Electromagnetic resonance risk prediction method and device for chip power supply network, and medium

Abstract

The invention discloses an electromagnetic resonance risk prediction method and device for a chip power supply network, and a medium, and relates to the technical field of chip power supply networks, and the method comprises the steps: firstly obtaining the historical operation data of a target chip power supply network, including a grid unit and an electromagnetic parameter thereof, and a first impact factor affecting the change of the electromagnetic parameter; meanwhile, obtaining the resonant frequency of a region corresponding to each grid unit and a second influence factor influencing the change of the resonant frequency, constructing a target model based on the operation data and the first influence factor, and predicting a resonance sensitive region in combination with a quantum bit mapping rule; the resonance intensity is predicted through the resonance frequency, the second influence factor and the physical constraint neural network model, a resonance risk value is calculated through synchronous iteration till the resonance risk value is larger than a preset target value, and a prediction result is output. And the risk prediction accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of chip power delivery network, more particularly, to a chip power delivery network electromagnetic resonance risk prediction method, device and medium. BACKGROUND

[0002] With the rapid development of integrated circuit (IC) technology, the integration and operating frequency of chips are continuously improving, and the electromagnetic compatibility (EMC) problem of chip power delivery network (PDN) has gradually become one of the key factors restricting the performance improvement of chips. The main function of PDN is to provide stable power supply for chips, but its complex electromagnetic environment and high-frequency signal transmission characteristics may cause electromagnetic resonance phenomenon, which in turn causes power supply noise, signal interference and chip performance degradation and other problems. Therefore, accurately predicting the electromagnetic resonance risk of PDN is of great significance for chip design and optimization.

[0003] In the prior art, when predicting the electromagnetic resonance risk of PDN, a simplified electromagnetic model or an analysis method based on a single parameter is usually used. Although these methods can provide preliminary evaluation results to some extent, they have many limitations: first, they ignore the differences in electromagnetic parameters in different regions of PDN and the interaction between these parameters, resulting in inaccurate prediction results; second, traditional methods are difficult to effectively identify resonance sensitive regions in PDN, i.e. regions that are more susceptible to electromagnetic resonance under certain conditions, so they cannot be optimized for design; in addition, there is no effective means to predict the resonance propagation path, so that the overall propagation process of electromagnetic resonance is not fully understood; finally, the existing technology fails to fully combine physical constraints when predicting resonance strength, resulting in a large deviation between the predicted results and the actual situation, which makes it difficult to meet the high requirements of modern chip design for electromagnetic compatibility.

[0004] Therefore, the prior art has insufficient prediction accuracy, cannot accurately identify resonance sensitive regions and propagation paths, and does not fully combine physical constraints. SUMMARY

[0005] In order to overcome the problems of insufficient prediction accuracy, inability to accurately identify resonance sensitive regions and propagation paths, and failure to fully combine physical constraints in the prior art, the present application discloses a chip power delivery network electromagnetic resonance risk prediction method, device and medium, which can effectively solve the above technical problems.

[0006] To solve the above technical problems, the technical solution of the present application is as follows:

[0007] A chip power delivery network electromagnetic resonance risk prediction method, the method comprising:

[0008] obtain running data of the target chip power supply network at a historical time point, the running data comprising electromagnetic parameters of each grid unit and a region corresponding to each grid unit, each grid unit being a unit of a PDN grid of the target chip power supply network at the historical time point;

[0009] obtain a first influence factor of each grid unit, the first influence factor being used to indicate a factor influencing a change in electromagnetic parameters of a region corresponding to the grid unit;

[0010] obtain a resonance frequency of a region corresponding to each grid unit and a second influence factor of each grid unit, the second influence factor being used to indicate a factor influencing a change in the resonance frequency of the region corresponding to the grid unit;

[0011] synchronously and iteratively predict local resonance and global resonance of the target chip power supply network according to the running data, the first influence factor of each grid unit, the resonance frequency of each grid unit, and the second influence factor of each grid unit, until a resonance risk of the target chip power supply network meets a preset condition, to obtain an electromagnetic resonance risk prediction result of the target chip power supply network.

[0012] Preferably, the obtaining of the electromagnetic resonance risk prediction result of the target chip power supply network comprises:

[0013] predicting a resonance sensitive region on a PDN grid of a target region according to the running data and the first influence factor of each grid unit, the target region being a region comprising the target chip power supply network;

[0014] predicting a resonance intensity of the target region according to the resonance frequency of each grid unit and the second influence factor of each grid unit;

[0015] determining a resonance risk value of the target region at a current iteration according to the resonance sensitive region and the resonance intensity;

[0016] if the resonance risk value is greater than a preset target risk value, obtaining the electromagnetic resonance risk prediction result of the target chip power supply network, and if the resonance risk value is less than the preset target risk value, continuing iteration until the resonance risk value is greater than the preset target risk value, to obtain the electromagnetic resonance risk prediction result of the target chip power supply network.

[0017] Preferably, the predicting of the resonance sensitive region on the PDN grid of the target region according to the running data and the first influence factor of each grid unit comprises:

[0018] input each of the grid cells and the first influence factor corresponding to each of the grid cells into a target model constructed based on a preset mapping rule of a PDN grid to a qubit, and predict a sensitive probability of each of the grid cells for different resonance levels;

[0019] The target model of the mapping rule of the PDN grid to the qubit is trained according to a plurality of sampling grid cells, electromagnetic parameters corresponding to the plurality of sampling grid cells, and resonance levels.

[0020] According to the sensitive probability of each of the grid cells for different resonance levels, the resonance level of each of the grid cells is predicted.

[0021] The grid cell with the high-sensitive level is determined as a target grid cell, and a resonance-sensitive area on a PDN grid of the target chip power supply network is obtained according to the area corresponding to each of the target grid cells.

[0022] Preferably, the resonance level of each of the grid cells is predicted according to the sensitive probability of each of the grid cells for different resonance levels, including:

[0023] The neighborhood coupling degree, the parameter conversion cost, and the risk inertia coefficient of each grid cell are obtained, the neighborhood coupling degree is the correlation degree of each resonance level of the grid cell in the preset area to which the grid cell belongs, the parameter conversion cost is used to indicate the difficulty of changing the resonance level of the area corresponding to the grid cell, and the risk inertia coefficient is determined according to the target risk value.

[0024] The target overall sensitive probability of the grid cell for different resonance levels is obtained according to the sensitive probability of the grid cell for different resonance levels, the neighborhood coupling degree, the parameter conversion cost, and the risk inertia coefficient.

[0025] The resonance level of each of the grid cells is predicted according to the target overall sensitive probability of the grid cell for different resonance levels and a preset screening algorithm.

[0026] Preferably, after the resonance level of each of the grid cells is predicted according to the target overall sensitive probability of the grid cell for different resonance levels and the preset screening algorithm, the method further includes:

[0027] Each of the target grid cells is input into a model trained based on a spatio-temporal compressed sensing measurement matrix, and a resonance propagation path of the target grid cell is predicted.

[0028] The model based on the spatio-temporal compressive sensing measurement matrix is trained according to a sampling target grid unit and a third influence factor corresponding to the sampling target grid unit, and the third influence factor is used to indicate an influence factor of a region corresponding to the sampling target grid unit forming different resonant propagation paths.

[0029] Preferably, the predicting the resonant intensity of the target region according to each of the resonant frequencies and each of the second influence factors comprises: for each of the target grid units, inputting each of the target grid units, a resonant propagation path corresponding to the target grid unit and a second influence factor into a preset neural network model based on physical constraints respectively, to predict a reference resonant intensity of a region corresponding to each of the target grid units.

[0030] Preferably, the neural network model based on physical constraints is trained according to a sampling target grid unit and a second influence factor corresponding to the sampling target grid unit.

[0031] According to the reference resonant intensity, the resonant intensity of the target region is predicted.

[0032] Preferably, the predicting the resonant intensity of the target region according to the reference resonant intensity comprises:

[0033] Obtaining a domain interference value of the resonant intensity of a region corresponding to each of the target grid units;

[0034] According to the reference resonant intensity of each of the target grid units and the domain interference value, the resonant intensity of the target region is predicted.

[0035] Preferably, an electronic device comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to implement the prediction method described above.

[0036] Preferably, a computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the prediction method described above.

[0037] Compared with the prior art, the beneficial effects of the present application are: the present application aims at the problem that the prior art cannot accurately identify the resonance sensitive area, uses a target model based on a PDN grid to qubit mapping rule, combines operation data and a first influence factor, accurately predicts the resonance level of each grid unit, and further screens out grid units with high sensitivity level by comprehensively considering the neighborhood coupling degree, parameter conversion cost and risk inertia coefficient, so as to accurately locate the resonance sensitive area; in view of the problem that the prior art is difficult to accurately predict the resonance propagation path, the present application introduces a model based on a spatiotemporal compressed sensing measurement matrix, accurately predicts the resonance propagation path by analyzing the target grid unit and the corresponding third influence factor, and makes up for the shortcomings of the prior art; in addition, in view of the problem that the prior art does not fully combine physical constraint conditions, resulting in insufficient prediction accuracy, the present application adopts a neural network model based on physical constraints, takes the resonance propagation path of the target grid unit and the second influence factor as input, combines the reference resonance intensity and the field interference value, predicts the resonance intensity of the target area, and ensures that the prediction result conforms to the actual physical law; by synchronously iterating to predict local resonance and global resonance, combining the resonance sensitive area and the resonance intensity, and finally determining the resonance risk value, the prediction accuracy is further improved, and in summary, the present application solves the shortcomings of the prior art by multi-dimensional data fusion, advanced model construction and sufficient combination of physical constraint conditions, and provides a more accurate and reliable technical means for electromagnetic resonance risk prediction of a chip power supply network. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only exemplary, and other drawings can be derived from the provided drawings without creative labor for those skilled in the art.

[0039] Figure 1 A flowchart of a chip power supply network electromagnetic resonance risk prediction method provided by the embodiment of the present application is shown in the figure.

[0040] Figure 2 A flowchart of predicting a resonance sensitive area provided by the embodiment of the present application is shown in the figure.

[0041] Figure 3 A flowchart of predicting a resonance intensity provided by the embodiment of the present application is shown in the figure.

[0042] Figure 4 A structural schematic diagram of an electronic device provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0043] The drawings are only used for illustrative description and cannot be understood as a limitation of the present patent;

[0044] For better illustrating the embodiments, some components in the drawings can be omitted, enlarged or reduced, which do not represent the actual product size.

[0045] It is understandable for those skilled in the art that some well-known structures and their descriptions in the drawings can be omitted.

[0046] The technical solutions of the present application will be further described below in combination with the drawings and embodiments.

[0047] Embodiment 1

[0048] Please refer to Figure 1 , Figure 1 A flowchart of a chip power supply network electromagnetic resonance risk prediction method provided by the embodiment of the present application, the method comprises the following steps:

[0049] A chip power supply network electromagnetic resonance risk prediction method, the method comprises:

[0050] Obtain the running data of the target chip power supply network at a historical time, the running data comprising electromagnetic parameters of each grid unit and the region corresponding to each grid unit, each grid unit being a unit of the PDN grid of the target chip power supply network at the historical time;

[0051] Obtain the first influence factor of each grid unit, the first influence factor being used to indicate the factors affecting the change of electromagnetic parameters of the region corresponding to the grid unit;

[0052] Obtain the resonance frequency of the region corresponding to each grid unit and the second influence factor of each grid unit, the second influence factor being used to indicate the factors affecting the change of resonance frequency of the region corresponding to the grid unit;

[0053] According to the running data, each first influence factor, each resonance frequency and each second influence factor, the local resonance and global resonance of the target chip power supply network are synchronously iteratively predicted until the resonance risk of the target chip power supply network meets the preset condition, and the electromagnetic resonance risk prediction result of the target chip power supply network is obtained.

[0054] The electromagnetic resonance risk prediction result of the target chip power supply network comprises:

[0055] According to the running data and each first influence factor, the resonance sensitive region on the PDN grid of the target region is predicted, the target region being a region comprising the target chip power supply network;

[0056] According to each resonance frequency and each second influence factor, the resonance intensity of the target region is predicted.

[0057] determining a resonance risk value of the target region in the current iteration according to the resonance sensitive region and the resonance intensity;

[0058] if the resonance risk value is greater than a preset target risk value, obtaining an electromagnetic resonance risk prediction result of the target chip power supply network, and if the resonance risk value is less than the preset target risk value, continuing iteration until the resonance risk value is greater than the preset target risk value, and obtaining the electromagnetic resonance risk prediction result of the target chip power supply network.

[0059] The resonance sensitive region on the PDN grid of the target region is predicted according to the running data and each first influence factor.

[0060] Each grid unit and the corresponding first influence factor of each grid unit are input into a target model constructed based on a preset mapping rule from a PDN grid to a qubit, and a sensitive probability of each grid unit for different resonance levels is predicted.

[0061] The target model of the mapping rule from the PDN grid to the qubit is trained according to a plurality of sampling grid units, corresponding electromagnetic parameters and resonance levels of the plurality of sampling grid units.

[0062] The resonance level of each grid unit is predicted according to the sensitive probability of each grid unit for different resonance levels.

[0063] The grid unit with a high sensitive level is determined as a target grid unit, and a resonance sensitive region on a PDN grid of the target chip power supply network is obtained according to a region corresponding to each target grid unit.

[0064] The resonance level of each grid unit is predicted according to the sensitive probability of each grid unit for different resonance levels.

[0065] The neighborhood coupling degree, the parameter conversion cost and the risk inertia coefficient of each grid unit are obtained, the neighborhood coupling degree is the correlation degree of each resonance level of the grid unit in a preset region, the parameter conversion cost is used to indicate the difficulty of changing the resonance level of the corresponding region of the grid unit, and the risk inertia coefficient is determined according to the target risk value.

[0066] The target overall sensitive probability of the grid unit for different resonance levels is obtained according to the sensitive probability of the grid unit for different resonance levels, the neighborhood coupling degree, the parameter conversion cost and the risk inertia coefficient.

[0067] According to the target overall sensitive probability of each grid unit for different resonance levels and the preset screening algorithm, the resonance level of each grid unit is predicted.

[0068] After the resonance level of each grid unit is predicted according to the target overall sensitive probability of each grid unit for different resonance levels and the preset screening algorithm, the method further comprises:

[0069] Each target grid unit is input into a trained model based on a spatio-temporal compressed sensing measurement matrix to predict the resonance propagation path of the target grid unit.

[0070] The model based on the spatio-temporal compressed sensing measurement matrix is trained according to a sampling target grid unit and a third influence factor corresponding to the sampling target grid unit, and the third influence factor is used to indicate the influence factor of the area corresponding to the sampling target grid unit forming different resonance propagation paths.

[0071] The resonance intensity of the target area is predicted according to each resonance frequency and each second influence factor, which comprises: for each target grid unit, each target grid unit and the resonance propagation path and second influence factor corresponding to the target grid unit are respectively input into a preset neural network model based on physical constraints to predict the reference resonance intensity of the area corresponding to each target grid unit.

[0072] The neural network model based on physical constraints is trained according to a sampling target grid unit and a second influence factor corresponding to the sampling target grid unit.

[0073] The resonance intensity of the target area is predicted according to the reference resonance intensity.

[0074] Preferably, the resonance intensity of the target area is predicted according to the reference resonance intensity, which comprises:

[0075] The domain interference value of the resonance intensity of the area corresponding to each target grid unit is obtained.

[0076] The resonance intensity of the target area is predicted according to the reference resonance intensity and the domain interference value of each target grid unit.

[0077] First, start the data collection process, link multiple sensors and data collection equipment, monitor the target chip power supply network in all directions, run data collection to cover different working modes of the chip, such as standby mode and high load operation mode electromagnetic parameters, ensure that the data can reflect the complex working conditions of the actual operation of the chip, the first influence factor collection also pays attention to comprehensiveness, through the temperature sensor network to monitor the chip surface and surrounding environment temperature distribution, humidity sensor is arranged at the key position of the circuit board, material characteristic parameters can be obtained from chip design documents and material test reports. Integrate these multi-dimensional data to build an initial data set for iterative prediction.

[0078] For the collected operation data, the system calls the spectrum analysis module, uses the fast Fourier transform (FFT) algorithm to convert the time domain signals such as voltage and current of each grid unit into frequency domain, extracts the resonance frequency component, and determines the resonance frequency value corresponding to each grid unit. Collect the second influence factor, the frequency change rate is obtained by calculating the difference between the resonance frequencies of adjacent time points and the ratio of the time interval; external electromagnetic interference is measured by an electromagnetic interference detector, and the interference signal strength, frequency and other parameters in different frequency bands are measured. The resonance frequency and the second influence factor are associated with the grid unit to form a grid unit-resonance frequency-second influence factor data group, which provides input for synchronous iterative prediction.

[0079] Based on the operation data, the first influence factor, the resonance frequency and the second influence factor, the prediction model of local resonance and global resonance is constructed, the local resonance model focuses on the resonance characteristics of a single grid unit or a small area grid, and considers the internal electromagnetic interaction of the grid unit; The global resonance model focuses on the entire chip power supply network, analyzes the coupling, propagation and superposition effects of different area resonances, and recalculates the resonance risk value according to the current data in each iteration during the iteration process. Use numerical iterative algorithms such as Newton iteration method and gradient descent method to adjust model parameters and optimize prediction results. For example, in the first iteration, the model predicts the local resonance strength and global resonance distribution based on the initial data, and in the second iteration, the input parameter weight and model internal coefficient are corrected according to the error feedback of the first iteration, so that the prediction gradually approaches the true resonance situation, and the dynamic synchronous optimization of local and global resonance prediction is realized.

[0080] After each iteration is completed, the resonant risk value under the current iteration is calculated according to the preset resonant risk evaluation algorithm. The risk value calculation comprehensively considers the range of the resonant sensitive area, the proportion of the high sensitive area, the amplitude and distribution of the resonant intensity, and other factors, and constructs a risk evaluation function, such as R = αS + βI (where R is the resonant risk value, S is the risk contribution degree of the resonant sensitive area, I is the risk contribution degree of the resonant intensity, and α and β are weight coefficients. According to the chip application scene adjustment, the calculated resonant risk value is compared with the preset target risk value (determined according to the chip reliability requirement, industry standard, etc. For example, for a high-performance computing chip, the target risk value may be set to be low to strictly control the resonant influence). If the risk value is greater than the target risk value, it means that the current prediction can clearly reflect the electromagnetic resonant risk faced by the chip, and meets the demand of risk warning in actual application, and the iteration can be stopped. If it is less than the target risk value, it means that the prediction result is conservative, or the potential risk is not fully explored, and the iteration needs to be continued, the model parameters are adjusted, and the data details are supplemented until the risk value meets the standard.

[0081] When the resonant risk meets the condition (the risk value is greater than the target risk value), the system integrates the resonant sensitive area (clearly defines the boundaries, area and position of the high, medium and low sensitive areas) and the resonant intensity (gives the resonant intensity values and distribution heat map of different areas) obtained in the iteration process, generates an electromagnetic resonant risk prediction result document according to the preset data report template. The document includes text description, such as risk level judgment, key risk area description, data table (sensitive probability and resonant intensity value of each grid unit), graphical display, such as resonant sensitive area distribution graph and resonant intensity three-dimensional graph, and is presented to the chip design engineer through the display, or is stored in the database for long-term analysis, to provide accurate and comprehensive reference basis for subsequent design optimization and fault troubleshooting of the chip.

[0082] Embodiment 2

[0083] Please refer to Figure 2 , Figure 2 The flowchart for predicting the resonant sensitive area provided by the embodiment of the application further includes, as an embodiment of the embodiment of the application, the following steps:

[0084] The detailed information of each grid unit collected in the early stage (including the physical size of the grid unit, the coordinate position in the chip, the connection relationship with the surrounding circuit, etc.) and the corresponding first influence factor, such as temperature value accurate to 0.1℃, humidity value in percentage form, material characteristic parameter in conductivity, dielectric constant, etc. are input through a data input interface, such as file import, real-time sensor data transmission input system. The system checks the input data for completeness (whether there is missing grid unit data, abnormal influence factor value) and accuracy (whether the temperature value is within a reasonable physical range, whether the material parameter is consistent with the actual chip), providing a data basis for model operation.

[0085] According to the mapping rule of the PDN grid to the qubit, the basic framework of the target model is built. The qubit has the characteristics of quantum state superposition and entanglement. The electromagnetic parameters of the PDN grid, such as voltage and current, can be quantized and encoded into the state vector of the qubit, establishing the correspondence between the quantum state and the grid electromagnetic state. A large amount of sampled grid unit data (covering grid unit samples under different chip types and different working conditions) are collected, and the electromagnetic parameters and resonance levels (through actual testing, such as detecting the resonance amplitude with a spectrum analyzer, and determining the resonance level according to expert experience) of these samples are extracted as a training data set. Quantum machine learning algorithms such as quantum support vector machine and quantum neural network are used to input the training data set into the model, adjust the quantum gate parameters and quantum state evolution rules in the model, and make the model learn the potential correlation between the PDN grid electromagnetic parameters and the resonance level. The model training is completed. During the training process, the loss function such as mean square error loss is used to measure the deviation between the predicted resonance level and the actual resonance level to monitor the model performance. The parameters are continuously optimized until the prediction accuracy of the model meets the preset requirements, and a stable and accurate target model is constructed.

[0086] The verified grid unit and first influence factor data are converted according to the input format requirements of the target model, such as quantum state encoding format and data normalization processing, and input into the trained target model. The model uses quantum state evolution algorithm to simulate the changes of grid unit electromagnetic state under the action of different first influence factors, calculates the sensitive probability of each grid unit for different resonance levels such as slight resonance (small resonance amplitude, short duration), moderate resonance (resonance affecting part of the chip function), and severe resonance (leading to chip failure), and uses the advantages of quantum parallel computing to process the probability prediction of multiple resonance levels simultaneously, improving the calculation efficiency. The output sensitive probability is presented in the form of probability distribution table.

[0087] To more accurately determine the resonance level of the grid cell, the system additionally obtains the neighborhood coupling degree, parameter conversion cost and risk inertia coefficient of each grid cell. The neighborhood coupling degree is calculated by analyzing the electromagnetic signal interaction between the grid cell and other grid cells within a certain range, such as the adjacent 3x3 grid region, and the correlation frequency and correlation strength of the resonance level of the adjacent grid cells are counted to obtain the coupling degree value. The parameter conversion cost considers the difficulty and cost of modifying the hardware design corresponding to the voltage adjustment amplitude and current change amount required to change the resonance level of the grid cell, and a cost evaluation model is constructed to calculate the cost based on chip manufacturing process knowledge. The risk inertia coefficient is inversely deduced by a risk assessment algorithm based on a preset target risk value. The system fuses these parameters with the sensitive probability to construct a resonance level prediction function, for example, using a weighted summation method to multiply the sensitive probability of different resonance levels by the corresponding weight (the weight is dynamically adjusted according to the neighborhood coupling degree, parameter conversion cost and risk inertia coefficient), to obtain the target overall sensitive probability of each grid cell for different resonance levels. Then, a preset screening algorithm, such as the maximum probability method, is used to select the resonance level with the maximum target overall sensitive probability as the resonance level of the grid cell. Or, the threshold determination method is used to determine the corresponding level when the target overall sensitive probability of a certain resonance level exceeds the set threshold. The resonance level of each grid cell is accurately predicted, and the grid cells with "high resonance level", "medium resonance level" and "low resonance level" are marked.

[0088] The system iterates through the resonance level determination results of all grid cells, selects the grid cells with high sensitive level as target grid cells, and uses a region clustering algorithm, such as the density-based clustering algorithm DBSCAN, to identify the high-density target grid cell region based on the coordinate position and distribution range of the target grid cells in the chip power distribution network PDN grid. The continuous region corresponding to these target grid cells is determined and marked as the resonance sensitive region on the target chip power distribution network PDN grid. The detailed information of the sensitive region, including the region boundary coordinates, area size and specific position in the chip, such as the corresponding chip function module region, is output to provide clear spatial guidance for chip design optimization.

[0089] Embodiment 3

[0090] Please refer to Figure 3 , Figure 3 The flowchart for predicting the resonance strength provided by the embodiments of the present application further includes, as an embodiment of the embodiments of the present application:

[0091] Collect the precise resonant frequency data and the second influence factor of each grid unit, and input the data into the system through data collection software, hardware interface, etc. according to the system data input specification. Perform validity check on the input data to exclude abnormal values, such as resonant frequency exceeding the chip operating frequency range and unreasonable influence factor value, to ensure data quality.

[0092] For each target grid unit (i.e. high-sensitive grid unit selected in the resonant sensitive area), the system calls a neural network model based on physical constraints. The model takes the sampled target grid unit as the training sample, inputs the resonant frequency and the second influence factor of the sample, and outputs the measured resonant intensity value of the corresponding area. The model parameters are trained through back propagation algorithm, and physical laws such as Maxwell's equations are integrated as constraint conditions to ensure that the model prediction conforms to the basic principles of electromagnetism. The resonant frequency and the second influence factor of the target grid unit are preprocessed according to the model input requirements, such as data standardization and feature engineering processing, and then input into the model. The model uses convolutional neural networks, recurrent neural networks, etc. to simulate the physical processes of electromagnetic resonance generation, propagation, and attenuation, taking into account the second influence factor such as external electromagnetic interference which may introduce additional resonance sources and change the resonant intensity. The model predicts the reference resonant intensity of the corresponding area of each target grid unit, and the output reference resonant intensity is presented in numerical form, marking the resonant intensity of each target grid unit area.

[0093] To predict the resonant intensity in line with the actual scenario, the system needs to obtain the domain interference value of the resonant intensity of the corresponding area of each target grid unit. The domain interference value is derived from electromagnetic interference between different functional modules inside the chip, such as interference of high-speed digital circuits on analog power supply networks, external environmental electromagnetic interference such as radio frequency interference in laboratory environment and electromagnetic noise in industrial environment. By arranging multiple groups of electromagnetic interference sensors, the system collects interference signals of the target grid unit area, uses frequency spectrum analysis and noise separation algorithm to extract interference components related to resonant intensity, and quantitatively calculates the domain interference value. For example, Fourier transform is used to separate the amplitude of interference signals at a specific frequency, and the coupling coefficient of interference signals and resonant signals is used to calculate the influence of interference on resonant intensity to obtain the domain interference value.

[0094] The system fuses and calculates the reference resonant intensity of each target grid unit with the corresponding domain interference value, using linear superposition, nonlinear correction, etc. to consider the enhancement or weakening effect of domain interference on resonant intensity, such as when the interference signal and the resonant signal are at the same frequency and in the same direction, the resonant intensity is enhanced; when they are in opposite directions, the intensity is weakened. The reference resonant intensity is corrected to obtain the resonant intensity prediction value of the target area (the chip power supply network area covering all target grid units). The final output prediction result is displayed in the form of a resonant intensity distribution map, presenting the spatial distribution of resonant intensity in the target area, providing data support for electromagnetic resonance risk assessment and chip design optimization.

[0095] Referring to Figure 4 , Figure 4 The application further provides an electronic device, and a processor is a control unit of an entire electromagnetic resonance risk prediction process. A program code specially written for electromagnetic resonance risk prediction needs to be called from a memory. The program code is pre-stored in the memory and covers data processing algorithms, model calling logic, iteration control rules and the like. The processor reads program instructions through a bus, loads the program instructions into an internal instruction cache, decodes and executes the instructions in sequence, provides software logic support for each step, and ensures that each link is orderly developed.

[0096] The processor establishes a connection with a data acquisition unit of a chip power supply network by relying on a hardware interface such as a high-speed serial interface, a parallel data bus or a communication module such as a board-level communication protocol module. For a target chip power supply network, running data of a historical moment is collected according to a preset time interval such as a millisecond level or a second level and according to a chip running scene requirement. The running data includes electromagnetic parameters of each grid unit and a corresponding area, wherein:

[0097] A voltage parameter is collected by a voltage sensor.

[0098] A current parameter is obtained by means of a current transformer and the like.

[0099] An impedance parameter is calculated by injecting a specific frequency test signal and combining voltage and current changes.

[0100] The processor uses the obtained running data and a first influence factor (temperature influences conductivity and dielectric constant of a chip material, humidity influences insulation performance of a circuit, material characteristics such as a metal wire material and a semiconductor substrate doping concentration and the like) to call a target model constructed based on a PDN grid to qubit mapping rule. The model is trained by a large number of samples, associates and maps PDN grid physical parameters and qubit quantum state characteristics, and converts grid unit electromagnetic parameters into a quantum state representation recognizable by the model.

[0101] The processor encodes input data into a quantum state and then inputs the quantum state into a model operation to simulate resonance responses of grid units in different electromagnetic environments and to predict sensitive probabilities of the grid units for different levels such as slight resonance, moderate resonance and severe resonance. An operation process involves complex mathematical operations such as matrix operation and quantum state evolution simulation and is completed by relying on an arithmetic logic unit (ALU) and a floating point unit (FPU) of the processor. A sensitive probability distribution of each grid unit is output, and potential resonance sensitive grid units are marked.

[0102] Based on the resonance frequency of each grid unit (extracted by Fourier transform of operation data through spectrum analysis algorithm, etc.) and the second influence factor (frequency change rate reflecting the fluctuation of resonance frequency over time, external electromagnetic interference such as signal interference of surrounding radio frequency equipment, etc.), the processor calls a neural network model based on physical constraints, which integrates the basic laws of electromagnetics such as Maxwell's equations, to learn the mapping relationship between resonance intensity and input parameters with the training samples of sampled grid unit data and corresponding second influence factors.

[0103] The processor inputs the resonance frequency and the second influence factor after normalization and other pretreatments into the model, which simulates the transmission, superposition and attenuation process of resonance energy in the actual electromagnetic environment through a multi-layer neural network, predicts the resonance intensity of the target area (covering the chip power supply network and the surrounding associated circuit area), and for example, when considering external electromagnetic interference, the model simulates the coupling effect of interference signals and internal resonance signals of the chip, and corrects the resonance intensity prediction value to fit the actual scenario.

[0104] The processor integrates the predicted resonance sensitive area (marked with the distribution of high, medium and low sensitive grid units) and the resonance intensity (presented with quantitative indicators such as resonance energy density and amplitude), and outputs it to the display through the display interface, stores it in the memory, or uploads it to the chip design optimization system through the communication interface in a preset data format, such as a table of sensitive area coordinates and resonance levels, a resonance intensity spatial distribution chart. These results provide a reference for chip designers, who can avoid high sensitive areas during layout and routing, or optimize the filter circuit for high resonance intensity areas during power module design, reducing the impact of electromagnetic resonance on chip performance.

[0105] The same or similar reference signs correspond to the same or similar components;

[0106] The terms used to describe the positional relationship in the drawings are only for illustrative purposes and should not be construed as a limitation on the patent;

[0107] Obviously, the above embodiments of the present application are only examples for clearly illustrating the present application, and are not intended to limit the embodiments of the present application. For those skilled in the art, on the basis of the above description, other different forms of changes or variations can also be made, and here it is not necessary or impossible to exhaust all the embodiments, and any modification, equivalent replacement and improvement made within the spirit and principles of the present application should be included in the protection scope of the claims of the present application.

Claims

1. A method for predicting electromagnetic resonance risk in a chip power supply network, characterized in that, The method includes: Obtain the operating data of the target chip power supply network at a historical time. The operating data includes the electromagnetic parameters of each grid cell and the region corresponding to each grid cell. Each grid cell is a cell of the PDN grid of the target chip power supply network at the historical time. Obtain a first influence factor for each of the grid cells, wherein the first influence factor is used to indicate the factors affecting the change of electromagnetic parameters in the corresponding region of the grid cell; Obtain the resonant frequency of the region corresponding to each grid cell and the second influence factor of each grid cell. The second influence factor is used to indicate the factors that affect the change of the resonant frequency of the region corresponding to the grid cell. Based on the operating data, each of the first influencing factors, each of the resonant frequencies, and each of the second influencing factors, the local resonance and global resonance of the target chip power supply network are synchronously iteratively predicted until the resonance risk of the target chip power supply network meets the preset conditions, thereby obtaining the electromagnetic resonance risk prediction result of the target chip power supply network.

2. The prediction method according to claim 1, characterized in that, The electromagnetic resonance risk prediction results obtained for the power supply network of the target chip include: Based on the operational data and each of the first influencing factors, the resonant sensitive region on the PDN grid of the target region is predicted, wherein the target region is the region including the power supply network of the target chip. Based on each of the resonant frequencies and each of the second influencing factors, the resonant intensity of the target region is predicted; Determine the resonance risk value of the target region under the current iteration based on the resonance sensitive region and the resonance intensity; If the resonance risk value is greater than the preset target risk value, the electromagnetic resonance risk prediction result of the target chip power supply network is obtained. If the resonance risk value is less than the preset target risk value, the iteration continues until the resonance risk value is greater than the preset target risk value, and the electromagnetic resonance risk prediction result of the target chip power supply network is obtained.

3. The prediction method according to claim 2, characterized in that, The step of predicting the resonance-sensitive region on the PDN grid of the target region based on the operational data and each of the first influencing factors includes: Each of the grid cells and the first influence factor corresponding to each of the grid cells are input into a target model constructed based on a preset mapping rule from PDN grid to qubit, and the sensitivity probability of each of the grid cells for different resonance levels is predicted. The target model of the PDN grid-to-qubit mapping rule is trained based on multiple sampling grid cells, the electromagnetic parameters and resonance levels corresponding to the multiple sampling grid cells; The resonance level of each grid cell is predicted based on the sensitivity probability of each grid cell to different resonance levels. The grid cells with high resonance sensitivity are identified as target grid cells, and the resonance sensitive regions on the PDN grid of the target chip power supply network are obtained according to the regions corresponding to each target grid cell.

4. The prediction method according to claim 3, characterized in that, The step of predicting the resonance level of each grid cell based on the sensitivity probability of each grid cell to different resonance levels includes: The neighborhood coupling degree, parameter conversion cost, and risk inertia coefficient of each grid cell are obtained. The neighborhood coupling degree is the degree of correlation between the resonance levels of the grid cell in its preset region. The parameter conversion cost is used to indicate the ease or difficulty of the resonance level change in the corresponding region of the grid cell. The risk inertia coefficient is determined based on the target risk value. The overall target sensitivity probability of the grid cell for different resonance levels is obtained based on the sensitivity probability of the grid cell for different resonance levels, the neighborhood coupling degree, the parameter conversion cost, and the risk inertia coefficient. Based on the overall sensitivity probability of the grid cells to different resonance levels and the preset screening algorithm, the resonance level of each grid cell is predicted.

5. The prediction method according to claim 4, characterized in that, After predicting the resonance level of each grid cell based on the overall sensitivity probability of the grid cell to different resonance levels and a preset screening algorithm, the method further includes: Each of the target grid cells is input into a trained model based on a spatiotemporal compressed sensing measurement matrix to predict the resonant propagation path of the target grid cell; The model based on the spatiotemporal compressed sensing measurement matrix is ​​trained based on the sampling target grid cell and the third influence factor corresponding to the sampling target grid cell. The third influence factor is used to indicate the factors that influence the formation of different resonance propagation paths in the region corresponding to the sampling target grid cell.

6. The prediction method according to claim 5, characterized in that, The step of predicting the resonance intensity of the target region based on each of the resonant frequencies and each of the second influencing factors includes: for each of the target grid cells, inputting each of the target grid cells and the corresponding resonance propagation path and the second influencing factor into a preset physical constraint-based neural network model to predict the reference resonance intensity of the region corresponding to each of the target grid cells.

7. The prediction method according to claim 6, characterized in that, The physical constraint-based neural network model is trained based on the sampled target grid cell and the second influence factor corresponding to the sampled target grid cell. Based on the reference resonance intensity, the resonance intensity of the target region is predicted.

8. The prediction method according to claim 7, characterized in that, The step of predicting the resonance intensity of the target region based on the reference resonance intensity includes: Obtain the area interference value of the resonance intensity of the region corresponding to each target grid cell; The resonance intensity of the target region is predicted based on the reference resonance intensity of each target grid cell and the surrounding interference value.

9. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, implements the prediction method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the prediction method according to any one of claims 1-7.

Images