Chip power consumption optimization regulation method and system

Abstract

The application relates to the technical field of chip power consumption management, and discloses a chip power consumption optimization regulation method and system. The method identifies a cascaded mutation cluster sequence through multi-scale variable point detection, analyzes an evolution mode to generate a steady-state prediction result, executes a penetration regulation strategy, calculates a cascaded response deviation index, detects aging drift and triggers online micro-calibration to update power consumption genome parameters, and corrects the cascaded response model and then applies the corrected cascaded response model to subsequent regulation, so that the technical problems of regulation oscillation caused by cascaded load mutation, prediction misalignment caused by aging drift and cascaded response and aging detection conflict are solved.

Description

Technical Field

[0001] This invention relates to the field of chip power management technology, and more specifically, to a chip power optimization and control method and system. Background Technology

[0002] In the long-term operation of server environments in data centers, chips face a dual challenge when carrying out complex interactive applications: multi-level cascaded load changes triggered by user operations form multiple load steps in a short period of time, with the amplitude and interval of each change exhibiting irregular characteristics; the physical characteristics of the chip continue to drift due to aging effects such as negative bias temperature instability and hot carrier injection, which manifests as increased threshold voltage and increased leakage current, thereby changing the characteristic parameters of the power consumption-frequency response curve.

[0003] Existing power consumption control methods employ multi-scale change point detection algorithms to identify load abrupt changes and trigger control responses. They also rely on power consumption-frequency response models for steady-state prediction and control decisions, and monitor chip characteristic drift through an independent aging detection mechanism.

[0004] Existing technologies have the following drawbacks: When the interval between multi-level load mutations is short, the multi-scale change point detection algorithm identifies each mutation as an independent event and triggers a control response separately, resulting in multiple control commands accumulating and conflicting, producing an oscillating control phenomenon of first increasing the frequency, then decreasing the frequency, and then increasing the frequency again; when aging causes the power consumption-frequency response model parameters to drift, the classification of cascade evolution modes exhibits systematic deviations, and the target steady-state calculation of the penetrating control strategy becomes inaccurate; independent aging detection relies on steady-state operation data acquisition, while the drastic power consumption fluctuations during cascade mutations severely interfere with the accuracy of drift detection, and the conflicting and uncoordinated operation of the cascade response and aging adaptation mechanisms lead to a significant decrease in the power consumption control effect of the aging chip under cascade load mutation scenarios. Summary of the Invention

[0005] This invention provides a chip power consumption optimization and control method and system, which solves the technical problems in related technologies such as inaccurate power consumption response model caused by chip aging, lag in control strategy response under cascaded load change scenarios, and drift of power consumption prediction accuracy over time.

[0006] This invention discloses a chip power consumption optimization and control method, comprising the following steps: acquiring the load time series of the target chip; identifying a set of phase transition points using a multi-scale change point detection algorithm; performing time-series clustering on adjacent phase transition points with time intervals less than a preset clustering threshold to generate a cascaded mutation cluster sequence; for each cascaded mutation cluster, analyzing the evolution pattern of the load change sequence within the cluster based on power consumption-frequency response parameters stored in the power consumption genome, generating a cascaded evolution pattern label, and calculating the target load state based on the cascaded evolution pattern label to generate a steady-state prediction result; generating and executing a penetrating control strategy based on the steady-state prediction result, wherein the penetrating control strategy directly uses the predicted steady state of the cascaded mutation cluster as the control target; in After the cascaded mutation clusters converge to a steady-state operating state, the deviation between the actual power consumption measurement value and the predicted power consumption value is calculated to generate a cascaded response deviation index. Trend analysis is performed on the cascaded response deviation index sequence corresponding to multiple consecutive cascaded mutation clusters to detect continuous monotonic drift patterns and generate cascaded sensing aging drift detection results. If significant aging drift exists, an online micro-calibration program is triggered to collect a sparse calibration dataset, and the drifting parameters in the power consumption genome are updated using an incremental learning algorithm. Based on the updated power consumption genome vector, the feature threshold for determining the cascade evolution mode and the coefficient parameters of the steady-state prediction formula are recalculated to generate a corrected cascaded response model. The corrected cascaded response model is then applied to subsequent cascaded mutation clusters to generate control strategies.

[0007] Furthermore, the multi-scale change point detection algorithm includes: setting a set of time window scales; for each scale, calculating the change amplitude of the mean load within the window using a sliding window method on the load time series; when the change amplitude exceeds the detection threshold corresponding to the scale, recording that moment as a candidate phase transition point; performing time alignment and confidence-weighted fusion on the candidate phase transition points detected at each scale, and outputting a set of phase transition points; wherein, the time alignment refers to unifying the time positions of candidate phase transition points with a time difference less than the minimum window scale to the time median of the group of candidate phase transition points; the confidence-weighted fusion refers to counting the number of times the candidate phase transition point is detected at each scale, using the ratio of the number of detections to the total number of scales as the confidence score, and retaining candidate phase transition points whose confidence scores exceed a preset confidence threshold.

[0008] Furthermore, the cascaded evolution mode labels include monotonically increasing, monotonically decreasing, first increasing then decreasing, and oscillating types. The determination rules are as follows: if the change direction of all phase transition points within a cluster is positive and the amplitude increases or remains constant, it is determined to be monotonically increasing; if the change direction of all phase transition points within a cluster is negative and the amplitude increases or remains constant, it is determined to be monotonically decreasing; if the phase transition points within a cluster exhibit a change sequence of first positive and then negative, it is determined to be first increasing then decreasing; if the change direction of the phase transition points within a cluster alternates more than a preset number of times, it is determined to be oscillating. Among them, the change direction is determined based on the difference between the average load before and after the phase transition point, with a positive difference indicating a positive change and a negative difference indicating a negative change.

[0009] Furthermore, the steady-state prediction result is calculated as follows: for the monotonically increasing type, the steady-state prediction value is the current load plus the sum of all positive amplitudes within the cluster; for the monotonically decreasing type, the steady-state prediction value is the current load minus the sum of all negative amplitudes within the cluster; for the type that first increases and then decreases, the steady-state prediction value is the current load plus the sum of positive amplitudes and then minus the sum of negative amplitudes; for the oscillating type, the steady-state prediction value is the average load at the terminal phase transition points within the cluster, and the number of terminal phase transition points is taken as the last 30% to 50% of the total number of phase transition points within the cluster.

[0010] Furthermore, the generation of the penetration-type regulation strategy includes: querying the power consumption-frequency mapping relationship in the power consumption genome based on the steady-state prediction results, traversing the discrete frequency ranges supported by the chip, calculating the predicted power consumption value that meets the load requirements at each frequency range, and selecting the frequency range with the smallest predicted power consumption value as the target frequency value; determining the target voltage value based on the target frequency value and the voltage-frequency constraint relationship; when the difference between the target voltage value and the current voltage value exceeds the safety threshold, decomposing the regulation process into several intermediate stages, with the voltage adjustment amplitude of each intermediate stage not exceeding the safety threshold, and setting a stabilization waiting time between each intermediate stage.

[0011] Furthermore, the calculation of the cascade response deviation index includes: dividing the difference between the actual power consumption measurement value and the predicted power consumption value by the predicted power consumption value to obtain a dimensionless percentage deviation value; wherein, the determination condition for steady-state operation is: within a preset observation window, the load fluctuation amplitude is less than the fluctuation threshold, and no new phase transition point is detected; the length of the preset observation window is 2 to 5 times the time span of the cascade mutation cluster, and the fluctuation threshold is 3% to 8% of the steady-state predicted load value.

[0012] Furthermore, the detection of a persistent monotonic drift pattern includes: calculating the linear regression slope of the cascade response deviation index sequence; if the absolute value of the slope exceeds a slope threshold and the regression fit exceeds a fit threshold, a persistent monotonic drift pattern is determined to exist; the Mann-Kendall trend test is used to distinguish between aging drift and random fluctuations; when the significance level is less than a preset significance threshold, the detected trend is determined to be aging drift; wherein the regression fit is characterized by the coefficient of determination, the fit threshold is 0.7 to 0.9, and the preset significance threshold is 0.05 to 0.1.

[0013] Furthermore, the online micro-calibration program includes: triggering the process during a time period in which the chip load rate is lower than a preset low load threshold and the duration exceeds a preset duration threshold; selecting representative sampling points in the voltage-frequency operating space according to the boundary priority principle; measuring the stable actual power consumption value at each sampling point to form a sparse calibration dataset; wherein, the boundary priority principle refers to dividing the voltage-frequency operating space into four boundary quadrants: high voltage high frequency region, low voltage low frequency region, high voltage low frequency region, and low voltage high frequency region; selecting sampling points in each boundary quadrant; and selecting a small number of sampling points in the central region.

[0014] Furthermore, the incremental learning algorithm includes: calculating the prediction error distribution on each parameter dimension based on the sparse calibration dataset and the current power genome; identifying parameter dimensions whose prediction errors exceed an error threshold and marking them as drift parameters; calculating update values ​​for the drift parameters based on the sparse calibration dataset using the least squares method; combining the updated values ​​of the drift parameters with the original values ​​of the non-drift parameters to generate an updated power genome vector; wherein, an upper limit threshold for the parameter update magnitude is set, and when the calculated update magnitude exceeds the upper limit threshold, the update magnitude is limited to the threshold boundary value.

[0015] This invention discloses a chip power consumption optimization and control system, comprising: a cascade mutation cluster identification module, used to acquire the load time series of the target chip, identify the phase transition point set using a multi-scale change point detection algorithm, perform time-series clustering processing on adjacent phase transition points, and generate a cascade mutation cluster sequence; an evolutionary pattern analysis module, used to analyze the evolutionary pattern of the load change sequence within the cluster based on the power consumption-frequency response parameters in the power consumption genome, and generate cascade evolutionary pattern labels and steady-state prediction results; a penetrating control module, used to generate and execute a penetrating control strategy based on the steady-state prediction results; a deviation calculation module, used to calculate the cascade response deviation index after the cascade mutation cluster converges to a steady state; an aging drift detection module, used to perform trend analysis on the cascade response deviation index sequence and generate cascade-aware aging drift detection results; an online micro-calibration module, used to collect a sparse calibration dataset and update the drifting parameters in the power consumption genome using an incremental learning algorithm; and a model correction module, used to recalculate the feature thresholds and coefficient parameters based on the updated power consumption genome vector to generate a corrected cascade response model.

[0016] This invention identifies phase transition points at different time granularities using a multi-scale change point detection algorithm and performs temporal clustering on adjacent phase transition points with time intervals less than a preset clustering threshold to generate a cascaded mutation cluster sequence. This integrates multiple load mutations within a short interval into a unified processing unit, avoiding the accumulation and conflict of multiple control commands caused by independent responses to each load mutation, thus solving the technical problem of oscillation control. By performing aging drift detection based on the steady-state power consumption deviation after the convergence of the cascaded mutation clusters, the drift observation window is set during steady-state operation rather than during cascaded mutations, avoiding the interference of drastic power consumption fluctuations during cascaded mutations on drift detection, thus solving the technical problem of conflict between cascaded response and aging detection. By updating the drift parameters in the power consumption genome using an incremental learning algorithm and recalculating the feature threshold for cascaded evolution mode determination and the coefficient parameters of the steady-state prediction formula based on the updated power consumption genome vector, the parameters of the cascaded response model are kept consistent with the current aging state of the chip, solving the technical problems of cascaded evolution mode classification bias and steady-state prediction inaccuracy caused by aging. This achieves the technical effect of improving the accuracy and stability of power consumption control of aging chips under cascaded load mutation scenarios. Attached Figure Description

[0017] Figure 1 This is a flowchart of the chip power consumption optimization and control method provided in the embodiments of the present invention. Detailed Implementation

[0018] According to an embodiment of this invention, a chip power consumption optimization and control method is provided. The chip power consumption optimization and control method runs on the power management controller of a data center server, collects load index data through a processor performance monitoring unit, and executes control commands through a voltage and frequency adjustment interface.

[0019] like Figure 1 As shown, the chip power consumption optimization and control method provided in this embodiment includes the following steps: Step 1: Obtain the load time series and identify cascaded mutation cluster sequences. The load time series of the target chip within a preset sampling period is obtained. The load time series is analyzed at different time granularities using a multi-scale change point detection algorithm to identify the set of phase transition points where the load changes significantly. Temporal clustering is performed on adjacent phase transition points in the set of phase transition points whose time interval is less than a preset clustering threshold. Phase transition points that meet the clustering conditions are grouped into the same cluster to generate a cascaded mutation cluster sequence.

[0020] It should be noted that the input to the above multi-scale change point detection algorithm is a load time series, and the output is a set of phase transition points. Specifically, the time window scale set is defined. ,in Indicates the first The length of a time window This is the sequence number of the time window. For each time window scale, the total number of time window scales is given. The change in the mean load within the time series is calculated using a sliding window. When the change exceeds the detection threshold corresponding to that scale, that moment is recorded as a candidate phase transition point. The candidate phase transition points detected at each scale are then time-aligned and weighted by confidence to output the final set of phase transition points.

[0021] Furthermore, the aforementioned time alignment refers to the process of mapping candidate phase transition points with similar time positions detected by windows of different scales to a unified time point. Specifically, for candidate phase transition points with a time difference smaller than the minimum window scale, the time position of the candidate phase transition points is unified to the time median of the group of candidate phase transition points.

[0022] Furthermore, the aforementioned confidence-weighted fusion refers to the process of screening candidate phase transition points based on the reliability of detection results at each scale. Specifically, for each aligned candidate phase transition point, the number of times the candidate phase transition point is detected at each scale is counted, and the ratio of the number of detections to the total number of scales is used as the confidence score of the candidate phase transition point. Candidate phase transition points with confidence scores exceeding a preset confidence threshold are retained as elements of the final phase transition point set. The typical preset confidence threshold is set to 0.5 to 0.7 to ensure that the output phase transition points are consistent across multiple time granularities.

[0023] Furthermore, the aforementioned detection thresholds are adaptively set according to the time window scale. Larger time windows correspond to higher detection thresholds, while smaller time windows correspond to lower detection thresholds, thereby balancing detection sensitivity and noise resistance at different time granularities.

[0024] It should be noted that the above-mentioned temporal clustering processing refers to the process of clustering and grouping based on the time interval between phase transition points. Specifically, for any two adjacent phase transition points in the set of phase transition points, if the time interval between the adjacent phase transition points is less than a preset clustering threshold... If these two adjacent phase transition points are indeed related, they are grouped into the same cascaded mutation cluster; through transitivity expansion, a complete cascaded mutation cluster partition is formed. The above-mentioned preset clustering threshold... It can be configured according to the response time requirements of the application scenario, for example, it can be set to a multiple of the execution cycle of a single control command.

[0025] Step 2: Analyze the evolutionary patterns of cascaded mutation clusters and generate steady-state prediction results. For each cascaded mutation cluster, the occurrence time, load change amplitude, and change direction of each phase transition point within the cluster are extracted. Based on the power-frequency response parameters stored in the current power consumption genome, the evolution mode of the load change sequence formed by each phase transition point within the cluster is analyzed. According to the characteristics of the evolution mode, the cascaded mutation cluster is determined to be one of the following: monotonically increasing, monotonically decreasing, first increasing then decreasing, or oscillating, and a cascaded evolution mode label is generated. Based on the cascaded evolution mode label and the power-frequency response parameters, the target load state after the cascaded mutation cluster converges is calculated, and the final steady-state prediction result is generated.

[0026] It should be noted that the aforementioned power consumption genome refers to a parameter vector describing the chip's power consumption characteristics, including parameters such as static power consumption coefficient, dynamic power consumption coefficient, frequency response slope, and voltage sensitivity index. The power consumption genome obtains initial values ​​through a comprehensive calibration procedure during the initial chip deployment phase and is updated based on aging conditions during operation.

[0027] Furthermore, the aforementioned comprehensive calibration procedure refers to the process of systematically measuring the chip power consumption characteristics in the early stages of chip deployment. Specifically, a uniformly distributed sampling point grid is selected within a preset voltage-frequency operating space, and corresponding voltage and frequency values ​​are set at each sampling point. A standard load mode is applied, and the power consumption value after stabilization is measured to form a three-dimensional mapping dataset of voltage-frequency-power consumption. By fitting the three-dimensional mapping dataset with a power consumption model, parameters such as static power consumption coefficient, dynamic power consumption coefficient, frequency response slope, and voltage sensitivity index are extracted and assembled into an initial power consumption genome vector.

[0028] It should be noted that the determination rules for the above evolutionary modes are as follows: if the change direction of all phase transition points in the cluster is positive and the amplitude increases or remains constant, it is determined to be a monotonically increasing type; if the change direction of all phase transition points in the cluster is negative and the amplitude increases or remains constant, it is determined to be a monotonically decreasing type; if the phase transition points in the cluster show a change sequence of first positive and then negative, it is determined to be an first increasing and then decreasing type; if the change direction of the phase transition points in the cluster changes alternately more than a preset number of times, it is determined to be an oscillating type.

[0029] Furthermore, the determination of the direction of change is based on the difference between the average load before and after the phase transition point. A positive difference indicates a positive change, a negative difference indicates a negative change, and the absolute value of the difference indicates the magnitude of the change.

[0030] Furthermore, the aforementioned preset number of times is dynamically adjusted according to the time span of the cascaded mutation clusters. Cascaded mutation clusters with a longer time span correspond to a higher preset number of times threshold, while cascaded mutation clusters with a shorter time span correspond to a lower preset number of times threshold, thereby avoiding misjudging short-term disturbances as oscillating evolutionary modes.

[0031] It should be noted that the above steady-state prediction results are calculated as follows: for monotonically increasing type, the steady-state prediction value is the current load plus the sum of all positive amplitudes within the cluster; for monotonically decreasing type, the steady-state prediction value is the current load minus the sum of all negative amplitudes within the cluster; for type that first increases and then decreases, the steady-state prediction value is the current load plus the sum of positive amplitudes and then minus the sum of negative amplitudes; for oscillating type, the steady-state prediction value is the average load at several phase transition points at the end of the cluster.

[0032] Furthermore, in the above-mentioned oscillating steady-state prediction, the number of several phase transition points at the end is determined based on the total number of phase transition points within the cluster. Specifically, the last 30% to 50% of the total number of phase transition points within the cluster is taken as the final sampling interval, thereby eliminating the impact of the violent fluctuations in the early stage of oscillation on the steady-state prediction.

[0033] Step 3: Generate and execute a penetration-based control strategy. Based on the steady-state prediction results and the current chip operating state, calculate the control path that directly transitions from the current voltage and frequency state to the target voltage and frequency state that matches the predicted steady state; skip the step-by-step response of the intermediate transition state and generate a penetrating control strategy that includes the target voltage value and the target frequency value; send the penetrating control strategy to the chip's voltage and frequency adjustment interface for execution, and simultaneously record the control execution trajectory and the actual power consumption response sequence.

[0034] It should be noted that the aforementioned penetrating regulation strategy refers to directly using the predicted steady state of the cascaded mutation cluster as the regulatory target, rather than responding to each phase transition point within the cluster individually. The generation process of the penetrating regulation strategy includes: querying the power-frequency mapping relationship in the power consumption genome based on the steady-state prediction results, and determining the target frequency value that satisfies the predicted steady-state load requirements and has optimal power consumption. Determine the target voltage value based on the target frequency value and the voltage-frequency constraint relationship. The target voltage and target frequency values ​​are encapsulated into a penetration control strategy.

[0035] Furthermore, the aforementioned power consumption-frequency mapping relationship refers to the correspondence between the load requirements defined by the power consumption genome parameters and the optimal operating frequency. Specifically, for a given steady-state predicted load value, the discrete frequency ranges supported by the chip are traversed, and the predicted power consumption value that meets the load requirements under each frequency range is calculated using the dynamic power consumption coefficient and static power consumption coefficient in the power consumption genome. The frequency range with the smallest predicted power consumption value is selected as the target frequency value.

[0036] Furthermore, the aforementioned voltage-frequency constraint relationship is defined by the safe operating boundary provided by the chip manufacturer, ensuring that the voltage value corresponding to any frequency value meets the circuit timing and reliability requirements.

[0037] In this embodiment, to ensure system stability during the control process, the implementation of the penetrating control strategy adopts a phased transition approach. Specifically, when the difference between the target voltage value and the current voltage value exceeds a safety threshold, the control process is decomposed into several intermediate stages. The voltage adjustment amplitude of each intermediate stage does not exceed the safety threshold, and a stabilization waiting time is set between each intermediate stage. This achieves the goal of the penetrating control strategy while avoiding system anomalies caused by sudden voltage changes.

[0038] Furthermore, the aforementioned safety threshold is set according to the voltage regulator response characteristics of the chip, with a typical range of 5% to 10% of the rated voltage for a single adjustment. The aforementioned stabilization waiting time is set according to the voltage regulator's setup time, with a typical range of 10 microseconds to 100 microseconds.

[0039] Step 4: Calculate the cascade response deviation index After the cascaded mutation cluster converges to a steady-state operating state, the actual power consumption measurement value of the chip during steady-state operation is obtained; the actual power consumption measurement value is compared with the predicted power consumption value corresponding to the steady-state prediction result generated in step 2, and the deviation between the two is calculated; a cascaded response deviation index is generated based on the deviation value. .

[0040] It should be noted that the formula for calculating the above-mentioned cascade response deviation index is as follows:

[0041] in, This represents the actual power consumption measurement value. This represents the predicted power consumption value calculated based on steady-state prediction results and power consumption genome analysis. The aforementioned cascade response deviation index... It is a dimensionless percentage value, eliminating the influence of the absolute value of power consumption.

[0042] Furthermore, the above predicted power consumption values The calculation process is as follows: the target load state corresponding to the steady-state prediction result generated in step 2 is input into the power calculation model in the power consumption genome. The power calculation model calculates the theoretical power consumption value under steady-state operation as the predicted power consumption value based on the current static power consumption coefficient, dynamic power consumption coefficient, target frequency value and target voltage value.

[0043] It should be noted that the above-mentioned steady-state operation condition is determined as follows: within the preset observation window, the load fluctuation amplitude is less than the fluctuation threshold, and no new phase transition point is detected.

[0044] Furthermore, the length of the aforementioned preset observation window is set to 2 to 5 times the time span of the cascaded mutation cluster, and the aforementioned fluctuation threshold is set to 3% to 8% of the steady-state predicted load value, thereby ensuring that the actual power consumption measurement value collected reflects the true steady-state operating characteristics.

[0045] Step 5: Detect aging drift and generate cascaded sensing aging drift detection results. The cascade response deviation index sequence corresponding to multiple consecutive cascaded mutation clusters is obtained, and trend analysis is performed on the cascade response deviation index sequence. The presence of a continuous monotonic drift pattern related to aging in the cascade response deviation index sequence is detected, and cascade sensing aging drift detection results are generated.

[0046] It should be noted that the aforementioned persistent monotonic drift pattern refers to the cascade response bias index exhibiting a continuously increasing or decreasing trend across multiple consecutive observations, with the number of such occurrences exceeding a preset drift confirmation threshold. Specifically, the linear regression slope is calculated for the cascade response bias index sequence. ,like If the slope threshold is exceeded and the regression fit exceeds the fit threshold, then a persistent monotonic drift pattern is determined to exist.

[0047] Furthermore, the slope of the linear regression above is calculated using the least squares method, with the observation order of the cascade response bias index sequence as the independent variable and the cascade response bias index value as the dependent variable, to fit a univariate linear equation. The slope coefficient of the univariate linear equation is then obtained. The sign of the slope coefficient reflects the drift direction, and the absolute value of the slope coefficient reflects the drift rate.

[0048] Furthermore, the regression fit is characterized by the coefficient of determination, which ranges from 0 to 1. The closer the coefficient of determination is to 1, the more significant the linear trend is, and the closer the coefficient of determination is to 0, the more dispersed the data points are. The typical fit threshold is 0.7 to 0.9 to ensure that the detected continuous monotonic drift pattern is statistically significant.

[0049] Furthermore, the aforementioned drift confirmation threshold is set according to the expected aging rate of the chip. A lower drift confirmation threshold is set for aging-sensitive chips to improve detection sensitivity, while a higher drift confirmation threshold is set for aging-stable chips to reduce the false positive rate.

[0050] Furthermore, the aforementioned slope threshold is set based on the allowable drift range of the power consumption genome parameters, with a typical value of no more than 1% to 3% change in deviation for each observation.

[0051] In this embodiment of the application, in order to distinguish between aging drift and random fluctuations, the Mann-Kendall trend test is used in the trend analysis process. The input of the Mann-Kendall trend test is the cascade response deviation index sequence, and the output is the test statistic and significance level. Only when the significance level is less than the preset significance threshold is the detected trend judged as aging drift, thereby reducing the probability of misjudgment.

[0052] Furthermore, the aforementioned preset significance threshold is set to 0.05 to 0.1, corresponding to a confidence level of 95% to 90%, to ensure that the trend identified as aging drift is statistically significant.

[0053] Step 6: Perform online micro-calibration and update power consumption genomic parameters. If the cascaded sensing aging drift detection results indicate significant aging drift, an online micro-calibration program is triggered when the chip enters a low-load operating phase. The actual power consumption response of the chip is measured at a limited number of voltage and frequency sampling points to generate a sparse calibration dataset. An incremental learning algorithm is used to update the drifted parameters in the power genome based on the sparse calibration dataset, while retaining the original values ​​of the parameters that have not drifted, to generate an updated power genome vector.

[0054] It should be noted that the aforementioned low-load runtime period refers to the time period during which the chip load rate is lower than the preset low-load threshold and the duration exceeds the preset duration threshold. Selecting the low-load runtime period to execute the online micro-calibration program is to reduce the impact of the online micro-calibration process on normal business operations.

[0055] Furthermore, the aforementioned preset low load threshold is set to 20% to 40% of the rated load, and the aforementioned preset continuous threshold is set to 1.5 to 2 times the time required to complete one online micro-calibration procedure, typically 30 to 120 seconds, thereby ensuring that there is a sufficient time window for data acquisition during the online micro-calibration procedure.

[0056] It should be noted that the above sparse calibration dataset was acquired as follows: several representative sampling points were selected in the voltage-frequency operating space, and corresponding voltage and frequency values ​​were set at each representative sampling point. The actual power consumption value after stabilization was measured to form a set of voltage-frequency-power consumption triples. The selection of representative sampling points followed the boundary priority principle, prioritizing the acquisition of data at the boundaries and inflection points of the voltage-frequency operating space.

[0057] Furthermore, the number of representative sampling points is determined based on the parameter dimensions of the power consumption genome, typically taking a value that is 2 to 4 times the number of parameter dimensions. This ensures that the number of representative sampling points is sufficient to support the statistical reliability of parameter updates, while controlling the time overhead of the online micro-calibration program.

[0058] Furthermore, the aforementioned boundary priority principle refers to the strategy of prioritizing the selection of sampling points in the boundary regions within the voltage-frequency operating space. Specifically, the voltage-frequency operating space is divided into four boundary quadrants: high voltage high frequency region, low voltage low frequency region, high voltage low frequency region, and low voltage high frequency region. Several sampling points are selected in each boundary quadrant, while a small number of sampling points are selected in the central region of the voltage-frequency operating space, forming a sparse sampling point distribution that covers both boundary and central features. This sampling strategy can capture the key morphological features of the power consumption characteristic curve with a smaller number of sampling points.

[0059] It should be noted that the input to the above incremental learning algorithm is the sparse calibration dataset and the current power genome, and the output is the updated power genome vector. The incremental learning algorithm only updates a portion of the parameters in the power genome based on the newly acquired sparse calibration dataset, rather than re-performing the entire calibration process. Specifically, the incremental learning algorithm includes the following sub-steps: Step 601: Based on the sparse calibration dataset and the current power consumption genome, calculate the prediction error distribution across each parameter dimension. Specifically, for each voltage-frequency-power consumption triplet in the sparse calibration dataset, calculate the predicted power consumption value under that voltage and frequency state using the parameters in the current power consumption genome. Record the difference between the predicted power consumption value and the actual measured power consumption value as the prediction error. Group and statistically analyze the prediction errors of each voltage-frequency-power consumption triplet according to the parameter dimension to obtain the prediction error distribution across each parameter dimension.

[0060] Furthermore, the grouping of the above parameter dimensions is based on the contribution of each parameter in the power consumption calculation model to power consumption prediction. The static power consumption coefficient affects the prediction error of the constant power consumption part, the dynamic power consumption coefficient affects the prediction error of the load-related power consumption part, the frequency response slope affects the power consumption prediction error caused by frequency changes, and the voltage sensitivity index affects the power consumption prediction error caused by voltage changes.

[0061] Step 602: Identify parameter dimensions whose prediction errors exceed the error threshold and mark them as drift parameters. Specifically, calculate the mean and standard deviation of the prediction error distribution for each parameter dimension. If the absolute value of the mean exceeds the preset error threshold, then the parameter dimension is marked as a drift parameter.

[0062] Furthermore, the aforementioned error thresholds are set based on the power consumption measurement accuracy and parameter sensitivity. Lower error thresholds are set for highly sensitive parameters to improve detection accuracy, while higher error thresholds are set for low-sensitivity parameters to avoid over-updating. The typical error threshold range is 5% to 15% of the predicted power consumption value.

[0063] Step 603: For the parameter dimensions labeled as drift parameters, calculate the updated values ​​based on the sparse calibration dataset using the least squares method. Specifically, for each drift parameter, extract a subset of voltage-frequency-power consumption triples related to the drift parameter from the sparse calibration dataset, construct a regression equation with the drift parameter as the independent variable and the actual power consumption as the dependent variable, and solve the regression equation using the least squares method to obtain the updated values ​​of the drift parameters.

[0064] Step 604: Combine the updated values ​​of the drift parameters with the original values ​​of the non-drift parameters to generate the updated power consumption genome vector.

[0065] In this embodiment, to improve the robustness of the incremental learning algorithm, a regularization constraint is introduced during the parameter update process. Specifically, an upper limit threshold is set for the parameter update magnitude. When the calculated update magnitude exceeds the upper limit threshold, the update magnitude is limited to the threshold boundary value, thereby preventing drastic parameter changes caused by anomalies in the sparse calibration dataset.

[0066] Furthermore, the above-mentioned update magnitude is defined as the ratio of the absolute value of the difference between the updated parameter value and the parameter value before the update to the parameter value before the update. The above-mentioned threshold boundary value refers to the adjustment of the updated parameter value to the result of adding or subtracting the product of the upper threshold and the parameter value before the update when the update magnitude exceeds the upper threshold, keeping the update direction unchanged but limiting the update magnitude.

[0067] Furthermore, the aforementioned upper limit threshold Based on the chip's aging rate and calibration cycle, a higher upper limit threshold is set for chips that age rapidly and have long calibration cycles, while a lower upper limit threshold is set for chips that age slowly and have short calibration cycles. The typical value is 10% to 30% of the initial parameter value.

[0068] Step 7: Correct the parameters of the cascade response model Based on the updated power genomic vector, the feature thresholds used in the cascade evolution mode determination process are recalculated; based on the updated power genomic vector, the coefficient parameters in the steady-state prediction formula are recalculated; the updated feature thresholds and coefficient parameters are written into the cascade response model to generate the corrected cascade response model.

[0069] It should be noted that the aforementioned feature thresholds include the amplitude threshold, direction threshold, and oscillation count threshold used in evolutionary pattern determination. These feature thresholds are related to the slope and nonlinearity of the power consumption-frequency response curve and need to be synchronously corrected when the power consumption genome parameters are updated.

[0070] Furthermore, the aforementioned amplitude threshold is recalculated based on the updated dynamic power consumption coefficient. When the dynamic power consumption coefficient increases, the amplitude threshold is increased accordingly to avoid misjudging normal load fluctuations as phase transition points. When the dynamic power consumption coefficient decreases, the amplitude threshold is decreased accordingly to improve the sensitivity of sudden change detection.

[0071] Furthermore, the aforementioned direction determination threshold is recalculated based on the updated static power consumption coefficient. When the proportion of static power consumption increases, the direction determination threshold needs to be increased to eliminate the interference of static power consumption fluctuations. When the proportion of static power consumption decreases, the direction determination threshold can be decreased to enhance the direction recognition capability.

[0072] It should be noted that the coefficient parameters in the above steady-state prediction formula include the proportionality coefficient and bias coefficient in the steady-state calculation formula corresponding to each evolution model. Let the general form of the steady-state prediction formula be:

[0073] in, This represents the predicted power consumption value. Indicates the dynamic power consumption coefficient. Indicates the frequency index. Indicates the voltage index. Indicates static power consumption. Indicates the target frequency. This represents the target voltage. When the power consumption genome is updated, the above coefficient parameters need to be recalculated based on the new parameter values.

[0074] Furthermore, the recalculation process for the aforementioned coefficient parameters is as follows: extract the corresponding parameter values ​​from the updated power consumption genome vector, directly replace the coefficient parameters in the steady-state prediction formula, and for parameter drift caused by aging, adjust the dynamic power consumption coefficient. Static power consumption typically shows an increasing trend. The frequency index typically shows an increasing trend. and voltage index Relatively stable.

[0075] Step 8: Apply the corrected cascade response model to generate a control strategy. For the subsequently detected cascaded mutation clusters, an evolutionary pattern analysis and steady-state prediction are performed using a corrected cascaded response model. Based on the steady-state prediction results of the corrected cascaded response model, a penetrating control strategy matching the current chip aging state is generated and executed.

[0076] In this embodiment, to achieve continuous optimization of the control strategy, steps 4 to 8 form a closed-loop iterative process. Specifically, after each execution of the penetrating control strategy, a cascaded response deviation index is calculated and accumulated into a cascaded response deviation index sequence. When the accumulated cascaded response deviation index sequence meets the drift detection condition, a new round of online micro-calibration program and cascaded response model correction is triggered, thereby enabling the penetrating control strategy to continuously adapt to the aging and evolution process of the chip.

[0077] Furthermore, the triggering conditions for the above-mentioned closed-loop iterative process include two types: the first type is time-triggered, that is, drift detection and cascade response model correction are automatically performed every preset calibration period, with a typical calibration period of 24 hours to 168 hours; the second type is event-triggered, that is, when the length of the accumulated cascade response deviation index sequence reaches the preset length and significant aging drift is detected, the cascade response model correction is immediately triggered, thereby achieving timely response during the rapid aging stage of the chip.

[0078] This implementation method identifies phase transition points at different time granularities using a multi-scale change point detection algorithm, and performs temporal clustering processing on adjacent phase transition points with time intervals less than a preset clustering threshold to generate a cascaded mutation cluster sequence. This integrates multiple load mutations within a short interval into a unified processing unit, thus avoiding the accumulation and conflict of multiple control commands caused by independent responses to each load mutation, thereby solving the problem of oscillation control.

[0079] This implementation method detects aging drift by performing steady-state power consumption deviation based on the convergence of cascaded mutation clusters. By setting the drift observation window during steady-state operation rather than during cascaded mutations, the interference of drastic power consumption fluctuations during cascaded mutations on drift detection is avoided, thus solving the problem of conflict between cascaded response and aging detection.

[0080] This implementation updates the drift parameters in the power consumption genome using an incremental learning algorithm, and recalculates the feature threshold for cascade evolution mode determination and the coefficient parameters of the steady-state prediction formula based on the updated power consumption genome vector. Therefore, the parameters of the cascade response model are kept consistent with the current aging state of the chip, thereby solving the problems of cascade evolution mode classification bias and steady-state prediction inaccuracy caused by aging.

[0081] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

[0082] A large cloud computing data center has deployed a server cluster based on Intel Xeon series processors. Server SVR-DC3-A127 hosts an online video transcoding service and has been running continuously for 42 months. The typical operating scenario for this server is as follows: after a user uploads a video file, the system sequentially triggers multiple tasks, including file reception, format parsing, video decoding, multi-resolution transcoding, quality detection, storage writing, cache updating, and CDN distribution notification. These tasks have data dependencies, creating a cascading load fluctuation pattern.

[0083] The server is equipped with a Xeon E5-2680 v4 chip, with a rated TDP of 120W, supporting 14 frequency adjustment levels (1.2GHz to 3.3GHz) and a voltage adjustment range of 0.8V to 1.35V. Due to long-term operation under high temperature and high load conditions, the chip exhibits obvious aging characteristics: the threshold voltage drifted from the initial 0.42V to 0.49V, the leakage current increased from the initial 850mA to 1120mA, and the static power consumption increased from the initial 18W to 24.3W.

[0084] March 15, 20xx The power management controller begins collecting the server's load time series via the performance monitoring unit (PMU), with a sampling period set to 100ms. to Within a 27-second time window, a typical cascading load surge event triggered by a video transcoding task was detected.

[0085] Power management controller in to 270 load sampling points were collected within the time period, forming a load time series. The system was configured with three time window scales: ms ms The corresponding detection thresholds are 8%, 12%, and 18%, respectively. The multi-scale change point detection algorithm performs sliding window analysis at each scale to identify the moment when the change in the average load exceeds the threshold.

[0086] Table 1 Input data for step 1: Partial original load time series for multi-scale change point detection Table 2 Output data of Step 1: Set of phase transition points and cascaded mutation clusters output by multi-scale change point detection The system sets a preset clustering threshold. Seconds. Through temporal clustering, phase transition points. to The time intervals were 0.3 seconds, 3.2 seconds, 3.3 seconds, 3.6 seconds, and 4.5 seconds, respectively, all less than 8 seconds. Therefore, these six phase transition points were grouped into the same cascaded mutation cluster and numbered as follows: The time span of this cascaded mutation cluster is... to It lasted for 14.9 seconds.

[0087] Targeting cascaded mutation clusters The system extracts characteristic parameters of each phase transition point within the cluster and reads the current power-frequency response parameters from the power consumption genome. The power consumption genome currently stores the following parameters: static power consumption coefficient. W, dynamic power consumption factor Frequency response slope Voltage sensitivity index .

[0088] Table 3 Input data for step 2: Phase transition point feature extraction results of cascaded mutation clusters

[0089] Evolutionary pattern determination process: the first 4 phase transition points within the cluster ( to All of them are positive changes, and the last two phase transition points ( and The change is negative, exhibiting a sequence pattern of first rising and then falling. Since the number of directional changes is 1 (from positive to negative), which does not exceed the preset threshold for oscillation determination (the threshold for this time span is 3 times), it is determined to be an evolutionary pattern of first rising and then falling, and the cascaded evolutionary pattern is labeled "first rising and then falling".

[0090] Steady-state prediction calculation: Sum of positive amplitudes = %, sum of negative magnitudes = %, cluster initial load rate %.

[0091]

[0092] Table 4 Output data of Step 2: Steady-state prediction results and corresponding target operating points

[0093] The system is Upon detecting cascading mutation clusters, the penetrating modulation strategy generation process is immediately initiated. The current chip operating status is: frequency... GHz, voltage V, power consumption W.

[0094] Based on the steady-state prediction results from step 2, the target operating point is: GHz V. Voltage difference V, accounting for 10.8% of the rated voltage of 1.2V, exceeds the safety threshold (8%), therefore the regulation process is decomposed into two intermediate stages.

[0095] Table 5 Output data of step 3: Phased implementation plan of the penetrating control strategy The system sequentially issues control commands through the voltage and frequency adjustment interface (ACPI P-state interface), and the entire penetrating control process is completed in 187μs. The control execution trajectory record shows that the voltage and frequency smoothly transition to the target values ​​without any system abnormalities or performance fluctuations.

[0096] Cascade Mutation Clusters exist The system converges to a steady-state operating state at all times. to Steady-state power consumption data was collected within a 30-second observation window (twice the cluster time span of 14.9 seconds). The load fluctuation amplitude within the observation window was 2.1% (less than the fluctuation threshold of 5.4%). No new phase transition point was detected, confirming that the system had entered a steady-state operating state.

[0097] Table 6 Input data for step 4: Actual power consumption measurement data during steady-state operation The average measured power consumption during steady-state operation is:

[0098] The predicted power consumption value calculated in step 2 is W. Calculate the cascade response deviation index:

[0099] This deviation index indicates that the actual power consumption is 3.42% higher than the predicted power consumption, and the deviation value is recorded in the cascade response deviation index sequence.

[0100] The power management controller recorded cascade response deviation indices for 15 consecutive cascaded mutation clusters over a seven-day period from March 15th to March 22nd, 20xx. The system performed trend analysis on these 15 deviation indices to detect the presence of persistent monotonic drift patterns related to aging.

[0101] Table 7 Input data for step 5: Cascade response bias index sequence of 15 consecutive cascaded mutation clusters Linear regression analysis was performed on the cascade response deviation index sequence, with the observation order (1 to 15) as the independent variable. Cascade response bias Dependent variable The least squares method is used to fit a univariate linear equation:

[0102] The regression slope was calculated. % / time, intercept %, coefficient of determination .

[0103] Table 8 Output data of step 5: Cascade sensing aging drift detection results absolute value of slope % > Slope threshold 0.15%, and coefficient of determination The fit threshold is 0.80, which meets the criteria for a sustained monotonic drift pattern. Further Mann-Kendall trend test was performed, and the test statistic was... significance level < Preset significance threshold 0.05, confirming the existence of a statistically significant upward trend.

[0104] Generate cascaded sensing aging drift detection results: Significant aging drift was detected, with the drift direction being positive (power consumption higher than predicted), and the drift rate increasing by 0.189% with each observation.

[0105] After detecting significant aging drift, the system on March 23, 20xx... The online micro-calibration program was triggered during the low-load period in the early morning. The current chip load rate is 27.3% (lower than the preset low-load threshold of 35%), and this low-load state has lasted for 92 seconds (exceeding the preset duration threshold of 60 seconds).

[0106] The system selects 12 representative sampling points in the voltage-frequency operating space (the power consumption genome parameter dimension is 4, and the number of sampling points is 3 times the parameter dimension), and distributes them according to the boundary priority principle in the high voltage high frequency region (3 points), low voltage low frequency region (3 points), high voltage low frequency region (2 points), low voltage high frequency region (2 points), and central region (2 points).

[0107] At each sampling point, a corresponding voltage frequency value is set, a standard load mode (a composite load with a fixed IPC of 1.35) is applied, and the actual power consumption value after stabilization is measured. The entire online micro-calibration process takes 78 seconds to complete and generates a sparse calibration dataset.

[0108] Table 9 Input data for step 6: Sparse calibration dataset and prediction error analysis (partial sampling points) Based on a sparse calibration dataset and the current power consumption genome, an incremental learning algorithm calculates the prediction error distribution for each parameter dimension. The prediction errors for the 12 sampling points are grouped and statistically analyzed by parameter dimension: the mean error for static power consumption is +2.3W (error rate 6.8%), the mean error for dynamic power consumption is +0.4W (error rate 1.2%), the mean error for frequency response is +0.1W (error rate 0.3%), and the mean error for voltage sensitivity is +0.2W (error rate 0.5%).

[0109] Table 10 Output data of step 6: Power consumption genome parameter update results An error threshold of 5% was set. Static power consumption error rates of 6.8% (> 5%) were marked as drift parameters; dynamic power consumption error rates of 1.2% (< 5%) were not marked as drift parameters. Updated values ​​for static power consumption parameters were calculated using the least squares method based on a sparse calibration dataset. W.

[0110] Update magnitude calculation: < Upper limit threshold 25%, update range not exceeded, update value accepted. Combine the updated values ​​of the drift parameters with the original values ​​of the undrift parameters to generate the updated power consumption genome vector: W, , , .

[0111] Based on the updated power consumption genome vector, the system recalculates the feature thresholds used in the cascade evolution mode determination process. Since the static power consumption increases from 24.3W to 26.8W, and the static power consumption percentage increases from 35.4% to 38.9%, the direction determination threshold needs to be increased to eliminate interference from static power consumption fluctuations.

[0112] Table 11 Output data of step 7: Cascade response model parameter correction results The original direction determination threshold is the load change amplitude. Adjusted according to the change in static power consumption ratio The dynamic power consumption coefficient remained unchanged, and the amplitude threshold remained the same. The oscillation number threshold remains unchanged. The sequence remains unchanged.

[0113] The coefficients of the steady-state prediction formula are recalculated, and the updated steady-state prediction formula is as follows:

[0114] The updated feature threshold (direction determination threshold 6.2%) and coefficient parameters (static power consumption 26.8W) are written into the cascade response model to generate the corrected cascade response model, thus completing the cascade response model correction process.

[0115] March 23, 20xx The system detected a new cascade of mutations. The time span is to It contains 5 phase transition points and is determined to be a monotonically increasing evolutionary mode. Steady-state prediction is performed using a corrected cascaded response model. The initial cluster load rate is 29.7%, the sum of positive amplitudes is 48.3%, and the predicted steady-state load rate is... .

Claims

1. A method for optimizing and controlling chip power consumption, characterized in that, Includes the following steps: The load time series of the target chip is obtained, and the set of phase transition points is identified by a multi-scale change point detection algorithm. Temporal clustering is performed on adjacent phase transition points with a time interval less than a preset clustering threshold to generate a cascaded mutation cluster sequence. For each cascaded mutation cluster, the evolution pattern of the load change sequence within the cluster is analyzed based on the power consumption-frequency response parameters stored in the power consumption genome, cascaded evolution pattern tags are generated, and the target load state is calculated based on the cascaded evolution pattern tags to generate steady-state prediction results. A penetrating regulation strategy is generated and executed based on the steady-state prediction results. The penetrating regulation strategy directly uses the predicted steady state of the cascaded mutation cluster as the regulation target. After the cascaded mutation cluster converges to a steady-state operating state, the deviation between the actual power consumption measurement value and the predicted power consumption value is calculated to generate a cascaded response deviation index. Trend analysis is performed on the cascade response deviation index sequence corresponding to multiple consecutive cascaded mutation clusters to detect continuous monotonic drift patterns and generate cascaded sensing aging drift detection results. If significant aging drift exists, trigger the online micro-calibration program to collect a sparse calibration dataset and use an incremental learning algorithm to update the drifting parameters in the power consumption genome; Based on the updated power consumption genome vector, the feature threshold for determining the cascade evolution mode and the coefficient parameters of the steady-state prediction formula are recalculated to generate a corrected cascade response model. The corrected cascade response model is then applied to subsequent cascade mutation clusters to generate regulatory strategies.

2. The chip power consumption optimization and control method according to claim 1, characterized in that, The multi-scale change point detection algorithm includes: Set a set of time window scales. For each scale, calculate the change in the mean load within the window using a sliding window method on the load time series. When the change exceeds the detection threshold corresponding to that scale, record that moment as a candidate phase transition point. Candidate phase transition points detected at each scale are time-aligned and weighted by confidence to output a set of phase transition points; The time alignment refers to unifying the time positions of candidate phase transition points with a time difference smaller than the minimum window scale to the time median of the group of candidate phase transition points; the confidence weighted fusion refers to counting the number of times a candidate phase transition point is detected at each scale, using the ratio of the number of detections to the total number of scales as the confidence score, and retaining candidate phase transition points whose confidence scores exceed a preset confidence threshold.

3. The chip power consumption optimization and control method according to claim 1, characterized in that, The cascaded evolution mode labels include monotonically increasing, monotonically decreasing, first increasing then decreasing, and oscillating types, and the determination rule is as follows: If the change direction of all phase transition points within a cluster is positive and the amplitude increases or remains constant, it is determined to be a monotonically increasing type. If the change direction of all phase transition points within a cluster is negative and the amplitude increases or remains constant, it is determined to be a monotonically decreasing type. If the phase transition points within a cluster exhibit a sequence of changes from positive to negative, then it is determined to be of the type that first rises and then falls. If the direction of change of phase transition points within a cluster changes alternately more than a preset number of times, it is determined to be an oscillating type; The direction of change is determined based on the difference between the average load before and after the phase transition point. A positive difference indicates a positive change, and a negative difference indicates a negative change.

4. The chip power consumption optimization and control method according to claim 1, characterized in that, The steady-state prediction result is calculated as follows: For monotonically increasing types, the steady-state prediction is the current load plus the sum of all positive amplitudes within the cluster; For monotonically decreasing types, the steady-state prediction is the current load minus the sum of all negative magnitudes within the cluster; For the type of load that increases first and then decreases, the steady-state prediction value is the current load plus the sum of positive amplitudes and then minus the sum of negative amplitudes; For oscillating types, the steady-state prediction value is the average load of the terminal phase transition points within the cluster, and the number of terminal phase transition points is taken as the last 30% to 50% of the total number of phase transition points within the cluster.

5. The chip power consumption optimization and control method according to claim 1, characterized in that, The generation of the penetration-type control strategy includes: Based on the steady-state prediction results, query the power consumption-frequency mapping relationship in the power consumption genome, traverse the discrete frequency ranges supported by the chip, calculate the predicted power consumption value that meets the load requirements under each frequency range, and select the frequency range with the smallest predicted power consumption value as the target frequency value. The target voltage value is determined based on the target frequency value and the voltage-frequency constraint relationship; When the difference between the target voltage value and the current voltage value exceeds the safety threshold, the control process is decomposed into several intermediate stages. The voltage adjustment range of each intermediate stage does not exceed the safety threshold, and a stabilization waiting time is set between each intermediate stage.

6. The chip power consumption optimization and control method according to claim 1, characterized in that, The calculation of the cascade response deviation index includes: Divide the difference between the actual power consumption measurement and the predicted power consumption by the predicted power consumption value to obtain the dimensionless percentage deviation value. The steady-state operating condition is determined as follows: within a preset observation window, the load fluctuation amplitude is less than the fluctuation threshold, and no new phase transition point is detected; the length of the preset observation window is 2 to 5 times the time span of the cascaded mutation cluster, and the fluctuation threshold is 3% to 8% of the steady-state predicted load value.

7. The chip power consumption optimization and control method according to claim 1, characterized in that, The detection of the continuous monotonic drift mode includes: Calculate the linear regression slope for the cascade response deviation index sequence. If the absolute value of the slope exceeds the slope threshold and the regression fit exceeds the fit threshold, then a continuous monotonic drift pattern is determined to exist. The Mann-Kendall trend test is used to distinguish between aging drift and random fluctuations. When the significance level is less than the preset significance threshold, the detected trend is judged as aging drift. The regression fit is characterized by the coefficient of determination, with a fit threshold of 0.7 to 0.9 and a preset significance threshold of 0.05 to 0.

1.

8. The chip power consumption optimization and control method according to claim 1, characterized in that, The online micro-calibration program includes: Triggered during a period when the chip load rate is below a preset low load threshold and the duration exceeds a preset duration threshold; Representative sampling points are selected in the voltage-frequency operating space according to the boundary priority principle. The actual power consumption value after stabilization is measured at each sampling point to form a sparse calibration dataset. The boundary priority principle refers to dividing the voltage-frequency operating space into four boundary quadrants: high voltage high frequency region, low voltage low frequency region, high voltage low frequency region, and low voltage high frequency region. Sampling points are selected in each boundary quadrant, while a small number of sampling points are selected in the central region.

9. The chip power consumption optimization and control method according to claim 1, characterized in that, The incremental learning algorithm includes: Based on the sparse calibration dataset and the current power genome, the prediction error distribution across each parameter dimension is calculated. Identify parameter dimensions whose prediction errors exceed the error threshold and mark them as drift parameters; The drift parameters are updated using the least squares method based on a sparse calibration dataset. The updated values ​​of the drift parameters are combined with the original values ​​of the undrifted parameters to generate the updated power consumption genome vector; Among them, an upper limit threshold is set for the parameter update magnitude. When the calculated update magnitude exceeds the upper limit threshold, the update magnitude is limited to the threshold boundary value.

10. A chip power consumption optimization and control system, used to execute the chip power consumption optimization and control method according to any one of claims 1 to 9, characterized in that, include: The cascaded mutation cluster identification module is used to acquire the load time series of the target chip, identify the set of phase transition points using a multi-scale change point detection algorithm, perform time-series clustering processing on adjacent phase transition points, and generate a cascaded mutation cluster sequence. The evolution pattern analysis module is used to analyze the evolution pattern of load change sequences within a cluster based on power consumption-frequency response parameters in the power consumption genome, and generate cascaded evolution pattern labels and steady-state prediction results. The penetrating control module is used to generate and execute penetrating control strategies based on steady-state prediction results. The deviation calculation module is used to calculate the cascade response deviation index after the cascaded mutation cluster converges to a steady state. The aging drift detection module is used to perform trend analysis on the cascade response deviation index sequence and generate cascade sensing aging drift detection results. An online micro-calibration module is used to collect sparse calibration datasets and update drifting parameters in the power genome using an incremental learning algorithm; The model calibration module is used to recalculate feature thresholds and coefficient parameters based on the updated power genomic vector to generate a calibrated cascade response model.

Images