitoring is performed during the standardization process. The signal integrity verifier detects the validity of the data of each processing node in real time, and generates an error reprocessing instruction when the data verification fails. The power spectral density monitoring unit continuously calculates the energy distribution index of the feature tensor, and triggers the sampling parameter reset process when the abnormal fluctuation exceeds the preset tolerance. The resource occupancy rate of all processing links is managed by a special monitoring chip, and the processing thread priority is dynamically adjusted according to the workload of the hardware acceleration unit to ensure the real-time requirements of the standardization process. The final output noise frame sequence meets the industrial real-time data transmission protocol specification and has a determined transmission rate and error recovery mechanism.

[0082] Example 2: see Figure 3The multi-dimensional entropy calculation module built-in the temperature control accelerator continuously receives the temperature gradient data stream of the three thermal sensitive areas of the server hardware acceleration unit. The area division is determined according to the thermal distribution characteristics of the hardware architecture, including the calculation core area, the cache area and the power management area. A distributed temperature sensing network is deployed in each area, and the calculation core area is configured with a sixteen-point thermocouple array, the cache area is equipped with an eight-channel infrared thermal imaging unit, and the power management area is provided with four groups of contact temperature sensors. The temperature data is transmitted at a sampling frequency of 200 Hz and input into the entropy calculation module through a special low-latency bus. Three independent processing channels are set inside the module to correspond to different areas, and a sliding variance calculation unit is built-in each channel. The unit uses a 256-point time window to calculate the first derivative of the temperature change rate in real time, and the variance value of the temperature derivative is calculated every 10 milliseconds. The variance calculation result is defined as the thermodynamic instability coefficient, which is stored in the dual-port memory in floating-point format, and reflects the intensity of the temperature fluctuation of the corresponding area in real time.

[0083] The historical operation record of the voltage frequency modulation module is acquired by an event-driven recorder. The recorder captures the precise timestamp of the frequency switching event, the frequency value before the jump, and the frequency value after the jump. The entropy calculation module sets a 10-second time window to statistically analyze the frequency switching behavior. Every time a frequency jump event occurs in the window, a special counter accumulates the number of jumps, and a difference calculator calculates the frequency adjustment amplitude between adjacent jump events. The frequency disturbance entropy generation unit uses the information entropy principle to process the statistical results: first, a frequency jump amplitude distribution histogram is constructed, dividing the 0-200 MHz amplitude range into 32 equal intervals; then the probability distribution of jump events in each interval is calculated; finally, the Shannon entropy formula is used to generate an entropy value index quantifying the uncertainty of frequency adjustment. The entropy value is updated every 5 seconds and output through a 12-bit precision register.

[0084] The multi-dimensional fusion processor receives the thermodynamic instability coefficient and the frequency disturbance entropy data stream. The processor uses a three-layer neural network structure to perform nonlinear weighted superposition operation: the first layer is a normalization layer, which maps the three types of thermodynamic instability coefficients and frequency disturbance entropy to the [0, 1] interval; the second layer is a feature weighting layer, which gives the calculation core area a weight of 0.45, the cache area a weight of 0.3, the power management area a weight of 0.15, and the frequency disturbance entropy a weight of 0.1; the third layer is a nonlinear activation layer, which uses the Sigmoid function to generate a 32-bit temperature control acceleration strategy index code. The index code accesses the pre-stored current pulse parameter library through a query engine, which contains 1024 parameter records, each record storing the corresponding relationship between the refrigeration pulse width and the voltage bias. The query engine uses a binary search algorithm to match the index code, outputs the pulse width control word to the thermoelectric refrigeration array, and outputs the 12-bit voltage bias to the voltage frequency modulation module.

[0085] The execution delay check rule of the noise temperature coupling analyzer contains three groups of independently running logic circuits. The trigger mechanism of rule one is realized by a double threshold comparator: the noise energy distribution feature vector input 32-channel band-pass filter bank, each filter corresponding to a specific frequency band and connected to a digital comparator. When the amplitude of a certain frequency band exceeds the preset threshold for 5 consecutive analysis periods, the comparator outputs a high level to trigger the time sequence controller. At the same time, the temperature gradient change rate of the mapping area of this frequency band (preset frequency band-area mapping table) input differential circuit generates a confirmation signal when the differential value is lower than the second threshold. The two signals trigger the frequency scaling factor down command through the AND gate, and the command is sent to the acceleration decision master unit through the priority arbiter.

[0086] The execution circuit of rule two contains a temperature peak detection unit and a noise stability analysis unit. After the temperature sensor data stream input moves average filter, the sustained over-limit is detected by the window comparator: when the temperature peak of a certain area exceeds the threshold for 3 seconds, the area over-temperature flag is generated. The energy distribution data of the associated noise frequency band (preset area-frequency mapping table) input standard deviation calculator generates a noise stability flag when the standard deviation of 20 consecutive sampling points is lower than the threshold. The two flag signals trigger the voltage bias compensation mechanism, and the compensation value is calculated by the compensation generator according to the temperature over-limit amplitude, and is output to the voltage frequency modulation module through the D / A converter.

[0087] The composite judgment of rule three is managed by the state machine controller. When the rule one trigger signal and the rule two trigger signal are simultaneously valid, the controller starts a three-level buffer queue: the first level caches the current noise temperature coupling coefficient matrix; the second level stores the historical matrix of the last five analysis periods; the third level preserves the matrix trend data. The buffer queue adopts a circular storage structure with a total capacity of 1MB. The controller sets the state register to the overload protection mode at the same time, which activates the special control strategy: sends the forced instruction of frequency reduction coefficient 0.7 to the acceleration decision master unit, and at the same time instructs the temperature control accelerator to increase the power of the thermoelectric refrigeration array to the maximum gear. The mode state indicator light displays a red warning signal on the hardware monitoring panel, which is automatically released until the coupling coefficient anomaly disappears.

[0088] Embodiment 3: refer to Figure 4 The core feature sequence screening module of the acceleration decision master unit continuously receives the noise temperature coupling coefficient matrix output by the noise temperature coupling analyzer. The matrix is a multi-dimensional data structure, with row index corresponding to different noise frequency bands and column index associated with the temperature sensing area of the server hardware acceleration unit. The screening module is built-in correlation analysis engine, which calculates the frequency domain correlation coefficient between each noise frequency band feature vector and temperature gradient data stream in the matrix. Noise frequency bands with correlation coefficient higher than the dynamic threshold are classified as high weight feature subsets. Each feature subset is assigned an initial frequency domain correlation weight value (k is the index of feature subset), which reflects the coupling strength between the noise frequency band and the temperature parameter.

[0089] The module also monitors the information entropy change characteristics of the feature subsets on the time axis. The sliding time window analyzer calculates the decay rate of the information entropy of each feature subset with a window length of 5 seconds and a step of 1 second . The decay rate is obtained by least square fitting the slope of the entropy value change within the time window. The comprehensive ranking weight of the feature subset is calculated by the following formula:

[0090]

[0091] where λ is the thermodynamic instability coefficient weight factor, and the value range is fixed at 0.6; is the frequency domain correlation weight (dimensionless); is the information entropy decay rate (bit / s); is the feature subset comprehensive ranking weight (dimensionless).

[0092] The thermodynamic instability coefficient input feature weight distributor provided by the temperature control accelerator. The distributor includes three groups of configurable gain amplifiers, respectively corresponding to the instability coefficients of the calculation core area, the cache area, and the power management area. The area coefficient mapper distributes the area thermodynamic instability coefficients to the corresponding feature subsets according to the pre-set area-frequency band association table. The feature ranking processor receives the values of all feature subsets , and performs secondary weighting ranking combined with the area thermodynamic instability coefficients. The ranking is realized by using the maximum heap data structure, and the top K feature subsets with the highest weights (K=12) are output in real time.

[0093] The selected feature subsets input the core feature sequence construction unit. The unit includes a feature reorganization processor that extracts the following elements from each feature subset: noise frequency band center frequency, frequency band width, energy mean, temperature correlation coefficient, and entropy decay rate. The reorganized data structure is packaged as a 128-bit wide feature vector, and twelve feature vectors are arranged in descending order of weight to form a noise temperature core feature sequence. The sequence header adds a time stamp and an area distribution marker, and is transmitted to the convolution kernel remapping operation unit through a 64-bit data bus.

[0094] The convolution kernel remapping operation unit is configured with a double-channel residual learning network structure. The first input channel is connected to a noise energy distribution feature vector buffer that stores noise feature data of the last eight analysis periods; the second input channel is connected to a temperature gradient data stream first-in-first-out memory with a depth of twelve groups of temperature sampling values. The double-channel data is sent to the convolution kernel remapping operation layer after being processed by a normalization layer. The operation layer includes three groups of reconfigurable processing units: the first unit performs 1×3 convolution operation to extract time dimension features; the second unit performs 3×1 convolution operation to extract frequency domain dimension features; and the third unit fuses the outputs of the previous two levels to generate a frequency scaling factor prediction value through a full connection layer and a voltage bias amount prediction value .

[0095] After the prediction value is output, it enters the residual calculation link. The actual frequency scaling factor is read from the control parameter register, and the actual voltage bias amount is sampled from the output end of the digital-to-analog converter. The residual calculator performs the following operations: frequency scaling residual , voltage bias residual . The residual weighted accumulator adopts an exponential weighted moving average algorithm: cumulative residual , where takes a fixed value of 0.85, is a voltage / frequency residual ratio coefficient fixed at 0.7.

[0096] The network parameter update controller monitors the residual cumulative value . When continues to exist, the back propagation engine is activated: the gradient calculation module calculates the partial derivatives of the loss function with respect to each network parameter according to and ; the parameter updater adjusts the convolution kernel weight using a stochastic gradient descent algorithm with momentum, with a learning rate fixed at 0.001 and a momentum coefficient of 0.9. During the network training process, the change characteristics of the residual cumulative value are continuously monitored: when and are satisfied for three consecutive iterations (n is the iteration index), the freeze signal generator outputs a high level. This signal triggers the weight latch to write all parameters of the current convolution kernel remapping operation layer to the non-volatile memory, and simultaneously disconnects the parameter update path. The freeze state indicator light displays a blue signal on the control panel, and at this time the operation unit switches to pure prediction mode operation.

[0097] The optimized hardware acceleration unit control parameters are output through independent channels: the frequency scaling factor correction coefficient generates an analog control signal through a 12-bit digital-to-analog converter, directly driving the reference clock circuit of the voltage frequency modulation module; the voltage bias compensation vector is input to a 32-bit digital potentiometer array, adjusting the DC bias voltage of the power management module. The output interface is configured with a real-time verification mechanism, which performs range verification and redundancy coding before parameter transmission. If the verification fails, it automatically switches to the safe default value output mode.

[0098] Embodiment 4: refer to Figure 5 The initial topology connection relationship is established according to the physical distance and the heat conduction path, and the connection edges are automatically established between the nodes with a distance less than 15 mm.

[0099] The temperature gradient data stream drives the topology dynamic update. The temperature sampling values of the three thermal sensitive areas are input into the topology update engine after being converted by 12-bit ADC. The engine performs weight update every 10 seconds: calculates the temperature difference between nodes , adjusts the connection weight according to the formula , where is the initial weight coefficient 0.3, is the temperature sensitive coefficient 0.05. The updated thermodynamic coupling adjacency matrix is stored in a dual-port RAM, and the matrix dimension is fixed at 6x6. The following table shows the adjacency matrix fragment after a typical update period:

[0100]

[0101] The heat noise sensitive path analysis adopts the improved Dijkstra algorithm. The path finder takes the heat source node as the starting point and the heat dissipation node as the ending point, and calculates the minimum thermal resistance path according to the adjacency matrix weight. The propagation delay calculator introduces the material thermal conductivity variable: the copper path basic delay is 0.8 ms / cm, the aluminum path is 1.2 ms / cm, and each increase of 1 unit weight adds 0.3 ms delay. The identified key path contains the following characteristics: path length greater than 30 mm, weight higher than 0.35, and delay exceeding 5 ms. The path marker encodes the sensitive path information into a 32-bit data structure, containing four fields: starting point ID, ending point ID, path weight and propagation delay.

[0102] The acoustic signal conversion unit is equipped with an eight-channel electromagnetic noise acquisition system. The center frequencies of the quadrature mixer array are configured as follows: channels 1-2 cover the 1-100MHz band, channels 3-5 cover 101-500MHz, and channels 6-8 cover 501-3000MHz. The mixer local oscillator signal is generated by a direct digital frequency synthesizer with a frequency resolution of 0.1Hz. The mechanical vibration channel uses a triaxial MEMS accelerometer with a sensor sensitivity set to 100mV / g and a range of ±50g. The sampling rate generator locks the hardware acceleration unit's base clock (100MHz) and generates a sampling clock that is an integer multiple of 200MHz.

[0103] The clock synchronization system employs a tree topology. The master clock source is a temperature-compensated crystal oscillator, which generates three co-source signals via a low-jitter clock distribution chip: the first drives the pulse controller of the thermoelectric cooling array, the second drives the acoustic signal ADC sampling clock, and the third provides a reference clock for the noise processing unit. A synchronization error monitor continuously detects the phase difference between each branch clock, automatically triggering a delay compensation circuit when it exceeds 0.5 ns. The cooling pulse gap is defined as a 1.8 μs silence period after the pulse ends. During this period, the acoustic acquisition system initiates a calibration sequence: first, all unnecessary power supplies are cut off; second, a standard test signal is injected for gain correction; and finally, background noise baseline updates are performed.

[0104] The adaptive gain controller implements dynamic range adjustment. The controller receives the frequency scaling factor from the acceleration decision-making master control unit. (Value range 0.5-1.2), the gain coefficient is determined by looking up a table: For every 0.1 dB decrease, the gain increases by 3 dB. The voltage-controlled amplifier uses digital control, with a 256-level gain adjustment range covering 20-60 dB. The signal chain is equipped with an overload protection mechanism: when the peak-to-peak value of the output signal exceeds 85% of the reference voltage, a 6 dB attenuator is automatically inserted. The output stage is equipped with an anti-aliasing filter bank, with a transition band roll-off of 120 dB / octave and a stopband rejection ratio greater than 80 dB.

[0105] The spatial mapping between the physical topology and acoustic acquisition is achieved through a coordinate transformer. This device stores the 3D model data of the hardware acceleration unit, mapping the physical location of each acoustic sensor to the nearest topology node. When the distributed topology update module marks thermally noise-sensitive paths, the association mapper automatically activates all acoustic acquisition channels within a 200μm radius of the path. The spatial filter adjusts the beamforming parameters according to the path orientation, setting the main lobe width to ±30 degrees and sidelobe suppression to no less than 25dB. In the final generated spatial distribution list of frequency scaling factors, the sensitive path association region is assigned the highest priority, and this list is transmitted in real-time to the acceleration decision control unit via a high-speed serial interface.

[0106] In embodiment 5, the cross-modal verification module continuously receives the noise temperature core feature sequence from the accelerated decision master unit and the set of hot noise sensitive paths from the distributed topology update module. The feature sequence parser extracts three key elements of each feature vector in the sequence: the frequency band center frequency value, the energy attenuation slope, and the temperature correlation degree parameter. The path set decoder parses the start and end point identifiers, weight coefficients, and propagation delay values of each sensitive path. The spatial mapping engine associates each sensitive path to a specific hardware region based on the pre-set correspondence between topology node coordinates and physical regions.

[0107] The delay-attenuation correlation analyzer establishes a double-input processing channel: the first channel inputs the sensitive path propagation delay data stream, recording the conduction time of the hot signal in the path in milliseconds; the second channel inputs the noise frequency band energy attenuation data of the associated hardware region, recording the unit in decibels per second. The correlation coefficient calculator uses a sliding window covariance algorithm, with a window width of 30 seconds, and calculates the Pearson correlation coefficient every 5 seconds. The dynamic threshold generator adjusts the judgment benchmark according to the real-time load level of the server: when the CPU utilization is less than 40%, the threshold is set to 0.65, between 40% and 70%, the threshold is set to 0.55, and when the CPU utilization is higher than 70%, the threshold is set to 0.45.

[0108] The anomaly determination circuit is activated when the absolute value of the correlation coefficient exceeds the dynamic threshold. The region locator determines the hardware partition that needs to be controlled based on the sensitive path identifier currently being analyzed. The drop amplitude calculator retrieves the frequency scaling factor correction coefficient register value, which is floating between 0.1 and 0.9 with a step size of 0.05. The partition frequency reduction amplitude is determined by the formula Δf = base frequency reduction value × (1 + correction coefficient), where the base frequency reduction value is fixed at 100MHz. The phase compensation amount generator monitors the target region clock signal and automatically generates a phase compensation amount output to the voltage frequency modulation module when it detects that the cumulative deviation of the clock period caused by frequency reduction exceeds 0.35 nanoseconds.

[0109] The phase compensation interface of the voltage frequency modulation module receives a digital compensation amount with 12-bit precision. The compensation amount is converted to an analog control voltage through digital-to-analog conversion and injected into the VCO tuning end of the phase-locked loop circuit. The matrix update trigger is started after the frequency reduction instruction is executed, sending a frequency domain resampling command to the noise temperature coupling analyzer. The analyzer activates the frequency band energy redistribution algorithm to perform energy value renormalization processing on the associated noise frequency bands of the frequency reduction region and updates the corresponding row data of the coupling coefficient matrix.

[0110] The fault prediction module configures dual data source input channels: a historical database stores a sequence of noise temperature coupling coefficient matrix changes in the last 72 hours, sampled at 5-minute intervals; a real-time data stream receives load monitoring signals of the current server hardware acceleration unit, including instruction throughput rate, cache hit rate, and power consumption values. A hidden Markov model processor defines three hidden states: normal state (S0), early warning state (S1), and fault state (S2). An observation value quantifier defines a noise frequency band energy mutation event as an instantaneous fluctuation exceeding the baseline value by 35%, and a temperature gradient anomaly event as a change exceeding 8°C within 10 seconds.

[0111] A state transition probability matrix is updated every 15 minutes. A probability calculation unit counts the transition frequencies of each event combination: for example, the transition probability from S0 to S1 is based on the event count of noise mutation accompanied by temperature anomaly. A critical path latency fluctuation analyzer performs Fourier transform to extract the main frequency components from the propagation latency history data provided by the distributed topology update module. A fluctuation feature encoder marks fluctuations with a period less than 30 seconds as high-frequency disturbances, a period of 30-120 seconds as medium-frequency fluctuations, and a period exceeding 120 seconds as low-frequency drift.

[0112] The fault probability fusion center adopts a three-level weighted processing: the first level gives a joint transition probability output by the hidden Markov model a weight of 0.6; the second level gives the high-frequency component in the latency fluctuation feature a weight of 0.3; the third level adds a weight of 0.1 according to real-time load data. The fusion result generates a six-dimensional fault probability vector, with vector elements corresponding to the six subsystems of the hardware acceleration unit. A threshold comparison array sets dynamic alert lines: the core area threshold is 0.82, the cache area is 0.75, the power area is 0.68, and the remaining areas are 0.65.

[0113] A pre-degradation instruction generation unit is triggered when any probability value exceeds the limit. An instruction encoder generates a binary control word: the high 6 bits identify the target subsystem, the middle 10 bits store the recommended frequency reduction amplitude, and the low 4 bits mark the urgency. The instruction is transmitted to the acceleration decision master control unit through a high-speed control bus, triggering a stepwise downshift of the frequency scaling factor: the first degradation executes 70% of the recommended amplitude, and subsequent evaluations are performed every 30 seconds, with the remaining amplitude added if there is no improvement.

[0114] The thermoelectric refrigeration array controller synchronously receives the frequency band identification code. The pulse width retrieval engine queries the parameter library according to the frequency band code to raise the heat exchange intensity of the corresponding area. The refrigeration pulse base width starts from the standard value of 350μs, and increases by 70μs at each level, and the maximum is not more than 800μs. The current amplitude regulator cooperates with the pulse width adjustment: 350-500μs pulse corresponds to 1.5A driving current, 501-650μs corresponds to 2.0A, and 651-800μs corresponds to 2.5A. The parameter update logger records the timestamp, frequency band code and adjustment parameter of each adjustment, and writes into the non-volatile memory for diagnostic analysis.

[0115] The system health status indicator light group reflects the processing state in real time: single green indicates normal operation; green flicker indicates cross-modal verification activation; yellow indicates low-risk warning of fault prediction module output; red indicates that the pre-reduction order instruction has been executed. The indicator signal drives the panel LED through the optocoupler isolator, and generates a system log event uploaded to the central monitoring platform at the same time. All control instructions need to be verified by the safety interlocking circuit before execution to prevent conflicts caused by concurrent operation of multiple modules.

[0116] It should be noted that, in the present document, relational terms such as first and second and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0117] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A noise temperature coupling based server hardware acceleration control system, comprising: The application relates to a noise temperature coupling analysis method for a server hardware acceleration unit. The application comprises: a multi-source noise collector arranged on the surface and inside key circuit nodes of a server hardware acceleration unit, which is used for capturing electromagnetic radiation noise spectrum features and mechanical vibration noise waveforms in real time and generating multi-modal noise original signal streams; a temperature control accelerator integrated with a thermoelectric refrigeration array and a voltage frequency modulation module, which receives temperature sensing data streams of the core area of the server hardware acceleration unit, generates pulse type refrigeration instructions and clock frequency correction amounts according to a dynamic frequency adjustment strategy; a noise temperature coupling analyzer connected with the multi-source noise collector and the temperature control accelerator, which is used for carrying out time-frequency domain decomposition on the multi-modal noise original signal streams, extracting noise energy distribution feature vectors, synchronously fusing the temperature sensing data streams and the clock frequency correction amounts and constructing a noise temperature coupling coefficient matrix; the noise temperature coupling coefficient matrix is a multi-dimensional data structure, the row index corresponds to different noise frequency bands and the column index is associated with the temperature sensing area of the server hardware acceleration unit; an acceleration decision master unit bidirectionally connected with the noise temperature coupling analyzer, which receives the noise temperature coupling coefficient matrix and executes convolution kernel remapping operation, outputs the frequency scaling factor and the voltage bias amount of the hardware acceleration unit to the temperature control accelerator and simultaneously feeds back the failure frequency band marking result to the noise temperature coupling analyzer for coefficient updating; the convolution kernel remapping operation comprises the following steps: constructing a double-channel residual learning network, a first channel inputting the noise energy distribution feature vector and a second channel inputting the temperature gradient data stream; fusing the double-channel features at the convolution kernel remapping operation layer and outputting the frequency scaling factor prediction value and the voltage bias amount prediction value; calculating the weighted residual accumulation value of the frequency scaling factor prediction value and the actual frequency scaling factor and reversely updating the double-channel residual learning network parameters; 2. The noise temperature coupling based server hardware acceleration control system of claim 1, wherein, when the weighted residual accumulation value continuously decreases for three times and the decreasing rate is lower than a set threshold, the convolution kernel remapping operation layer weight is frozen. the multi-source noise collector comprises a multi-source signal standardization unit, and the processing flow is as follows: separating high-frequency carrier components and low-frequency modulation components in the electromagnetic radiation noise spectrum features, carrying out wavelet packet decomposition on the mechanical vibration noise waveform, extracting acceleration harmonic features and resonance frequency point amplitudes and generating the noise energy distribution feature vector; performing band pass filtering and spectrum shifting operations on the high-frequency carrier components, carrying out base band reconstruction on the low-frequency modulation components and simultaneously converting the acceleration harmonic features into envelope phase sequences through Hilbert transformation; establishing a time scale alignment channel of the electromagnetic radiation noise and the mechanical vibration noise, adopting a sliding time window to check time sequence deviation and generating a standardized noise feature tensor; 3. The noise temperature coupling based server hardware acceleration control system of claim 1, wherein, associating the clock cycle marking points of the server hardware acceleration unit, cutting the standardized noise feature tensor into noise frame sequences according to the operation period and outputting the noise frame sequences to the noise temperature coupling analyzer. the temperature control accelerator comprises a multi-dimensional entropy value calculation module, and the operation steps are as follows: receiving temperature gradient data streams of three heat sensitive areas in the server hardware acceleration unit, calculating the sliding variance value of the temperature change rate of each area and generating a thermodynamic instability coefficient; Extract the frequency switching record of the voltage frequency modulation module, count the frequency jump times and amplitude difference in a unit time, and generate a frequency disturbance entropy value; Fuse the thermodynamic instability coefficient and the frequency disturbance entropy value, and generate a temperature control acceleration strategy index code through nonlinear weighted superposition; According to the temperature control acceleration strategy index code, call the current pulse width library of the thermoelectric refrigeration array to generate a corresponding relationship table of the refrigeration instruction duration and the voltage bias amount.

4. The noise temperature coupling based server hardware acceleration control system of claim 1, wherein, The noise temperature coupling analyzer includes a delay check rule, which is specifically: Rule one: when the amplitude of a specific frequency band in the noise energy distribution feature vector continuously exceeds the first threshold value and the temperature gradient data flow change rate of the corresponding area is lower than the second threshold value, a frequency scaling factor down command is triggered; Rule two: when the peak value of the temperature sensing data stream continuously exceeds the third threshold value and the associated noise frequency band energy distribution standard deviation is lower than the fourth threshold value, activate the voltage bias amount compensation mechanism; Rule three: when rule one and rule two are met at the same time, start the multi-level dynamic buffer queue to store the abnormal noise temperature coupling coefficient matrix, and mark the current server hardware acceleration unit running state as overload protection mode.

5. The noise temperature coupling based server hardware acceleration control system of claim 1, wherein, The acceleration decision master unit includes a core feature sequence screening module, which is implemented as: Extract the noise temperature coupling feature subset with the highest frequency domain correlation weight from the noise temperature coupling coefficient matrix, calculate the information entropy decay rate of each feature subset on the time axis; According to the thermodynamic instability coefficient, the feature subsets are weighted and sorted, and the top K feature subsets are selected to construct a noise temperature core feature sequence; Input the noise temperature core feature sequence into the convolution kernel remapping operation unit to generate a frequency scaling factor correction coefficient and a voltage bias amount compensation vector; According to the frequency disturbance entropy value, adjust the iteration step of the convolution kernel remapping operation, and output the optimized hardware acceleration unit control parameters.

6. The noise temperature coupling based server hardware acceleration control system of claim 1, wherein, It also includes a distributed topology update module, which operates as follows: Establish the physical location topology of the calculation core, cache unit and power module in the server hardware acceleration unit, and take the element service length as the node attribute; According to the temperature gradient data flow, dynamically update the connection weight between nodes to generate a thermodynamic coupling adjacency matrix; Based on the thermodynamic coupling adjacency matrix, calculate the critical path propagation delay, and mark the set of heat noise sensitive paths; Map the heat noise sensitive path set to the frequency domain space of the noise temperature coupling coefficient matrix to generate a spatial distribution priority list of the frequency scaling factor.

7. The noise temperature coupling based server hardware acceleration control system of claim 1, wherein, The multi-source noise collector includes an acoustic signal conversion unit, and the configuration parameters of the acoustic signal conversion unit include: The electromagnetic radiation noise collection channel is provided with eight groups of orthogonal mixers, and the center frequency covers a range from megahertz to gigahertz; The mechanical vibration noise collection channel adopts a three-axis accelerometer array, and the sampling rate is set to an integer multiple of the base clock frequency of the server hardware acceleration unit; The acoustic signal conversion unit and the thermoelectric refrigeration array share a clock synchronization signal, and perform noise sampling calibration during the refrigeration pulse gap; The output end of the acoustic signal conversion unit is configured with an adaptive gain controller, which dynamically adjusts the signal amplification multiple according to the frequency scaling factor.

8. The noise temperature coupling based server hardware acceleration control system of claim 1, wherein, It also includes a cross-modal verification module, and the interaction logic is as follows: Receive noise temperature core feature sequence and thermal noise sensitive path set, calculate the correlation coefficient of path propagation delay and noise frequency band energy attenuation; When the correlation coefficient exceeds the dynamic threshold, trigger the partition frequency reduction instruction of the server hardware acceleration unit; Adjust the partition frequency reduction amplitude according to the frequency scaling factor correction coefficient, generate the phase compensation amount of voltage bias; Inject the phase compensation amount into the voltage frequency modulation module of the temperature control accelerator, and update the frequency domain distribution of the noise temperature coupling coefficient matrix synchronously.

9. The noise temperature coupling based server hardware acceleration control system of claim 1, wherein, It also includes a fault prediction module, whose operation logic is: Receive the historical change sequence of the noise temperature coupling coefficient matrix and the real-time load data stream of the server hardware acceleration unit, calculate the joint transition probability of energy mutation and temperature gradient anomaly of each noise frequency band through hidden Markov model; Based on the dynamic update record of thermodynamic coupling adjacency matrix, extract the periodic fluctuation characteristics of the critical path propagation delay; Fusion joint transition probability and periodic fluctuation characteristics generate fault probability vector, when any element in the fault probability vector exceeds the dynamic threshold, send the pre-order reduction instruction of the frequency scaling factor to the acceleration decision master unit; Synchronously update the pulse width library of the thermoelectric refrigeration array of the temperature control accelerator, increase the heat exchange intensity of the frequency band corresponding to the pre-order reduction instruction.

Images