A device for estimating chip power consumption and a method thereof

TWI945551BActive Publication Date: 2026-11-11MEDIATEK INC
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
TW113126776
Authority / Receiving Office
TW · TW
Patent Type
Patents
Current Assignee / Owner
Priority Date
2024-07-10
Filing Date
2024-07-17
Publication Date
2026-11-11
Estimated Expiration
2044-07-16

Smart Images

  • Figure TWG2TB001920296_001
    Figure TWG2TB001920296_001
  • Figure TWG2TB001920296_002
    Figure TWG2TB001920296_002
  • Figure TWG2TB001920296_003
    Figure TWG2TB001920296_003
Patent Text Reader

Abstract

A computing device collects a plurality of data samples. Each data sample represents the signal activity of a plurality of signals on a chip. The computing device selects a subset of the plurality of signals as surrogate signals. These surrogate signals are correlated with the actual power consumption of the chip according to a criterion. The computing device uses the signal activity of the plurality of signals as input and the actual power consumption as output to train a power model. The computing device fine-tunes the coefficients of the surrogate signals in the power model. This fine-tuning adjusts the estimation error between the estimated power consumption output by the power model and the actual power consumption.
Need to check novelty before this filing date? Find Prior Art

Description

Machine Learning-Assisted In-Chip or Out-of-Chip Power Model Building Framework This disclosure generally relates to power modeling and estimation of integrated circuits, and more particularly, to techniques for using machine learning to accelerate data collection and training processes to build in-chip or out-of-chip power models. In chip design, various techniques and technologies have been used for power modeling and estimation. These methods aim to predict the power consumption of a chip based on its design and activities, enabling designers to optimize power efficiency and provide reliable operation within power limits. One common power modeling method is to use simulation-based methods. These methods involve simulating the chip design at various levels of abstraction, such as the register transfer level (RTL) or gate level, and estimating power consumption based on the switching activities of circuit components. However, simulation-based methods are time-consuming and computationally expensive, especially for large and complex chip designs. The process of generating a power report using a tool like PrimeTime PX (PTPX) can take several hours even for a small time interval, making it impractical for comprehensive power analysis. The following presents a simplified summary of one or more aspects to provide a basic understanding of these aspects. This summary is not an extensive overview of all contemplated aspects, and is not intended to identify key or critical elements of all aspects or to delineate the scope of any or all aspects. Its sole purpose is to present some concepts of one or more aspects in a simplified form as a prelude to the more detailed description provided later. In one aspect of the present disclosure, a method, a computer-readable medium, and a device are provided. The device may be a computing device. The computing device collects a plurality of data samples. Each data sample represents the signal activities of a plurality of signals of the chip. The computing device selects a subset of the plurality of signals as proxy signals. These proxy signals are related to the actual power consumption of the chip according to a criterion. The computing device uses the signal activities of the plurality of signals as inputs and the actual power consumption as an output to train a power model. The computing device fine-tunes the coefficients of the proxy signals in the power model. This fine-tuning adjusts the estimation error between the estimated power consumption output by the power model and the actual power consumption. To achieve the above and related purposes, one or more aspects include features that are fully described hereinafter and particularly pointed out in the claims. The following description and the drawings set forth certain illustrative features of one or more aspects in detail. However, these features are only a small part of the several different ways of using the principles of the various aspects, and this description is intended to include all such aspects and their equivalents. These and other objects of the present invention will no doubt become apparent to those skilled in the art after reading the detailed description of the preferred embodiments shown in the following figures and drawings. The following detailed description, along with the accompanying drawings, is intended to describe various configurations and is not meant to imply that these concepts can only be practiced in the configurations described. The detailed description includes specific details for the purpose of providing a thorough understanding of the various concepts. However, for those skilled in the art, these concepts may be practiced without these specific details. In some instances, well-known structures and components are shown in block diagram form to avoid obscuring the concepts. Aspects of a telecommunications system will now be described with reference to various devices and methods. These devices and methods will be described in the following detailed description and illustrated in the accompanying drawings by various blocks, components, circuits, processes, algorithms, etc. (collectively referred to as "elements"). These elements can be implemented using electronic hardware, computer software, or a combination of both. Whether an element is implemented as hardware or software depends on the particular application and the design constraints imposed on the overall system. For example, an element, any part of an element, or any combination of elements can be implemented as a "processing system" including one or more processors. Examples of processors include microprocessors, microcontrollers, graphics processing units (GPUs), central processing units (CPUs), application processors, digital signal processors (DSPs), reduced instruction set computing (RISC) processors, system on a chip (SoC), baseband processors, field programmable gate arrays (FPGAs), programmable logic devices (PLDs), state machines, gate logic, discrete hardware circuits, etc., as well as other suitable hardware configurations to perform the various functions described in this disclosure. One or more processors in the processing system may execute software. Software is to be broadly construed as instructions, instruction sets, code, code segments, program code, programs, subprograms, software components, applications, software applications, software packages, routines, subroutines, objects, executable files, threads, programs, functions, etc., whether referred to as software, firmware, middleware, microcode, hardware description language, or otherwise. Thus, in one or more example aspects, the described functions can be implemented in hardware, software, or a combination of both. If implemented in software, the functions can be stored on or encoded as one or more instructions or code on a computer-readable medium. Computer-readable media includes computer storage media. Storage media can be any available medium that can be accessed by a computer. By way of example and not limitation, such computer-readable media can include random access memory (RAM), read only memory (ROM), electrically erasable programmable ROM (EEPROM), optical disk storage, magnetic disk storage, other magnetic storage devices, combinations of the above types of computer-readable media, or any other medium that can be used to store computer-executable code in the form of instructions or data structures and that can be accessed by a computer. Figure 1 is Figure 100, which illustrates a schematic diagram of on-chip or off-chip power modeling. Chip 110 includes a plurality of internal circuit components that consume power during operation. The power consumption of chip 110 varies with the workload executed by chip 110. The on-chip power of chip 110 is measured to generate signals 120-1, 120-2, ... 120-n, which reflect the power consumption of different components within chip 110. Signals 120-1, 120-2, ... 120-n are input into power model 150 to establish a power model. Power model 150 can estimate the power consumption of chip 110 based on signals 120-1, 120-2, ... 120-n. Signals 120-1, 120-2, ... 120-n can represent the activities of a subset of signals within chip 110. Signals 120-1, 120-2, ... 120-n are selected to reduce the hardware overhead of power model 150. Signal activities can include , and signals, where represents the duration that the signal is at a low voltage level, represents the duration that the signal is at a high voltage level, and represents the number of signal transitions. More specifically, the number of transitions (TC) refers to the number of times the signal switches between a low voltage level and a high voltage level, or vice versa. In other words, it represents the number of transitions of the signal between its two states. Power model 150 can be implemented in software and executed on a computer, or can be implemented in hardware and integrated into chip 110. Power model 150 can be used to monitor the power consumption of chip 110 during operation, or can be used to estimate the power consumption of chip 110 at design time. Power model 150 can be trained using machine learning algorithms. A plurality of data samples are collected to train power model 150. Each data sample includes the values of signals 120-1, 120-2, ... 120-n, and the corresponding power consumption of chip 110. The training configuration of power model 150 can include two stages. In the first stage, a subset of signals is selected as a proxy. The proxy is the signal that is most relevant to the power consumption of chip 110. In the second stage, the coefficients of the proxy are fine-tuned to accurately predict the dynamic power consumption of chip 110. The dynamic power consumption of chip 110 can be expressed as a function of signal activities and proxy coefficients, and can be represented by the following equation: where represents the coefficients of the proxy, which are learned by an AI algorithm. The power model 150 can be used to accelerate the power estimation process, reduce the hardware cost of power meters on the chip, and enable energy efficiency optimization technologies such as dynamic voltage and frequency scaling (DVFS). FIG. 2 is FIG. 200, which illustrates the framework for generating a chip power report. In the first framework, the register transfer level (RTL) description 210 of the chip design and the design constraints 212 (such as clock rate) are fed into the front-end tool 222, which includes a synthesizer. The front-end tool 222 also takes the cell library 214 as input. The RTL description 210 can be written in a hardware description language such as VHDL. The front-end tool 222 performs RTL simulation and synthesis to generate a pre-layout netlist 230. The pre-layout netlist 230 is an idealized representation of the design and does not include the clock tree. The pre-layout netlist 230, together with the design rules 232 and the cell library 234, is fed into the back-end tool 242 for gate-level simulation. The back-end tool 242 generates a post-layout netlist 244, which is a more realistic representation of the actual chip activities and includes the clock tree, delays, and glitches. The post-layout netlist 244 is then input into the power analysis tool 260, which performs per-cycle power tracking to generate a detailed power report 270. The power report 270 contains the per-cycle power consumption breakdown for a given chip design. For example, the power report 270 may include a power trace that details the exact power usage for each clock cycle within the analysis time range. The power report 270 may contain the total power consumption. The report aggregates the per-cycle data to provide the total power consumed during the analysis period (e.g., for a specific workload or mode). The power report 270 may contain a breakdown by power domain / component. The report shows the power consumption of different functional blocks or power domains within the chip. The power report 270 may include statistical metrics of the power data, such as average power, peak power, and power distribution. The process of generating the power report 270 using the power analysis tool 260 can be very time-consuming. For example, it may take 5.5 hours to generate a power trace for just a 1 second time interval. This becomes a bottleneck in the power estimation process, especially when multiple modes need to be tested. To address this issue, in the second framework, an AI-based power model 150 can be used to replace the time-consuming power analysis step. As shown in FIG. 2, the power model 150 takes the pre-layout netlist 230 as input, selects a subset of signals most relevant to power consumption as proxies, and tunes the coefficients of these signals to estimate power consumption. This can significantly reduce the time required to generate power estimates compared to the traditional process using the power analysis tool 260. The power model 150 can be implemented on the chip to enable real-time power monitoring and optimization techniques such as dynamic voltage and frequency scaling (DVFS). By providing runtime power information, the on-chip power model allows the system to dynamically adjust the voltage and frequency according to the current power consumption, thereby improving energy efficiency. Power models are used to estimate the power consumption of a chip. There are two main types of power models: off-chip power models and on-chip power models. Off-chip power models are used to estimate the power consumption of a chip during design. On-chip power models are used to monitor the power consumption of a chip during runtime. The process of training a power model for chip design faces several challenges that can affect the efficiency and accuracy of the model. Two key challenges are the time-consuming data collection process and the bottleneck in proxy selection time. Collecting training data for the power model is a time-consuming process. As mentioned earlier, generating accurate power reports using the power analysis tool 260 can take a significant amount of time. For example, using the PrimeTime PX (PTPX) tool for RTL sign-off may take 5.5 hours to generate a power trace for just 1 second time interval. This becomes a bottleneck in the data collection process, especially when collecting training data that covers the entire power interval range. To train the power model 150, data samples including signal activities and their corresponding power consumption values need to be collected. The power consumption values are obtained from the power report 270 generated by the power analysis tool 260. However, the time required to generate these power reports for multiple modes and time intervals can be excessive. For example, consider a scenario of training a power model for a GPU. The GPU may have different power consumption levels, such as low, medium, and high, depending on the workload. Training data that covers the entire power interval range needs to be collected. This requires running various modes on the GPU and generating power reports for each mode, which can be very time-consuming. Another challenge in training the power model 150 is the bottleneck in proxy selection time. As mentioned earlier, the first step in the training configuration is to select a subset of signals that are most relevant to the chip's power consumption. These selected signals are called proxies. However, the process of selecting proxies becomes a bottleneck in training time, especially for large chip designs. In fact, a mobile chip may have more than 10 million signals. Selecting the most relevant proxies from such a large number of signals is computationally expensive and time-consuming. Figure 3 is a schematic diagram showing a proposed power model framework 300. The framework 300 includes two main parts: a data-related component and a power model construction component. The framework aims to reduce the time required for data collection and model training. The data-related components include a data collection component 314 and a data preprocessing component 316. The original data signal 312 from the hardware design is fed into the data collection component 314. The data collection component 314 adopts the minimum dataset method to reduce the number of PTPX (PrimeTime PX) power reports that need to be collected, thereby reducing the data collection time. The collected data is then passed to the data preprocessing component 316, which applies a similarity filtering method to remove highly similar data points. This step not only accelerates the training time by reducing the amount of training data, but also enhances the model's ability to learn from diverse data points rather than repeated and similar data. The power model construction components include an agent selection component 332, a coefficient tuning component 334, and a quantization component 336. The preprocessed data is fed into the agent selection component 332, which uses the PICASSO library to select a portion of the signals most relevant to power consumption from a large number of signals (e.g., 10 million signals) in the hardware design. The PICASSO library is described in Zhao, T., Liu, H., & Zhang, T. (2017); Pathwise coordinate optimization for sparse learning: Algorithm and theory; arXiv:1412.7477, 9 Feb 2017, the entire content of which is hereby incorporated by reference. The PICASSO library adopts various techniques to accelerate the agent selection process, which will be discussed in detail later. In addition, the agent selection component 332 applies the maximized warm start initialization method to further optimize the selection process. The selected agents are then passed to 334, which tunes the coefficients of each selected signal to accurately predict the dynamic power consumption. The coefficient tuning component 334 also adopts the regularization path method to accelerate the tuning process. The quantization component 336 quantizes the coefficients, signal activities, and power values of the model into fewer bits to reduce the hardware cost of the on-chip power model 350. Quantization is performed not only on the model coefficients, but also on the input signal activities and output power values. Figure 4(A) is a flowchart illustrating a procedure 400 for creating an orthogonal minimum dataset 410 for training the power model 150. Figure 4(B) is a diagram showing the orthogonal minimum dataset 410. The data collection element 314 obtains an array 402 of a plurality of PMU (Performance Monitoring Unit) signals from the design signal 312. of each PMU signal The array represents the duration for which the PMU signal is at a high voltage level. The PMU is a hardware unit integrating performance counters. A performance counter is a register that counts the number of occurrences of a specific event. The events counted by the performance counter can vary depending on the chip. For example, some performance counters may count the number of clock cycles, while others may count the number of instructions executed. Performance counters can be used to measure the activity of the chip. The activity of the chip is related to the power consumption of the chip. Therefore, the PMU signal can be used to estimate the power consumption of the chip. The data collection component 314 pairs of a plurality of PMU signals The array 402 applies a dimensionality reduction method to obtain a plurality of PMU embeddings 404. The data collection component 314 applies a clustering method to the plurality of PMU embeddings 404 to obtain orthogonal clusters 406. The data collection component 314 applies a sampling rule, such as a sampling rule based on group size, to sample the orthogonal clusters 406 to obtain an orthogonal minimum dataset 410. The dimensionality reduction and clustering operations use techniques such as UMAP (Uniform Manifold Approximation and Projection) for dimensionality reduction and DBSCAN (Density-Based Spatial Clustering of Applications with Noise) for clustering. These techniques are applied sequentially, first performing dimensionality reduction to transform the high-dimensional PMU signal array into a low-dimensional space (e.g., 2D), and then performing clustering to group similar data points according to their distribution in the low-dimensional space. As shown in Figure 4(B), the original dataset 422 contains orthogonal clusters 406-1, 406-2, 406-3, etc. The data selected from the original dataset 422 forms the orthogonal minimum dataset 410 and the validation dataset 426. The orthogonal minimum dataset 410 contains clusters 434-1, 434-2, 434-3, etc. The validation dataset 426 contains clusters 436-1, 436-2, 436-3, etc. The data collection component 314 can determine the group sizes of the orthogonal minimum dataset 410 and the validation dataset 426. For example, if cluster 406-1 constitutes 30% of the original dataset 422 and cluster 406-1 needs to be reduced by 50% for the orthogonal minimum dataset 410, then the data collection component 314 can randomly select 30% * 50% of the data points in the original dataset 422. These selected data points will be shown as cluster 434-1. The orthogonal minimum dataset 410 can be used to train the power model 150. The orthogonal minimum dataset 410 is smaller than the original dataset 422. This can reduce the time required to train the power model 150. This process uses the Array 402 clusters the data and then samples from each cluster by group size to create an orthogonal minimum dataset 410. The orthogonal minimum dataset 410 is diverse and covers the full range of power intervals. In some cases, this process may reduce the number of data points that need to be collected by up to 93%. Figure 5 is a chart numbered 500, illustrating another process for creating 410 to train the power model 150. This process includes similarity filtering 508 of the orthogonal clusters 406 to remove highly similar data points within each cluster, further providing diversity in the minimum dataset. The data collection component 314 obtains the array 402 of PMU signals. The data collection component 314 applies a dimensionality reduction method to the array 402 of a plurality of PMU signals to obtain a plurality of PMU embeddings 404. Then, the data collection component 314 applies a clustering method to the plurality of PMU embeddings 404 to obtain orthogonal clusters 406. In addition, the data collection component 314 applies a sampling rule to the orthogonal clusters 406 to obtain sampled data points. The data collection component 314 then performs similarity filtering 508 on the orthogonal clusters 406. The data collection component 314 calculates the similarity between each pair of sampled data points within each cluster 406. For each pair of sampled data points, the data collection component 314 concatenates the arrays of the PMU signals of each sampled data point into a 1D array. Then, the data collection component 314 calculates the similarity between the two 1D arrays. If the similarity between the two 1D arrays is higher than a threshold, the data collection component 314 removes one of the two sampled data points. The combination of the minimum dataset method and similarity filtering results in a smaller and more diverse dataset for training the power model 150. This reduces the training time and improves the generalization ability of the model by preventing the model from being biased towards highly similar data points. The data within each cluster is more orthogonal, meaning that the data points within each cluster are more independent of each other. This can further improve the performance of the model by reducing the correlation between input features. Figure 6 is a flowchart numbered 600, illustrating the process of implementing similarity filtering to preprocess the training data of the power model 150. This process aims to filter out highly similar data points to enhance the generalization ability of the model and accelerate the training process. In operation 602, the data preprocessing component 316 obtains an array of all signals from the collected data. Each signal array represents the activity of a signal in the hardware design. In operation 604, the data preprocessing component 316 calculates the similarity between each pair of data points. For each pair of data points, the corresponding signal scalars are concatenated into a 1D array. Then the similarity between the two 1D arrays is calculated. The similarity measurement can be based on various metrics, such as cosine similarity or Euclidean distance. In operation 606, the data preprocessing component 316 checks whether the similarity between the two 1D arrays exceeds a predefined threshold. If the similarity is greater than the threshold, one of the two data points is randomly removed. Thus, highly similar data points are filtered out, reducing redundancy in the training data. In operation 608, the data preprocessing component 316 obtains the filtered data without highly similar data points. The filtered data is more diverse and contains less redundancy, which can improve the efficiency and effectiveness of the training process of the power model 150. Figures 7(A) and 7(B) are diagrams that illustrate the impact of similarity filtering on the training and test data sets respectively. These diagrams show the power consumption patterns before and after applying similarity filtering. Figure 7(A) shows the power consumption pattern in the data set before similarity filtering. The horizontal axis represents the data points and the vertical axis represents the power consumption in watts. Sections 722-1, 722-2,..., 722-M represent different patterns in the training data set. Each pattern corresponds to a specific workload or activity executed by the chip. Before similarity filtering, the power consumption pattern contains many flat regions, indicating highly similar data points. These flat regions may bias the learning process of the power model, resulting in overfitting and poor generalization ability. Figure 7(B) shows the power consumption pattern in the data set after similarity filtering. Similar to Figure 7(A), the horizontal axis represents the data points and the vertical axis represents the power consumption in watts. Sections 732-1, 732-2,..., 732-K represent different patterns in the test data set. After similarity filtering, the flat regions are reduced, resulting in more diverse and independent data points in the test data set. This improves the performance of the power model on unseen data, thereby enhancing accuracy and robustness. The similarity filtering process is quantified in the following table: Impact of Similarity Filtering on Training Data The table shows that with a similarity threshold of 0.99, the number of original SAIF (Switching Activity Interchange Format) files is reduced from 4597 to 617, significantly reducing the training data size. The similarity filtering process takes only 4 minutes, demonstrating its efficiency. By implementing similarity filtering, in certain scenarios, the training time of the power model is accelerated by 26.3 times. This significant reduction in training time can be achieved without compromising the model's accuracy. In fact, removing highly similar data points enhances the generalization ability of the power model, resulting in better performance on both the training and test datasets. The similarity filtering process described in FIG. 6 and shown in FIGS. 7(A) and 7(B) addresses the challenge of redundant and highly similar data points in the training dataset. By filtering out these data points, the power model training process becomes more efficient and effective. The filtered data is more diverse and contains less redundancy, which prevents the power model from being biased towards repetitive data points. This enhances the model's generalization ability for unseen data, improving its accuracy and robustness. The similarity filtering process can be implemented in the data preprocessing component 316 of the proposed power model framework 302. The similarity filtering process works in conjunction with other components such as the data collection component 314 and the power model construction component to reduce the time required for data collection and model training. FIG. 8 shows multiple agent layouts 810, 820, 830, 840, demonstrating the bottom-up strategy adopted by the agent selection component 332 for agent selection. The agent selection algorithm faces challenges in memory usage and speed when selecting agents from large-scale mobile chip designs, which may contain over 10 million signals. The bottom-up strategy enables the agent selection component 332 to process the entirety of these large-scale designs within a single-chip on-chip power model (OPM), while reducing memory usage and execution time. The bottom-up approach sequentially selects agents and propagates the selections to the next level until the entire design is processed. Agent layout 810 shows the initial stage of the bottom-up strategy. The design hierarchy is represented by nested rectangles, with the outermost rectangle corresponding to the top-level module. Each layer in the hierarchy contains a set of signals, with deeper layers representing smaller modules with fewer signals. The agent selection component 332 starts agent selection from the deepest layer of the hierarchy and works its way up to the top level. In agent layout 810, the agent selection component 332 first selects agents from Group 1 and Group 2, which are the deepest modules. Each agent group contains a subset of signals that are most relevant to the power consumption at their corresponding hierarchical level. The selected agents from these groups are then combined with all the signals at the next level, i.e., Group 3, for another round of agent selection, as shown in agent layout 820. After this round of selection, the agent selection component 332 obtains the selected agents from Group 2, Group 1, and Group 3. These selected agents are then combined with the signals at the surrounding level to form Group 4, as shown in agent layout 830. This process continues iteratively until the entire design is processed and the final set of agents is selected, as shown in agent layout 840. The proxy layout 840 represents the final stage of the bottom-up strategy. In this layout, all previously selected proxies are combined into a single group, the proxy group 5. The proxy selection component 332 selects the final set of proxies from this group. This final set of proxies represents the signals that are most relevant for estimating the power consumption of the entire design. Therefore, the proxy selection component 332 does not need to process all the signals in the design simultaneously. Instead, it only needs to consider the selected proxies from the lower-level groups and the signals at the current level being processed. This approach significantly reduces memory usage and speeds up the proxy selection process because the proxy selection component 332 can work with a smaller subset of signals at each iteration. Figure 9 is a diagram 900 illustrating the process of proxy selection using the PICASSO library. The PICASSO library utilizes a path coordinate optimization algorithm to reduce the proxy selection time. The process includes an outer loop 950, a middle loop 956, and an inner loop 960. In operation 912, the proxy selection component 332 initializes a regularization parameter . The regularization parameter is used to control the sparsity of the solution. A larger value of encourages a sparser solution, meaning fewer proxies are selected. The middle loop 956 includes operations 914, 916, and 918. In operation 914, the proxy selection component 332 applies a strong rule to preselect coordinates. The strong rule includes calculating the gradient of each coordinate and comparing it with a threshold. If the absolute value of the gradient is greater than the threshold, the coordinate is included in the active set; otherwise, it is excluded from the active set. The active set includes the coordinates that are most likely to be relevant to the predicted power consumption. The inactive set includes the coordinates that are less likely to be relevant to the predicted power consumption. The inner loop 960 includes operation 916. In operation 916, the proxy selection component 332 performs active coordinate minimization on the active set. Active coordinate minimization is a technique that iteratively updates the values of the coordinates in the active set until the objective function is minimized. In operation 918, the proxy selection component 332 updates the active set. The active set is updated by adding the coordinates from the inactive set that have large absolute gradient values. Thus, the active set contains the coordinates that are most likely to be relevant to the predicted power consumption. The middle loop 956 (operations 914, 916, and 918) repeats until a convergence criterion is met. The convergence criterion may be based on a change in the objective function value or a change in the active set. Once the convergence criterion is met, the output solution 920 is obtained. In operation 922, the surrogate selection component 332 performs warm-start initialization, where the surrogate selection component 332 uses the solution of the previous iteration as the starting point for the current iteration. This helps to accelerate the convergence of the algorithm because the algorithm does not need to start from scratch in each iteration. The surrogate selection component 332 initializes a new regularization parameter . For example, the surrogate selection component 332 may decrease the regularization parameter to encourage the selection of more surrogates. The outer loop 950 (operations 920 and 922) is repeated until all regularization parameters are processed. Once all regularization parameters are processed, the surrogate selection component 332 selects a set of surrogate signals for the power model 150. The PICASSO algorithm starts with a large regularization parameter to suppress the over-selection of irrelevant coordinates. Gradually, the PICASSO algorithm restores the relevant coordinates to obtain a sparse output solution with optimal statistical properties, including parameter estimation and support recovery. The PICASSO library includes the following features: active coordinate minimization, strong rules, active set update, and warm-start initialization. Active coordinate minimization involves minimizing the coordinates only for the coordinates in the active set. This eliminates the need to compute the gradient vectors for all coordinates. Strong rules are used to preselect the coordinates of the active set. This involves computing the gradient of each coordinate and comparing it with a threshold. If the absolute value of the gradient of a coordinate is greater than the threshold, then the coordinate is included in the active set; otherwise, it is included in the inactive set. Active set update involves adding the coordinate with the largest absolute gradient value to the active set in each iteration. Warm-start initialization involves using the solution of the previous stage as the initialization for the current stage. This accelerates the convergence of the algorithm. Figure 10(A) is a schematic diagram illustrating the first type of process for surrogate selection and fine-tuning. The process involves multiple rounds of training, with each round consisting of two stages: stage-1 (surrogate selection) and stage-2 (fine-tuning). In the first round, all signals 1012 are fed into the outer loop 950 (the initial one) in stage-1 1021 from scratch. This outer loop 950 is the bottleneck of the training time because it has no previous results to use. After the initial outer loop 950, subsequent outer loops are performed. Then the selected surrogates and their coefficients are passed to stage-2 1031 for fine-tuning operation 1032. The output of the first round is a set of surrogates and their corresponding coefficients. In the second round, the process starts from the beginning, again starting from the initial outer loop 950 of stage - 1 1023, followed by subsequent outer loops and fine - tuning in stage - 2 1033. The output of the second round is another set of agents and coefficients. Similarly, the third round starts from the initial outer loop 950 in stage - 1 1025, followed by subsequent outer loops and fine - tuning in stage - 2 1035. The output of the third round is yet another set of agents and coefficients. In this process, each round starts from an initial outer loop, which is time - consuming and causes multiple bottlenecks throughout the training process. In each round, stage - 1 1021, 1023, 1025 uses the PICASSO library to select agents from the input signal. When there are no results from the previous round, the initial outer loop 950 is a bottleneck in the training time. Subsequent outer loops in the same round are faster because they are based on the results of the previous round. Then stage 2 (1031, 1033, 1035) fine - tunes the coefficients of the selected agents. Figure 10(B) illustrates a second type of process for agent selection and fine - tuning. This process utilizes the maximize warm - start method, aiming to reduce the number of bottlenecks by minimizing the occurrence of the initial outer loop. In the first round, all signals 1052 are fed into the initial outer loop 950 in stage - 1 1061. The result of this initial outer loop is stored in file 1082. In stage - 2 1071, a fine - tuning operation 1072 is applied to the agents and coefficients, and the output is a set of agents and coefficients. For the second round, the stored result of the first - round initial outer loop (in file 1082) is used to initialize the initial outer loop 950 in stage - 1 1063. This reduces the time required for the outer loop because it uses the previous result. The result is stored in file 1084. In stage - 2 1073, a fine - tuning operation 1074 is applied to the agents and coefficients. Similarly, the third round uses the stored result of the second round (stored in file 1084) to initialize the outer loop 950 in stage - 1 1065. The result is stored in file 1086, and the agents and coefficients are fine - tuned in stage - 2 1075. By storing and reusing the results of the previous - round initial outer loop, the second type of process reduces the number of bottlenecks to only one, significantly accelerating the training process. This method achieves a 5.01 - fold speedup, accompanied by only a slight decrease in accuracy. In the stage-1 of the disclosed invention 1061, 1062, 1063, the outer loop 950 utilizes a path coordinate optimization algorithm, particularly the PICASSO library, to select the surrogate for power model training. More specifically, the initial outer loop 950 in stage-1 1061 is executed only once at the beginning of the training process. It starts with a large regularization parameter and an empty active set. Specifically, a strong rule is applied to preselect the coordinates of the active set. Active coordinate minimization is performed in operation 916, iterating over the coordinates in the active set. Active set update is performed in operation 918, by adding the coordinates in the inactive set with large absolute gradient values. The middle loop 956 (strong rule, active coordinate minimization, and active set update) iterates until the convergence criterion is met. This criterion can be based on the change in the objective function value or the change in the active set. Once the middle loop 956 converges, the output solution of operation 920, containing information for fine-tuning, such as the selected surrogate and its coefficients, etc., is stored in file 1082. The subsequent outer loops 950 in stage-1 1063 and 1065 utilize the warm start initialization technique to accelerate the surrogate selection process. These outer loops do not start from scratch, but instead utilize the output solution of the previous outer loop as initialization. Specifically, the output solution from the previous outer loop (stored in files 1082, 1084) is loaded and used as the initial solution for the current outer loop. The regularization parameter is also updated, usually decreased, to encourage the selection of more surrogates. The strong rule is applied again to preselect the coordinates, but this time, it starts from the pre-filled active set of the previous iteration. Similar to the outer loop in stage-1 1061, active coordinate minimization in operation 916 and active set update in operation 918 are performed. The middle loop 956 continues according to the defined criterion until convergence is reached. The updated solution 920, containing the selected surrogates and their coefficients, is stored in files (1084, 1086). Figure 11(A) is a schematic diagram illustrating a first type of process for surrogate selection and coefficient fine-tuning using the regularization path method. The first type of process involves running multiple rounds to obtain the power model 150. Each round includes a stage-1 for surrogate selection and a stage-2 for coefficient fine-tuning. In the first round, all signals 1112 are input into stage 1 1121 for surrogate signal selection, using the regularization parameter . Then the selected surrogate signals and their coefficients are passed to stage 2 1131 for fine-tuning. The output of the first round is a set of surrogate signals and their corresponding coefficients. In the second round, the process restarts and enters stage 1 1123 for surrogate signal selection, but this time with a slightly reduced regularization parameter . Then the selected surrogate signals and coefficients are passed to stage 2 1133 for fine-tuning. The output of the second round is another set of surrogate signals and coefficients. Similarly, the third round starts from stage 1 1125 for surrogate signal selection, using a further reduced regularization parameter , and then proceeds to fine-tuning in stage 2 1135. The output of the third round is yet another set of surrogate signals and coefficients. During this process, each round starts with the surrogate selection stage, which is time-consuming and causes multiple bottlenecks throughout the training process. Figure 11(B) illustrates a second process for surrogate selection and coefficient fine-tuning using the regularization path method. This process aims to minimize the number of iterations of the surrogate selection in stage 1 by leveraging the regularization path and iteratively fine-tuning the coefficients in stage 2. In the first round, all signals 1152 are input into stage 1 1161 for surrogate selection, using the regularization parameter . Then the selected surrogate signals and their coefficients are passed to stage 2 1171 for iterative fine-tuning. In stage 2 1171, the coefficients are fine-tuned in multiple sub-rounds (e.g., the 1-1 round, the 1-2 round, ...) using a series of decreasing coefficient scaling parameters values (e.g., 1, 0.99, ..., 0.001). The output of each sub-round is a set of surrogate signals and their corresponding fine-tuned coefficients for each value. For the second round, the process starts from stage 1 1163 for surrogate selection, but with a significantly reduced regularization parameter . This is because the coefficients have been fine-tuned in stage 2 1171 of the first round. Then the selected surrogate signals and coefficients are passed to stage 2 1173 for another round of iterative fine-tuning, using the same decreasing value sequence. In some cases, by leveraging the regularization path and iteratively fine-tuning the coefficients in stage 2, the second process reduces the number of iterations of the surrogate selection in stage 1, potentially resulting in a 2.32-fold increase in speed with a small decrease in accuracy compared to the first process. The regularization path method addresses the bottleneck problem in the surrogate selection and coefficient fine-tuning process. Compared to the coefficient fine-tuning in stage 2, the surrogate selection in stage 1 is more time-consuming. Therefore, reducing the number of iterations in stage 1 can significantly speed up the entire training process. The first process, as shown in Figure 11(A), performs surrogate selection in each round, and the regularization parameter gradually decreases. However, this method still requires multiple Stage 1 iterations, resulting in bottlenecks during the entire training process. The second process, as shown in Figure 11(B), uses a regularization path and iterative fine-tuning in Stage 2 to minimize the number of Stage 1 iterations. By performing extensive coefficient fine-tuning in Stage 2, using a series of decreasing values, this process allows for a more aggressive decrease in the regularization parameter in subsequent Stage 1 surrogate selection rounds . This method effectively reduces the number of Stage 1 iterations, resulting in a significant acceleration of the entire training process. The regularization parameter controls the sparsity of the solution during the surrogate selection process (Stage 1). A higher enforces a sparser solution, resulting in fewer selected surrogate signals. determines the trade-off between model complexity and accuracy. A larger prioritizes a simple model with fewer surrogate signals (lower hardware overhead), but may come at the cost of slightly lower accuracy. A smaller allows for a more complex model with more surrogate signals, potentially achieving higher accuracy but increasing the hardware overhead. In Stage 2 (Figure 11(B)), is a scaling factor for the penalty term that limits the number of surrogate signals. A higher value introduces a higher penalty for using more surrogate signals. A lower value introduces a lower penalty for using more surrogate signals. In Figure 11(B), iterative fine-tuning is performed in Stage 2 using decreasing values, allowing for a more aggressive decrease in in subsequent rounds. The extensive fine-tuning thoroughly explores the solution space, enabling the algorithm to confidently use a significantly smaller in the next round, accelerating the entire training process. As shown, in Figures 11(A) and 11(B), the regularization parameter decreases in each subsequent surrogate selection round (Stage 1). A higher value strongly penalizes including many features (in this case, signals) in the model. This results in a very sparse solution, selecting only a few surrogate signals that have the strongest correlation with power consumption. A lower value relaxes this penalty, allowing more features to enter the model, resulting in a denser solution. The algorithm starts with a larger Start, with forced sparsity and to avoid overfitting, especially when dealing with millions of signals. As decreases, the algorithm allows more signals (proxy signals) to be included in the model, gradually increasing the complexity and potential accuracy of the model. "Warm start" initialization (operation 922 in Figure 9) uses the solution of the previous round as the starting point for optimization, making the optimization more efficient. The input to the proxy selection phase in each round is the signal set 1112 in Figure 11(A) and the signal set 1152 in Figure 11(B). As explained in operation 922, the solution of the previous round (the selected proxy signals and their coefficients) is used as the starting point for the optimization of the current round. This enables the coordinate optimization to converge faster. In Figure 11(B), the dense fine-tuning in stage 2 allows for a more aggressive decrease in in the subsequent stage 1. The algorithm benefits from the knowledge obtained in the previous rounds, enabling it to more effectively explore a wider range of proxy selections. Figure 12 is a flowchart 1200 that describes a process (method) for quantifying the power model 150 to reduce the hardware cost of implementing the power model 150. Quantification is necessary to reduce the hardware cost of the power model 150. Quantifying all coefficients to integers may significantly reduce the hardware cost with minimal loss of accuracy. To further save hardware costs, the signal activity that is the input to the power model 150 can also be quantified. The dynamic power output by the power model 150 is also quantified. All cases can be handled using full search or sequential search. As described above, the power model can be represented by the following equation: is the output of the power model, representing the estimated dynamic power consumption of the chip. is a function that maps the input and the model coefficients to the output of the function. is the input to the power model, representing the activity of a subset of signals within the chip. The signal activity can include and and signals. represents the coefficients of the power model, which are learned using artificial intelligence algorithms during the training process. These coefficients determine the weight or importance of each signal activity in the estimated dynamic power consumption. More specifically, in operation 1202, the quantization component 336 iteratively quantifies the power model coefficients . The coefficients are quantified to a fixed number of bits, such as 8 bits. This reduces the hardware cost of storing and processing the coefficients. This quantization process may find the minimum number of bits required to represent the coefficients with minimal loss of accuracy. In operation 1204, the quantization component 336 iteratively quantizes the power model input, using the coefficients quantized in operation 1202 . Signal activity may include , and signals. The signal activity is quantized into a fixed number of bits to reduce the hardware costs of the input interface and processing. The quantization process of the input may determine the minimum number of bits required to represent the signal activity with a minimum loss of precision. In operation 1206, the quantization component 336 iteratively quantizes the power model output, using the coefficients quantized in operations 1202 and 1204 and the quantized signal activity . The dynamic power output is quantized into a fixed number of bits to reduce the hardware costs of the output interface. The quantization order of operations 1202, 1204, and 1206 may be arbitrary. The quantization component 336 may identify the predetermined number of bits (e.g., the minimum number) for each quantization operation to result in an acceptable loss of precision. Thus, the quantized power model 150 minimizes the hardware costs while maintaining sufficient precision. The quantized power model 150 obtained after the quantization process has coefficients, input signal activity, and output dynamic power represented in fewer bits compared to the original power model 150. This results in a significant reduction in hardware costs when implementing the power model 150, making it more suitable for integration into mobile chips and other resource-constrained devices. The quantization process may use a full search or a sequential search to determine the optimal bit width for each component of the power model. FIG. 13 is a diagram 1300 illustrating a hardware embodiment of a device 1378 using a processing system 1314. The device 1378 may implement a collection component 1342, a training component 1346, and a fine-tuning component 1348. The processing system 1314 may be implemented using a bus architecture, generally represented by a bus 1324. The bus 1324 may include any number of interconnected buses and bridges, depending on the specific application and overall design constraints of the processing system 1314. The bus 1324 connects various circuits together, including one or more processors and / or hardware components, represented by a processor 1304, a network controller 1310, and a read / write media / memory 1306. The bus 1324 may also connect various other circuits, such as a timing source, peripherals, voltage regulators, and power management circuits, which are well known in the art and thus will not be described further. The processing system 1314 may be connected to a network controller 1310. The network controller 1310 provides means for communicating with various other devices on the network. The network controller 1310 receives signals from the network, extracts information from the received signals, and provides the extracted information to the processing system 1314, particularly the communication component 1320 of the device 1378. In addition, the network controller 1310 receives information from the processing system 1314, particularly the communication component 1320, and generates signals to be sent to the network based on the received information. The processing system 1314 includes a processor 1304 connected to a read / write medium / memory 1306. The processor 1304 is responsible for general processing, including executing software stored on the read / write medium / memory 1306. When the processor 1304 executes the software, it causes the processing system 1314 to perform the various functions described for any of the specific devices above. The read / write medium / memory 1306 is also used to store data that the processor 1304 operates on when executing the software. The processing system also includes a collection component 1342, a training component 1346, and a fine-tuning component 1348. These components may be software components running in the processor 1304, stored / stored in the read / write medium / memory 1306, one or more hardware components connected to the processor 1304, or some combination thereof. The device 1378 may include means for performing the operations described above with reference to FIGS. 10(A), 10(B), 11(A), 11(B), and 12. The above means may be one or more of the above components of the device 1378 and / or the processing system 1314 of the device 1378, configured to perform the functions described by the above means. It is understood that the specific order or hierarchical structure of the blocks in the disclosed processes / flowcharts is illustrative of example methods. Based on design preferences, it is understood that the specific order or hierarchical structure of the blocks in the processes / flowcharts can be rearranged. In addition, some blocks may be combined or omitted. The accompanying method claims present the elements of the various blocks in an example order and are not meant to be limited to the specific order or hierarchical structure presented. The foregoing description is provided to enable any person skilled in the art to practice the various aspects described herein. Various modifications to these aspects will be apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects. Thus, the claims are not intended to be limited to the aspects shown herein, but are to be accorded the full scope consistent with the claim language, where a singular reference to an element does not mean "only one" unless specifically so stated, but rather "one or more." The term "exemplary" is used herein to mean "serving as an example, instance, or illustration." Any aspect described as "exemplary" is not necessarily to be construed as preferred or advantageous over other aspects. Unless specifically stated otherwise, the term "some" refers to one or more. Combinations such as "at least one of A, B, or C," "one or more of A, B, or C," "at least one of A, B, and C," "one or more of A, B, and C," and "A, B, C, or any combination" include any combination of A, B, and / or C, and may include multiple A's, multiple B's, or multiple C's. Specifically, combinations such as "at least one of A, B, or C," "one or more of A, B, or C," "at least one of A, B, and C," "one or more of A, B, and C," and "A, B, C, or any combination" can be only A, only B, only C, A and B, A and C, B and C, or A and B and C, where any such combination may contain one or more members A, B, or C. All structural and functional equivalents of the various aspects described in this disclosure, those known or later coming to be known to those of ordinary skill in the art, are hereby incorporated herein by reference and are intended to be covered by the claims. In addition, nothing disclosed herein is intended to be dedicated to the public regardless of whether it is explicitly recited in the claims. The words "module," "mechanism," "element," "device," etc. are not intended to be substitutes for the word "means." Thus, no claim element shall be construed as a means-plus-function element unless the element is expressly recited using the phrase "means for." Those skilled in the art will readily observe that many modifications and changes can be made to the devices and methods while retaining the teachings of the invention. Accordingly, the above disclosure should be construed as limited only by the appended claims. 100: Schematic diagram of on-chip or off-chip power model building 110: Chip 110 120-1, 120-2,..., 120-n: Signals 150: Power model 214: Cell library 222: Front-end tool 210: Register Transfer Level (RTL) description 212: Design constraints 230: Pre-layout netlist 232: Design rules 232 and cell library 234 are sent to back-end tool 242 together 234: Cell library 242: Back-end tool 244: Post-layout netlist 260: Power analysis tool 270: Power report 300: Power model framework 314: Data collection component 316: Data preprocessing component 312: Data signal 332: Agent selection component 334: Coefficient tuning component 336: Quantization component 350: Power model 410: Orthogonal minimum data set 402: Array of performance monitoring unit signals 404: Embedded performance monitoring unit 406: Orthogonal clustering 422: Original data set 422 406-1, 406-2, 406-3: Orthogonal clustering 410: Orthogonal minimum data set 434-1, 434-2, 434-3: Clustering 426: Verification data set 436-1, 436-2, 436-3: Clustering 722-1, 722-2,..., 722-M: Sections 732-1, 732-2,..., 732-K: Sections 810, 820, 830, 840: Agent layout 950: Outer loop 1021, 1023, 1025, 1061, 1063, 1065, 1121, 1123, 1125, 1161, 1163: Phase-1 1031, 1033, 1035, 1071, 1073, 1075, 1131, 1133, 1135, 1171, 1173: Phase-2 1032: Tuning operation 1314: Processing system 1378: Device 1342: Collection component 1346: Training component 1348: Tuning component 1324: Bus 1304: Processor 1310: Network controller 1306: Read / write media / memory The present invention can be more comprehensively understood by combining the accompanying drawings and reading the following detailed description and examples, wherein: FIG. 1 is a schematic diagram illustrating the concept of on-chip or off-chip power model establishment. FIG. 2 is a schematic diagram illustrating the framework for generating a chip power consumption report. FIG. 3 is a schematic diagram illustrating the proposed power model framework. FIG. 4A is a flowchart illustrating the process of creating an orthogonal minimum dataset for training a power model. FIG. 4B is a schematic diagram illustrating the orthogonal minimum dataset. FIG. 5 is a schematic diagram illustrating another process of creating an orthogonal minimum dataset for training a power model. FIG. 6 is a flowchart illustrating the process of implementing similarity filtering to preprocess power model training data. FIG. 7A shows the power consumption patterns in the dataset before similarity filtering. FIG. 7B shows the power consumption patterns in the dataset after similarity filtering. FIG. 8 shows multiple agent layouts displaying bottom-up policy agent selection. FIG. 9 is a schematic diagram illustrating the process of agent selection using the PICASSO library. FIG. 10A is a schematic diagram illustrating the first type of process for agent selection and fine-tuning. FIG. 10B shows the second type of process for agent selection and fine-tuning. FIG. 11A is a schematic diagram illustrating the first type of process for agent selection and coefficient fine-tuning using the regularization path method. FIG. 11B shows the second type of process for agent selection and coefficient fine-tuning using the regularization path method. FIG. 12 is a flowchart illustrating a process (method) for quantifying a power model to reduce the hardware cost of implementing an on-chip power model. FIG. 13 is a schematic diagram illustrating a hardware implementation example of a device using a processing system. 100: Schematic diagram of on-chip or off-chip power model establishment 110: Chip 110 120-1, 120-2,... 120-n: Signals 150: Power model

Claims

1. A method for training a power model to estimate the power consumption of a chip, the method comprising: collecting a plurality of data samples, each data sample representing signal activity of a plurality of signals of the chip; selecting a subset of the plurality of signals as proxy signals related to the actual power consumption of the chip according to a criterion; training the power model using the signal activity of the plurality of signals as input and the actual power consumption as output; and fine-tuning the coefficients of the proxy signals in the power model to adjust the estimation error between the estimated power consumption output by the power model and the actual power consumption.

2. The method as claimed in claim 1, wherein collecting the plurality of data samples comprises: obtaining a T1 signal array from the wafer, the T1 array representing one or more signals at high voltage levels at one or more time points; applying a dimensionality reduction method to the T1 array to obtain a plurality of embeddings; applying a clustering method to the plurality of embeddings to obtain orthogonal clusters; and applying a sampling rule to the orthogonal clusters to obtain an orthogonal minimum dataset.

3. The method as described in claim 2, wherein the T1 array is obtained from the performance monitoring unit (PMU) of the wafer.

4. The method as described in claim 2, wherein the sampling rule is based on a group size.

5. The method as described in claim 2, further comprising: performing similarity filtering on the orthogonal clusters to remove highly similar data points within each cluster.

6. The method as claimed in claim 5, wherein similarity filtering comprises: calculating the similarity between each pair of data points within each cluster; concatenating the signal activity of each data point in the pair into a 1D array; calculating the similarity between the two 1D arrays; and removing one of the two data points if the similarity is higher than a threshold.

7. The method as described in claim 1, wherein selecting the subset as a proxy signal comprises: applying a bottom-up strategy to progressively select the proxy signal and propagating the selection to the next level until the entire design of the chip has been processed.

8. The method as described in claim 7, wherein applying the bottom-up strategy comprises: selecting proxy signals starting from the deepest level of the chip's design hierarchy; combining the proxy signal selected at the current level with a signal at the next level in the design hierarchy; and using the combined signal for another round of proxy signal selection.

9. The method as claimed in claim 1, wherein selecting the subset as the surrogate signal comprises: applying an optimization algorithm to select the surrogate signal, the optimization algorithm comprising: actively minimizing the coordinates in the active set; applying a strong rule to preselect the coordinates of the active set; updating the active set by adding the coordinates of the inactive set that have the largest absolute gradient value; and performing a warm-start initialization using the solution from the previous iteration as the starting point of the current iteration.

10. The method of claim 1, wherein training the power model comprises: performing a first phase comprising selecting the subset as a proxy signal using a regularization parameter; and performing a second phase comprising iteratively fine-tuning the coefficients of the proxy signal using a coefficient scaling parameter, wherein the regularization parameter decreases in one or more iterations of the first phase, and wherein the coefficient scaling parameter decreases in a plurality of sub-iterations corresponding to each iteration of the first phase.

11. The method as claimed in claim 1, wherein training the power model comprises: performing multiple rounds of training, each round including a first phase for agent signal selection and a second phase for coefficient fine-tuning; in the first phase, storing the initial outer loop results of agent selection in a file; and in the second phase, loading the stored results and performing coefficient fine-tuning.

12. The method as claimed in claim 1, wherein fine-tuning the coefficients of the proxy signal comprises: performing multiple rounds of training, each round comprising a first phase for proxy signal selection and a second phase for coefficient fine-tuning; in the first phase, decreasing a regularization parameter after each round; and in the second phase, iteratively fine-tuning the coefficients using a series of decreasing coefficient scaling parameters.

13. The method as claimed in claim 1, further comprising: quantizing the coefficients of the power model as integers; quantizing the signal activity used as input to the power model; and quantizing the power value output by the power model.

14. A computing apparatus for training a power model to estimate the power consumption of a chip, comprising: a memory; and at least one processor connected to the memory and configured to: collect a plurality of data samples, each data sample representing signal activity of a plurality of signals of the chip; select a subset of the plurality of signals as proxy signals related to the actual power consumption of the chip according to a criterion; train the power model using the signal activity of the plurality of signals as input and the actual power consumption as output; and fine-tune the coefficients of the proxy signals in the power model to adjust the estimation error between the estimated power consumption output by the power model and the actual power consumption.

15. The computing apparatus of claim 14, wherein, in order to collect a plurality of data samples, at least one processor is further configured to: acquire a T1 array from a wafer representing a plurality of signals at one or more time points that are at a high voltage level; apply a dimensionality reduction method to the T1 array to obtain a plurality of embeddings; apply a clustering method to the plurality of embeddings to obtain orthogonal clusters; and apply a sampling rule to the orthogonal clusters to obtain an orthogonal minimum data set.

16. The computing device as claimed in claim 15, wherein the T1 array is obtained from the performance monitoring unit (PMU) of the wafer.

17. The computing device as claimed in claim 15, wherein the sampling rule is based on the group size.

18. The computing apparatus of claim 15, wherein at least one processor is further configured to: perform similarity filtering on orthogonal clusters to remove highly similar data points within each cluster.

19. The computing apparatus of claim 18, wherein, for similarity filtering, at least one processor is further configured to: calculate the similarity between each pair of data points within each cluster; concatenate the signal activity of each data point into a 1D array; calculate the similarity between two 1D arrays; and remove one of the two data points if the similarity is higher than a threshold.

Citation Information

Patent Citations

  • Design method of electric energy quality monitoring application-specific integrated circuits based on soft nucleus CPU technology

    CN100517341C

  • Sampling chip activity for real time power estimation

    CN102272735B

  • Physical experiment platform for power consumption analysis of cryptographic chip

    CN202584139U

  • Generating SAIF efficiently from hardware platforms

    US10621296B1

  • On-Chip Power Proxy Based Architecture

    US20100268975A1