Characterization extraction method of circuit device and storage medium
Through a machine learning framework based on gradient enhancement, multiple sets of models are trained and the optimal super parameter combination is selected, which solves the problem of excessive consumption of feature extraction resources, and achieves acceleration and accuracy improvement of feature extraction, which significantly reduces the design cycle and cost.
Patent Information
- Application Number
- CN202510078938.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-17
- Publication Date
- 2025-04-25
AI Technical Summary
In the prior art, the consumption of featured extraction resources is too high, resulting in high time cost. Under advanced processes, static timing analysis requires enumeration of all possible states, resulting in excessive deviation between the library construction results and actual applications.
Using a machine learning framework based on gradient enhancement, we use the required files extracted by reading the time sequence arc feature, and generate the SPICE netlist, perform reproducible pseudo-random marking, form a training set and a prediction set, train multiple sets of models based on gradient enhancement, select the optimal super parameter combination, and generate the final model for timing or power consumption prediction.
It significantly accelerates the feature extraction process, reduces the consumption of computing resources, improves prediction accuracy and stability, reduces design cycle and material costs, and enhances the market competitiveness of the product.
Smart Images
Figure CN119886003A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of integrated circuit automated design, and in particular to a method for accelerating the characterization extraction of devices in a standard cell library. Background Art
[0002] At present, the main process of characterization extraction of standard cells of integrated circuits is to obtain the delay and power consumption information of different timing arcs of standard cells under different states (such as temperature, voltage, when conditions, input swing, output load, etc.) based on SPICE simulation. However, the full use of SPICE simulation requires a large amount of computing resources and consumes a high cost of time to complete timing convergence and the establishment of standard cell libraries.
[0003] With the rapid iteration of integrated circuit technology, some device effects and process effects cannot be ignored under advanced processes and may cause great impacts. These effects require the introduction of more states to describe the timing arcs in the process of static timing analysis of standard cells. When performing static timing analysis under advanced processes, if only the results under the most pessimistic conditions are considered, the library building results will deviate too much from the actual application. This excessive pessimism will increase the design margin, reduce the robustness of the design, extend the design cycle, and increase material costs, resulting in reduced product competitiveness in the market. Therefore, IC designs with high process precision need to enumerate all possible states when performing static timing analysis. This enumeration causes the time index required for characterization extraction based on SPICE simulation to increase dramatically.
[0004] Therefore, a method that can accelerate feature extraction is needed as a technical problem to be solved. Summary of the invention
[0005] In order to solve the technical problem of excessive resource consumption of characterization extraction in the prior art, the present invention proposes a method for characterization extraction of circuit devices and a storage medium.
[0006] The method for characterizing and extracting a circuit device proposed by the present invention comprises:
[0007] Step 1, reading the file required for characterization extraction of each timing arc, and generating a SPICE netlist required for characterization extraction according to the file;
[0008] Step 2, performing reproducible pseudo-random labeling on the SPICE netlist, simulating the SPICE netlist labeled as a simulation netlist, and forming a training set with the simulation result labels and the corresponding input parameter vectors; and forming a prediction set with the input parameter vectors corresponding to the unsimulated netlist;
[0009] Step 3, training multiple groups of gradient boosting-based models for each time series arc based on the training set, selecting optimal super parameters by cross-validation according to the super parameter combination, and training the final model according to the optimal super parameters;
[0010] Step 4, pass the input parameter vector in the prediction set to the final model, find the corresponding unsimulated netlist through pseudo-random reproduction of the timing or power consumption prediction vector output by the final model, and write it to the corresponding position of the standard cell library file in the specified format.
[0011] Further, the step 3 comprises:
[0012] Train the gradient boosting based model using the training set and multiple sets of hyper parameters;
[0013] Cross-validation is used to select the optimal hyperparameter combination;
[0014] The gradient boosting-based model is trained according to the optimal hyperparameter combination to obtain the final model.
[0015] Further, training the gradient boosting based model includes:
[0016] Initialize the first base model M0, x is the input parameter vector, N is the number of training samples in the training set, c is a constant for the first training, and c is the output result of the previous model for the second and subsequent trainings, y i represents the label of the i-th sample, and L is the loss function;
[0017] Calculate the i-th residual λ of the m-th iteration mi , Its value is the negative gradient, where M is the total model, x i Represents the input parameter vector of the i-th sample;
[0018] When the residual is greater than or equal to the threshold, a new model R(x) is trained based on the residual and the model M is updated. m (x) = M m-1 (x)+R(x), when the residual is less than the threshold, the final model in step 3 is
[0019]
[0020] Furthermore, the loss function is a first-order derivative or a second-order derivative of a mean absolute error, a mean absolute percentage error, or a root mean square error.
[0021] Furthermore, the reproducible pseudo-random marking is to uniformly select a shift within a parameter range of different input parameter vectors with a certain step length for marking, or to perform marking based on the input parameter vector with a shift of a reproducible pseudo-random step length under a fixed random seed.
[0022] Furthermore, the cross-validation selects K-fold or leave-P cross-validation to obtain the optimal super parameter combination.
[0023] The computer-readable storage medium proposed in the present invention is used to run a computer program, and when the computer program is run, the circuit device characterization extraction method of the above technical solution is executed.
[0024] The present invention proposes a machine learning framework based on gradient boosting, and the machine learning method based on gradient boosting is defined in the function domain (I don't quite understand what is defined in the function domain, and the material details scheme given at present cannot correspond to this point. If there is any missing, please add the technical solution details). Unlike most machine learning methods defined in the parameter domain, the method defined in the function domain corrects the errors caused by all old models by training a new weak model each iteration, and has the advantages of strong resistance to overfitting, high precision, and stable performance, and can effectively cope with the challenges brought by the regression task of discrete small sample data sets. In addition, the present invention can model each time series arc through a gradient boosting algorithm reconstructed based on C++, and the total training prediction speed is negligible compared to SPICE simulation. Moreover, the present invention is based on a pseudo-random method with reproducible consistency (such as uniform selection, pseudo-random step size) to perform data set partitioning and post-processing write-back, so that the independence of each step of the characterization extraction acceleration process is enhanced, the required memory and the dependence between steps are reduced, and it is easy to embed the original characterization extraction product without affecting the data structure of the original product code. In summary, compared with the prior art, the present invention has the advantages of being more efficient, accurate, and stable. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] The present invention is described in detail below with reference to the embodiments and accompanying drawings, wherein:
[0026] Figure 1 It is an overall flow chart of an embodiment of the present invention.
[0027] Figure 2 This is a training set extraction flow chart of an embodiment of the present invention.
[0028] Figure 3 It is a model training flow chart of an embodiment of the present invention. DETAILED DESCRIPTION
[0029] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.
[0030] Thus, a feature indicated in this specification will be used to illustrate one of the features of an embodiment of the present invention, rather than implying that each embodiment of the present invention must have the described feature. In addition, it should be noted that this specification describes many features. Although some features can be combined together to illustrate possible system designs, these features can also be used in other combinations that are not explicitly described. Thus, unless otherwise stated, the described combinations are not intended to be limiting.
[0031] like Figure 1 As shown, the method for characterizing and extracting circuit devices of the present invention includes the following steps.
[0032] Step 1: read the files required for characterization extraction of each timing arc, and generate the SPICE netlist required for characterization extraction based on the files.
[0033] In general characterization tasks, it is necessary to traverse and simulate multiple variables of the timing arc, such as input excitation, load capacitance, working environment, etc., in order to extract the necessary characteristic information. Therefore, after the characterization extraction initialization of each timing arc is completed, a SPICE simulation netlist will be generated in batches based on the input parameters (i.e. input excitation, load capacitance, working environment, etc.).
[0034] In the initialization phase, the working voltage is inferred and analyzed through the functional expression of the circuit unit and the initialization setting of the circuit is performed. The input signal of the circuit unit is converted and constructed through a specific formula, and then the SPICE simulation netlist is generated in batches based on the input parameters. The specific formula is provided by the characterization software.
[0035] Step 2, perform reproducible pseudo-random marking on the SPICE netlist, simulate the SPICE netlist marked as a simulation netlist, and form a training set with the simulation result labels and the corresponding input parameter vectors; and form a prediction set with the input parameter vectors corresponding to the unsimulated netlist.
[0036] The SPICE netlist marked as the simulation netlist performs SPICE simulation, obtains timing and power consumption information through the traditional characterization extraction process, that is, outputs the result vector, and combines the output result vector with the input parameter vector corresponding to the SPICE simulation netlist in a pseudo-random manner to form a training set.
[0037] Step 3: Based on the training set, multiple groups of gradient boosting-based models are trained for each time series arc, and the optimal super parameters are selected by cross validation according to the super parameter combination, and the final model is trained according to the optimal super parameters.
[0038] Step 4, pass the input parameter vector in the prediction set to the final model, find the corresponding unsimulated netlist through pseudo-random reproduction of the timing or power consumption prediction vector output by the final model, and write it to the corresponding position of the standard cell library file in the specified format.
[0039] Through the above technical solution, the gradient boosting algorithm is used to model each timing arc, and the total training prediction speed is negligible compared to the SPICE simulation, which can effectively improve the characterization extraction speed of the basic unit library. In addition, the present invention has the advantages of strong resistance to overfitting, high precision, and stable performance, and can effectively cope with the challenges brought by the regression task of discrete small sample data sets.
[0040] like Figure 2 As shown, in a specific embodiment, the training set of the present invention is extracted using the following detailed steps.
[0041] Read in various files required for characterization extraction, including netlist files, cell library files, and device parameter files, etc. Netlist files are files that describe circuit devices and structures; cell library files include target information to be extracted, such as standard cell power consumption, timing, and constraints, as well as driver models and receiver models; device parameter files update device models based on input parameters and corresponding device effects.
[0042] Generate the SPICE netlist required for characterization extraction through the above files.
[0043] Mark simulated and unsimulated netlists in a reproducible pseudo-random manner.
[0044] Perform SPICE simulation on the simulation netlist and collect the corresponding simulation results.
[0045] Combine the simulation result labels with the corresponding input parameter vectors to complete the extraction of the training set.
[0046] like Figure 3 As shown, in a specific embodiment, step 3 of the present invention includes the following steps.
[0047] Train the gradient boosting based model using the training set and multiple sets of hyper parameters;
[0048] Cross-validation is used to select the optimal hyperparameter combination;
[0049] The gradient boosting-based model is trained according to the optimal hyperparameter combination to obtain the final model.
[0050] When training a gradient boosting model using a training set and multiple sets of hyperparameters, only a small number of samples in the training set are needed. The purpose is not to train the model, but to get the optimal hyperparameter combination. Hyperparameters include but are not limited to the number of sub-models, the model learning rate used for training, the maximum depth of the model, etc.
[0051] After obtaining the optimal hyperparameter combination, the training set and the optimal hyperparameter combination are used to train the gradient boosting model. Although the training process is the same for the gradient boosting model, the number of samples in the training set used in the first selection of the optimal hyperparameter combination is much smaller than the number of samples used in the final model training.
[0052] In a specific embodiment, training a gradient boosting based model includes the following steps.
[0053] Initialize the first base model M0, x is the input parameter vector, y i Represents the label of the ith sample, that is, the simulated value corresponding to the ith sample. x is the ith sample, which can also be expressed as x i , that is, x is the input parameter vector corresponding to yi. The input parameter vector refers to the vector corresponding to a set of input parameter combinations. N is the number of training samples in the training set. c is a constant during the first training. During the second and subsequent trainings, c is the simulation result of the previous model, i.e., power consumption or delay. y i represents the label of the i-th sample. The label refers to the value of the simulation result corresponding to the sample. L is the loss function.
[0054] Calculate the i-th residual λ of the m-th iteration mi , Its value is the negative gradient, where M is the total model, x i Represents the input parameter vector of the i-th sample;
[0055] When the residual is greater than or equal to the threshold, a new model R(x) is trained based on the residual and the model M is updated. m (x) = M m-1 (x)+R(x), when the residual is less than the threshold, the final model in step 3 is
[0056]
[0057] In the above technical solution, the loss function L is the first-order derivative or the second-order derivative of the mean absolute error, the mean absolute percentage error, and the root mean square error. Of course, the loss function L can also be any other loss function that can be used for regression tasks.
[0058] In the above technical solution, a training set previously obtained through a small amount of SPICE simulation is used to train multiple groups of gradient-based models for each timing arc, that is, multiple groups of machine learning models using different super parameters. The optimal parameters are selected by cross-validation based on the empirically obtained super parameter combination. The cross-validation method can select K-fold or leave-P cross-validation. The cross-validation evaluation index can be the mean absolute error, root mean square error, mean absolute percentage error and other evaluation indicators that meet the requirements of the application scenario, or the results obtained by scoring multiple evaluation indicators with different weights, and retrain a set of final models based on the optimal hyperparameters. The super parameter combination comes from actual work experience. For circuit designs under different processes, there will be different selections of appropriate super parameter combinations. In order to fit the application scenario to the greatest extent, multiple groups of super parameter combinations are pre-set in the model, and the optimal combination is selected through verification.
[0059] The reproducible pseudo-random marking of the present invention is to uniformly select a shift within the parameter range of different input parameter vectors with a certain step length for marking, or to reproducibly shift the pseudo-random step length under a fixed random seed to mark based on the input parameter vector.
[0060] The main purpose of the present invention is to reduce the number of simulations in the characterization process, and the purpose of using pseudo-randomness is to distinguish between the training set that needs to be simulated and the prediction set that is not needed. In order to obtain accurate and reproducible results, a pseudo-random method is used to select the training set.
[0061] The timing arcs of different states of the present invention follow the naming rules of standard cells, that is, the cell names in a cell library cannot be repeated, the states corresponding to the timing arcs of the same path cannot be repeated, and the suffixes of the same cell names in different states are also different, and the suffixes are determined by the corresponding input parameter vectors. In addition to the above provisions, the format of the data written back to the standard cell library is the unique format of the characterized output results, and there is a unified standard in the industry. In order to ensure that the output results of the present invention can be directly embedded in the existing chip design process, the output results will be written back to the original unique format.
[0062] The type and number of input parameter vectors are determined by the application scenario and scope. For example, the traditional NLDM model includes temperature, voltage, etc., while the NLDM that takes aging into consideration needs to introduce two additional parameters: aging time and input signal duty cycle. The more types and numbers of input parameter vectors, the wider the application scenarios of the trained model, and the longer it takes to establish the training set through SPICE simulation. In addition, the more complex the dimensions and number of input parameter vectors, the wider the application scope of the trained model, and the more universal the model, so that it can be reused instead of just for a single acceleration.
[0063] The present invention also protects a computer-readable storage medium, which is used to run a computer program, and when the computer program is run, the circuit device characterization extraction method of the above technical solution is executed.
[0064] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A method for characterizing and extracting a circuit device, characterized in that: include: Step 1, reading the file required for characterization extraction of each timing arc, and generating a SPICE netlist required for characterization extraction according to the file; Step 2, performing reproducible pseudo-random labeling on the SPICE netlist, simulating the SPICE netlist labeled as a simulation netlist, and forming a training set with the simulation result labels and the corresponding input parameter vectors; and forming a prediction set with the input parameter vectors corresponding to the unsimulated netlist; Step 3, training multiple groups of gradient boosting-based models for each time series arc based on the training set, selecting optimal super parameters by cross-validation according to the super parameter combination, and training the final model according to the optimal super parameters; Step 4, pass the input parameter vector in the prediction set to the final model, find the corresponding unsimulated netlist through pseudo-random reproduction of the timing or power consumption prediction vector output by the final model, and write it to the corresponding position of the standard cell library file in the specified format.
2. The method for characterizing and extracting a circuit device according to claim 1, wherein: The step 3 comprises: Train the gradient boosting based model using the training set and multiple sets of hyper parameters; Cross-validation is used to select the optimal hyperparameter combination; The gradient boosting-based model is trained according to the optimal hyperparameter combination to obtain the final model.
3. The method for characterizing and extracting a circuit device according to claim 2, wherein: Training a gradient boosting based model involves: Initialize the first base model M0, x is the input parameter vector, N is the number of training samples in the training set, c is a constant for the first training, and c is the output result of the previous model for the second and subsequent trainings, y i represents the label of the i-th sample, and L is the loss function; Calculate the i-th residual λ of the m-th iteration mi , Its value is the negative gradient, where M is the total model, x i Represents the input parameter vector of the i-th sample, m ≥ 2; When the residual is greater than or equal to the threshold, a new model R(x) is trained based on the residual and the model M is updated. m (x) = M m-1 (x)+R(x), when the residual is less than the threshold, the final model in step 3 is 4. The method for characterizing and extracting a circuit device according to claim 3, wherein: The loss function is a first-order derivative or a second-order derivative of one of the mean absolute error, the mean absolute percentage error, and the root mean square error.
5. The method for characterizing and extracting a circuit device according to claim 1, wherein: The reproducible pseudo-random marking is to uniformly select a shift within the parameter range of different input parameter vectors with a certain step length for marking, or to mark based on the input parameter vector with a shift of a reproducible pseudo-random step length under a fixed random seed.
6. The method for characterizing and extracting a circuit device according to claim 2, wherein: The cross-validation selects K-fold or leave-P cross-validation to obtain the optimal super parameter combination.
7. A computer-readable storage medium for running a computer program, characterized in that: When the computer program is executed, the circuit device characterization extraction method according to any one of claims 1 to 6 is executed.