WO2023226423A1 - 一种芯片辅助设计方法、装置、设备及非易失性存储介质 - Google Patents

一种芯片辅助设计方法、装置、设备及非易失性存储介质 Download PDF

Info

Publication number
WO2023226423A1
WO2023226423A1 PCT/CN2022/141691 CN2022141691W WO2023226423A1 WO 2023226423 A1 WO2023226423 A1 WO 2023226423A1 CN 2022141691 W CN2022141691 W CN 2022141691W WO 2023226423 A1 WO2023226423 A1 WO 2023226423A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
chip
features
prediction model
original
Prior art date
Application number
PCT/CN2022/141691
Other languages
English (en)
French (fr)
Inventor
邹德强
Original Assignee
苏州元脑智能科技有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 苏州元脑智能科技有限公司 filed Critical 苏州元脑智能科技有限公司
Publication of WO2023226423A1 publication Critical patent/WO2023226423A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/30Circuit design
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00Machine learning
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/30Circuit design
    • G06F30/32Circuit design at the digital level
    • G06F30/33Design verification, e.g. functional simulation or model checking
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/30Circuit design
    • G06F30/32Circuit design at the digital level
    • G06F30/337Design optimisation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F30/00Computer-aided design [CAD]
    • G06F30/30Circuit design
    • G06F30/39Circuit design at the physical level
    • G06F30/398Design verification or optimisation, e.g. using design rule check [DRC], layout versus schematics [LVS] or finite element methods [FEM]

Definitions

  • the present application relates to the field of chip design technology, and in particular to a chip auxiliary design method, device, equipment and non-volatile storage medium.
  • Chip generally refers to the carrier of integrated circuits. Its design is divided into front-end design (logical design) and back-end design (physical design). Front-end design involves functional design, and back-end design involves process-related design, including RTL (Register Transfer Level, Register transfer level) writing, functional verification, logic synthesis, formal verification, DFT (Design for Testability, design for testability), placement and routing, Sign Off (signing), layout verification and other processes to make it a chip with manufacturing significance .
  • RTL Register Transfer Level, Register transfer level
  • DFT Design for Testability, design for testability
  • Sign Off Sign Off
  • EDA Electronic design automation
  • the purpose of the embodiments of the present application is to propose a chip auxiliary design method, device, equipment and non-volatile storage medium, which can enhance the understanding of chip data characteristics and improve the understanding of chip data characteristics by designing a compound weighted structure through feature engineering.
  • the effective utilization of chip data can output more reliable results.
  • the chip-assisted design ideas based on machine learning are more flexible, have fewer constraints, and have lower costs. From a technical perspective, they can be applied to changing application scenarios.
  • one aspect of the embodiment of the present application provides a chip auxiliary design method, which includes the following steps: obtaining original chip data, and preprocessing the above original chip data to obtain preprocessed chip data; based on chip experience Perform exploratory analysis on the above-mentioned preprocessed chip data to obtain the first characterized data, perform filtering analysis on the above-mentioned preprocessed chip data based on statistical tests to obtain the second characterized data, and perform the above-mentioned preprocessing based on the model direction.
  • the subsequent chip data is analyzed by the embedding method to obtain third characterized data; the above-mentioned first characterized data, the above-mentioned second characterized data and the above-mentioned third characterized data are weighted and analyzed to obtain a characterized data set; Select different algorithms to establish different prediction models based on the above-mentioned characterized data sets, and train and evaluate the above-mentioned prediction models to obtain the optimal prediction model; and analyze the data characteristics of the chip data based on the above-mentioned optimal prediction models to obtain the data Feature importance ranking, chip-assisted design based on the above importance ranking and through the above optimal prediction model.
  • obtaining the original chip data includes: obtaining the original chip data through an EDA tool; or obtaining the original chip data through test data; or obtaining the original chip data through server data.
  • preprocessing the above-mentioned original chip data to obtain preprocessed chip data includes: performing missing value processing on the above-mentioned original chip data; and/or performing repeated value processing on the above-mentioned original chip data; and/or Perform outlier processing on the above original chip data; and/or perform character data encoding processing on the above original chip data; and/or perform dimensionless processing on the above original chip data.
  • performing exploratory analysis on the above-mentioned preprocessed chip data based on chip experience to obtain the first characterized data includes: analyzing the trend changes of discrete chip features through histograms; analyzing the trend changes of continuous chip features through kernel density curves. Trend changes; analyze the trend of chip features compared to labels by calculating crosstabs; analyze the trend of matching features compared to labels through grouped box plots; analyze features and features, features and tags through correlation coefficients and heat maps Correlation; create features by analyzing principal component analysis.
  • performing filtering analysis on the above-mentioned preprocessed chip data based on statistical tests to obtain the second characterized data includes: filtering the above-mentioned preprocessed chip data based on variance filtering and/or correlation filtering to obtain the second characterized data. Obtain second characterization data.
  • weighting and analyzing the first characterized data, the second characterized data and the third characterized data to obtain a characterized data set includes: using statistical analysis software to analyze the weighted data.
  • the above-mentioned first characterization data, the above-mentioned second characterization data and the above-mentioned third characterization data are analyzed.
  • selecting different algorithms to establish different prediction models based on the above-mentioned characterized data sets includes: establishing a chip quality prediction model based on the above-mentioned characterized data sets; and/or establishing a chip performance scoring model based on the above-mentioned characterized data sets.
  • training and evaluating the above-mentioned prediction model to obtain the optimal prediction model includes: training the above-mentioned prediction model through at least one of K-fold cross-validation, learning curve, and grid search.
  • training and evaluating the above-mentioned prediction model to obtain the optimal prediction model includes: predicting the above-mentioned chip quality through at least one of confusion matrix, accuracy, precision, recall, F1 score, and ROC curve. Model is evaluated.
  • training and evaluating the above prediction model to obtain the optimal prediction model includes: evaluating the above chip performance scoring model through mean square error.
  • the method includes the following steps: feedback the above importance ranking as feedback information to adjust the chip design.
  • the method further includes: obtaining chip data of the above-mentioned newly manufactured chip, and analyzing the chip data of the above-mentioned newly manufactured chip based on the above-mentioned optimal prediction model.
  • the raw chip data obtained above includes at least one of the following: structured data of database, semi-structured data of file logs, and unstructured data of audio and video data.
  • the above-mentioned processing of missing values on the above-mentioned original chip data includes one of the following: deleting the missing values in the above-mentioned original chip data; using 0 to fill the missing values in the above-mentioned original chip data; using the mean value of the same feature to fill in Missing values in the above-mentioned original chip data; use other feature indicators to predict the missing values in the above-mentioned original chip data.
  • performing duplicate value processing on the original chip data includes: deleting duplicate values in the original chip data.
  • the above-mentioned outlier processing on the above-mentioned original chip data includes: deleting the outliers in the above-mentioned original chip data.
  • the above-mentioned embedding analysis of the above-mentioned preprocessed chip data based on the model direction to obtain the third characterized data includes: importing all chip features into the model, and after training and tuning, using the model attributes to evaluate Based on the importance of the above-mentioned chip features, all chip features are screened through a preset threshold to obtain a feature subset, wherein the third characterization data includes the feature subset, and the preprocessed Chip data are all chip characteristics described.
  • a chip auxiliary design device including: a first module configured to obtain original chip data, and preprocess the original chip data to obtain preprocessed chip data. ;
  • the second module is configured to perform exploratory analysis on the above-mentioned preprocessed chip data based on chip experience to obtain the first characterized data, and perform filtering analysis on the above-mentioned preprocessed chip data based on statistical testing to obtain the second
  • the characterization data performs an embedding analysis on the above-mentioned preprocessed chip data based on the model direction to obtain the third characterization data;
  • the third module is configured to analyze the above-mentioned first characterization data, the above-mentioned second characterization data and the above-mentioned characterization data.
  • the third characterized data is weighted and analyzed to obtain a characterized data set; the fourth module is configured to select different algorithms to establish different prediction models based on the above characterized data sets, and to train and evaluate the above prediction models to obtain the optimal prediction model; and the fifth module is configured to analyze the data characteristics of the chip data based on the above-mentioned optimal prediction model to obtain the importance ranking of the data features, based on the above-mentioned importance ranking and through the above-mentioned optimal prediction model Conduct chip-assisted design.
  • a computer device including: at least one processor; and a memory.
  • the memory stores computer instructions that can be run on the processor. When the instructions are executed by the processor, the above method is implemented. step.
  • a non-volatile storage medium stores a computer program that implements the above method steps when executed by a processor.
  • the embodiments of the present application at least have the following beneficial technical effects: designing a compound weighted structure through feature engineering can enhance the understanding of chip data characteristics, improve the effective utilization of chip data, and output more reliable results.
  • chips based on machine learning The auxiliary design idea has more flexibility, fewer constraints, and lower cost. From a technical perspective, it can be applied to changing application scenarios.
  • Figure 1 is a schematic diagram of an embodiment of a chip-assisted design method provided by an embodiment of the present application
  • Figure 2 is a schematic diagram of an embodiment of a chip auxiliary design device provided by an embodiment of the present application
  • Figure 3 is a schematic diagram of an embodiment of a computer device provided by an embodiment of the present application.
  • FIG. 4 is a schematic diagram of an embodiment of a non-volatile storage medium provided by an embodiment of the present application.
  • the first aspect of the embodiments of this application provides an embodiment of a chip-assisted design method.
  • Figure 1 shows a schematic diagram of an embodiment of the chip-assisted design method provided by this application.
  • the chip-assisted design method according to the embodiment of the present application includes the following steps:
  • the chip in feature engineering, can be solved through compound weighted structure, exploratory analysis, filtering method, embedding method, and professional statistical analysis by SPSS (Statistical Product Service Solutions).
  • SPSS Statistical Product Service Solutions
  • obtaining the original chip data includes: obtaining the original chip data through an EDA tool; or obtaining the original chip data through test data; or obtaining the original chip data through server data.
  • the chip data exists completely in the database and is tagged; the chip data can come from EDA tools, such as simulation data, variability measurements, or test data from the chip, or from server data, including But it is not limited to current, voltage, temperature, frequency, data flow, process library parameters, IP (Internet Protocol) selection, etc.
  • EDA tools such as simulation data, variability measurements, or test data from the chip, or from server data, including But it is not limited to current, voltage, temperature, frequency, data flow, process library parameters, IP (Internet Protocol) selection, etc.
  • the data collected includes: structured data of databases, semi-structured data of file logs, and unstructured data of audio and video data.
  • File logs are collected through Flume (a distributed system for collecting, aggregating and transmitting massive logs), and audio and video are collected through Kafka (an open source stream processing platform) message queue.
  • Flume a distributed system for collecting, aggregating and transmitting massive logs
  • Kafka an open source stream processing platform
  • preprocessing the original chip data to obtain preprocessed chip data includes: performing missing value processing on the original chip data; and/or performing repeated value processing on the original chip data; and/or Perform outlier processing on the original chip data; and/or perform character data encoding processing on the original chip data; and/or perform dimensionless processing on the original chip data.
  • the original complex chip data is cleaned so that it can be standardized and described using mathematical language.
  • the preprocessing module involves processing solutions such as missing values, repeated values, outliers, character data encoding, and non-dimensionalization.
  • processing of missing values includes: direct deletion; 0 filling, using 0 to fill missing values; mean filling, using the mean value under the feature to fill missing values; predictive filling, using other feature indicators to predict missing values.
  • the missing value in various embodiments of the present application is used to indicate that a missing value occurs in the value corresponding to the chip feature.
  • processing of duplicate values includes: direct deletion and deduplication.
  • processing of outliers includes: established rules, in accordance with physical laws, deletion of super abnormal items (items); 3 ⁇ principle, in accordance with statistical experience, if the chip data obeys the normal distribution, the numerical distribution is in ( ⁇ - The probability in 3 ⁇ , ⁇ +3 ⁇ ) is 0.9974, that is, P(
  • super abnormal items are items (items) that exceed the abnormal threshold, that is, exceed ( ⁇ -3 ⁇ , ⁇ +3 ⁇ ).
  • “super abnormality” and “abnormality” can be understood as having the same meaning
  • “deleting super abnormal items (items)” can be understood as “deleting abnormal values”.
  • the character data encoding processing of the original chip data includes: direct encoding, suitable for categorical variables; one-hot encoding, suitable for nominal variables; binning encoding, which can be based on distance binning and quantile as needed Binning, entropy value binning, etc.; word embedding (word embedding), using neural network coding.
  • dimensionless processing of original chip data includes: normalization, the formula is as follows:
  • chip data After chip data is preprocessed, it can be understood and transformed into high-quality data, which will help improve the accuracy of the results.
  • performing exploratory analysis on preprocessed chip data based on chip experience to obtain the first characterization data includes: analyzing trend changes of discrete chip features through histograms; analyzing continuous chips through kernel density curves Trend changes in features; analyze the trend of chip features compared to labels by calculating crosstabs; analyze the trend of matching features compared to labels through grouped box plots; analyze features and features, features and tags through correlation coefficients and heat maps Correlation between them; create features by analyzing principal component analysis.
  • Kernel density curve to analyze trend changes in continuous chip features
  • Correlation coefficient and heat map analyze the correlation between features and features, features and labels
  • the chip design engineer selects the main chip data features and obtains the characterized data T1.
  • performing filtering analysis on the preprocessed chip data based on statistical tests to obtain the second characterized data includes: filtering the preprocessed chip data based on variance filtering and/or correlation filtering. to obtain the second characterization data.
  • weighting and analyzing the above-mentioned first characterization data, the above-mentioned second characterization data and the above-mentioned third characterization data to obtain a characterized data set includes: using statistical analysis software to weight the weighted data.
  • the processed first characterized data, the above second characterized data and the above third characterized data are analyzed.
  • the filtering method mainly considers variance, correlation, etc. to select features, and a threshold can be set to control the margin of chip features.
  • Correlation pass rate filters out chip features that have low correlation with labels. It is generally believed that such features have little impact on labels, including: Chi-square test, which considers the independence between chip features and labels; F test, which considers the chip The linear relationship between features and labels; the mutual information method considers the linear relationship and non-linear relationship between digital chip features and labels;
  • embedding analysis is performed on the preprocessed chip data based on the model direction to obtain third characterization data.
  • a feature selection method commonly used in the industry can effectively improve the model score.
  • the embedding method has restrictions on the model. Generally speaking, the model should have one of coef_attributes, feature_importances_attributes, L1 or L2 penalty terms, etc.
  • the embedding method imports all chip features into the model (such as random forest). After training and tuning, the model attribute feature_importances_ is used to evaluate the importance of the features, set the threshold, and obtain the data feature subset (or feature subset). ), that is, obtain the characterized data T3.
  • the embedding method generally has model requirements, that is, there are feature_importances_attributes or penalty terms, etc. After modeling, the features can be quantitatively understood, thresholds can be set, and important features can be retained based on the thresholds to obtain feature subsets.
  • the above "importing all chip features into the model” can be, but is not limited to, understood as: importing preprocessed chip data into the model, and then using the model to filter features and construct a feature subset.
  • the above-mentioned preprocessed chip data can be all the above-mentioned chip characteristics.
  • a random forest model is an example, and other models can also be used, such as a logistic regression model and so on. It can be understood that for the models used above (for example, random forest models or logistic regression models), thresholds for filtering features can be preset.
  • the data features obtained from three angles are counted according to the number of occurrences, and weights of 1, 2, and 3 are given, and the weighted data are imported into statistical analysis software for analysis.
  • the weighted data is imported into SPSS (if the memory is not enough, you can sample) professional analysis software to perform systematic statistical analysis, such as homogeneity of variance test, correlation analysis, principal component analysis, factor analysis, Regression analysis, clustering, variance expansion factor method (removing multicollinearity), etc., are used to obtain the final characterized data set T.
  • SPSS analysis is to obtain the final characterized data set T about the chip data.
  • the final characterized data set T is obtained by weighting the characterized data T1, T2, and T3 of the chip and analyzed, and it belongs to the reprocessing of data. Utilize, for example, the cases described below in correlation analysis and principal component analysis.
  • the original high-dimensional space data is mainly projected into a low-dimensional data space.
  • the data feature value range of weight 3 can be increased to 3 times, and the data feature value range of weight 2 can be increased by 2. times, during the dimensionality reduction operation, the projection will be skewed towards the characteristics of this type of data, increasing attention;
  • selecting different algorithms to establish different prediction models based on the characterized data set includes: establishing a chip quality prediction model based on the characterized data set; and/or establishing a chip performance scoring model based on the characterized data set.
  • Training and evaluating the above prediction model to obtain the optimal prediction model includes: training the above prediction model through at least one of K-times cross-validation, learning curve, and grid search.
  • algorithm modeling is a process of model selection, training and optimization. Establish a chip quality prediction model 1 based on the characterized data T, and establish a chip performance prediction model 2 based on the characterized data T;
  • a two-classification model Classification algorithms such as logistic regression, random forest, support vector machine, and naive Bayes are considered;
  • the chip performance scoring model 2 is a regression problem, that is, giving a prediction score for the performance of a chip, considering a linear regression model, etc.
  • Chip quality prediction model 1 or chip performance prediction model 2 both require a training process. You can choose K-time cross-validation, learning curve, grid search and other techniques to improve the score of the model, including:
  • K times of cross-validation divide the data set into K parts, use K-1 parts as the training set, and the remaining 1 part as the test set, repeat K times, and observe the training results;
  • Grid search is a method for cross-validating and optimizing models by exhaustively enumerating parameters and combining parameters.
  • the chip quality prediction model can be evaluated through at least one of confusion matrix, accuracy, precision, recall, F1 score, and ROC curve. , or evaluate the chip performance scoring model through mean square error. After the evaluation is completed, the model will be analyzed, details of which will be described in subsequent optional embodiments.
  • the method may include the following steps: feedback the above importance ranking as feedback information to adjust the chip design.
  • the importance ranking can be fed back to the chip design engineer.
  • the chip design engineer can adjust the chip according to different situations.
  • a chip design engineer can design a new chip based on the above feedback information, obtain new chip data, and analyze the chip data of the new chip based on the optimal prediction model. For example, when a chip design engineer uses feedback information to design a new chip, the data of the new chip can be imported into the model for data simulation verification.
  • embodiments of the present application provide an open chip auxiliary design method based on machine learning.
  • designing a compound weighted structure through feature engineering can enhance the understanding of chip data characteristics and improve the accuracy of chip data. Effective utilization, output more reliable results.
  • the open chip-assisted design idea based on machine learning has better flexibility, fewer constraints, lower costs, and can be applied to changing application scenarios from a technical perspective. Details are as follows:
  • An open chip-assisted design method based on machine learning including chip design data, data preprocessing module, feature engineering module, algorithm modeling module, evaluation module, analysis, etc.
  • microarray data is assumed to be fully present in the database and tagged.
  • the data comes from EDA tools, such as simulation data, variability measurement, or test data from chips, or from server data, etc.;
  • the data preprocessing module cleans the original complex chip data so that it can be standardized and described using mathematical language.
  • the preprocessing module involves processing solutions for missing values, duplicate values, outliers, character data encoding, dimensionless, etc.;
  • the feature engineering module designs compound weighted structure innovation, extracts and creates valuable chip data, and provides characteristic data for subsequent algorithm modeling.
  • Data feature engineering is carried out from three perspectives: exploratory analysis in the direction of chip experience to obtain characterized data T1, filtering analysis in the direction of statistical testing to obtain characterized data T2, and embedding analysis in the direction of the model to obtain characterized data.
  • T3 count the data features obtained from three angles according to the number of occurrences, and assign weights of 1, 2, and 3. Import the weighted data into SPSS (if the memory is not enough, you can sample) professional analysis software to do systematic statistical analysis.
  • Chip quality prediction model 1 is a classification problem. It is assumed that only one chip is abnormal or not, that is, a two-class classification model. It considers classification algorithms such as logistic regression, random forest, support vector machine, and naive Bayes; chip performance prediction model 2, It is a regression problem, that is, giving a prediction score for the performance of a chip, considering a linear regression model, etc. Regardless of the chip quality prediction model 1 or the chip performance prediction model 2, a training process is required. You can choose K-time cross-validation, learning curve, grid search and other techniques to improve the score of the model;
  • the evaluation module evaluates the prediction effect of chip data under different algorithms and selects the optimal algorithm based on the evaluation results.
  • the evaluation indicators depend on the type of model. For the chip quality prediction model 1, confusion matrix (confusion_matrix), accuracy, precision, recall, f1score, roc curve (Receiver Operating Characteristic Curve, Receiver Operating Characteristic Curve), etc. can be used. Measure the effect of the model; for the chip performance prediction model 2, the mean square error MSE can be used to measure the effect of the model; including the following classification evaluation indicators:
  • Confusion matrix (confusion_matrix) is a multi-dimensional measurement index system for binary classification problems. 1 represents abnormal chips and 0 represents normal chips. In the confusion matrix, the true value comes first and the predicted value comes last, as shown in the following table:
  • Recall rate the proportion of all predicted anomaly categories in the total anomaly categories, capture anomaly categories, the formula is as follows:
  • the roc curve is a curve with the false positive rate as the abscissa and the recall rate as the ordinate.
  • the accuracy of the model is selected as the evaluation index; when facing imbalanced chip data, f1score (f1 score) will be considered.
  • MSE Mean Square Error
  • Regression evaluation indicators can also use root mean square error (RMSE, Root Mean Square Error), mean absolute error (MAE, Mean Absolute Error), coefficient of determination (Coefficient of Determination), etc.
  • RMSE Root Mean Square Error
  • MAE Mean Absolute Error
  • Coefficient of Determination Coefficient of Determination
  • the coefficient can be used to measure the importance of chip characteristics. That is, the greater the absolute value of the coefficient, the deeper the chip characteristics affect the quality factors of the chip. If the coefficient is positive, it means that the chip Characteristics have a positive effect. If it is negative, it means that the chip feature has a negative effect. If you use a random forest model, you can use the attribute feature_importances_ to measure the importance of chip features. The larger the value, the greater the influencing factors. Through induction and summary, you can learn the features that affect chip quality; for the chip performance prediction model 2, the higher the MSE. Small means the model is more accurate, and the score describes how confidently the chip's performance predictions are.
  • chip design engineer After we analyze the potential physical laws of the chip data, we can feedback information to the chip design engineer. If it is a quality classification model, we can gradually adjust the chip design details based on the summarized chip feature importance, such as by changing some circuit structures, process parameters, wiring, etc. , adjust the data under this feature or part of the feature to convert it from an abnormal chip to a normal chip to improve efficiency; in the case of a performance prediction model, mathematical optimization methods can be used to calculate the input feature data of the chip to maximize its performance, that is, the maximum performance of the chip. Based on the optimal design, high-performance chips can be designed, that is, auxiliary chip design based on machine learning is completed.
  • FIG. 2 shows a schematic diagram of an embodiment of a chip-assisted design device provided by an embodiment of the present application.
  • the chip auxiliary design device of the embodiment of the present application includes the following modules: a first module 011, configured to obtain original chip data and preprocess the original chip data to obtain preprocessed chip data;
  • the second module 012 is configured to perform exploratory analysis on the preprocessed chip data based on chip experience to obtain the first characterized data, and perform filtering analysis on the preprocessed chip data based on statistical testing to obtain the second characterized data.
  • the third module 013 is configured to perform analysis on the first characterized data, the second characterized data and the third characterized data. Weight processing and analysis to obtain a characterized data set; the fourth module 014 is configured to select different algorithms to establish different prediction models based on the characterized data set, and train and evaluate the prediction model to obtain the optimal prediction model; And the fifth module 015 is configured to analyze the data characteristics of the chip data based on the optimal prediction model to obtain the importance ranking of the data features, and perform chip-assisted design based on the importance ranking and through the above-mentioned optimal prediction model.
  • FIG. 3 shows a schematic diagram of an embodiment of a computer device provided by an embodiment of the present application.
  • the computer equipment of the embodiment of the present application includes the following devices: at least one processor 021; and a memory 022.
  • the memory 022 stores computer instructions 023 that can be run on the processor. When the instructions are executed by the processor, the above is achieved. Method steps.
  • FIG. 4 shows a schematic diagram of an embodiment of a non-volatile storage medium provided by an embodiment of the present application.
  • the non-volatile storage medium 031 stores a computer program 032 that performs the above method when executed by the processor.
  • the program of the chip-assisted design method can be stored in a computer-readable In the storage medium, when the program is executed, it may include the processes of the above method embodiments.
  • the storage medium of the program can be a magnetic disk, an optical disk, a read-only memory (ROM (Read-Only Memory), a read-only memory) or a random access memory (RAM (Random Access Memory, a random access memory)), etc.
  • ROM Read-Only Memory
  • RAM Random Access Memory, a random access memory
  • the method disclosed according to the embodiment of the present application can also be implemented as a computer program executed by a processor, and the computer program can be stored in a non-volatile storage medium.
  • the computer program is executed by the processor, the above functions defined in the method disclosed in the embodiment of the present application are performed.
  • the above-mentioned method steps and system units can also be implemented using a controller and a non-volatile storage medium used to store a computer program that enables the controller to implement the above-mentioned steps or unit functions.
  • Nonvolatile storage media includes computer storage media and communication media including any medium that facilitates transfer of a computer program from one place to another. Storage media can be any available media that can be accessed by a general purpose or special purpose computer.
  • the non-volatile storage medium may include RAM, ROM, EEPROM (Electrically Erasable Programmable read-only memory), CD-ROM (Compact Disc Read-Only Memory) , CD-ROM) or other optical disk storage device, magnetic disk storage device or other magnetic storage device, or can be used to carry or store the required program code in the form of instructions or data structures and can be used by a general or special purpose computer or a general or special purpose computer. Any other media accessed by the processor. Additionally, any connection is properly termed a non-volatile storage medium.
  • disks and optical disks include compact disks (CDs), laser disks, optical disks, digital versatile disks (DVDs), floppy disks, and Blu-ray disks, where disks typically reproduce data magnetically, while optical disks reproduce data using lasers. Reproduce data optically. Combinations of the above should also be included within the scope of non-volatile storage media.
  • the program can be stored in a non-volatile storage medium.
  • the storage medium can be read-only memory, magnetic disk or optical disk, etc.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computer Hardware Design (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Geometry (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Data Mining & Analysis (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)
  • Complex Calculations (AREA)

Abstract

本申请实施例公开了一种芯片辅助设计方法,包括:获取原始芯片数据并进行预处理以得到预处理后的芯片数据;基于复式加权结构进行特征提取,即,基于芯片经验对预处理后的芯片数据进行探索性分析、基于统计检验对预处理后的芯片数据进行过滤法分析、基于模型方向对预处理后的芯片数据进行嵌入法分析,以分别得到第一、第二和第三特征化数据;对第一、第二和第三特征化数据进行加权处理并进行分析以得到特征化数据集;选择不同的算法基于特征化数据集建立不同的预测模型,并对预测模型进行训练和评估以得到最优预测模型;基于最优预测模型对芯片数据的数据特征进行分析以得到数据特征的重要性排序,基于重要性排序并且通过最优预测模型进行芯片辅助设计。

Description

一种芯片辅助设计方法、装置、设备及非易失性存储介质
相关申请的交叉引用
本申请要求于2022年5月24日提交中国专利局,申请号为202210569483.1,申请名称为“一种芯片辅助设计方法、装置、设备及可读介质”的中国专利申请的优先权,其全部内容通过引用结合在本申请中。
技术领域
本申请涉及芯片设计技术领域,尤其涉及一种芯片辅助设计方法、装置、设备及非易失性存储介质。
背景技术
芯片一般是指集成电路的载体,其设计分为前端设计(逻辑设计)和后端设计(物理设计),前端设计涉及功能设计,后端设计涉及工艺有关的设计,包括RTL(Register Transfer Level,寄存器传输级)编写、功能验证、逻辑综合、形式验证、DFT(Design for Testability,可测试性设计)、布局布线、Sign Off(签发)、版图验证等多个流程,使其具备制造意义的芯片。
随着人工智能的发展,工业界正在经历着一场变革,朝着更加智能化的发展方向迈进。据统计,每一款芯片设计生成的数据量平均超过500TB,这对于数据科学来说是一笔不菲的财富,特别是开发具有不同功耗、性能和温度的芯片时,具有极高的利用价值。一般来说,芯片设计工程师面临大量的设计数据和可变性,往往只是利用主观经验浅度参考,导致过度设计或欠设计芯片,造成高昂的决策成本。或者说,怎样利用芯片数据,设计出在性能、功耗、面积等方面表现最佳,去实现更高性能的芯片,机器学习技术是极佳的领域。
目前在芯片设计领域,机器学习的主流发展形式是伴随EDA(Electronic design automation,电子设计自动化)工具,EDA厂商将机器学习模块嵌入到EDA工具中,利用内部的芯片设计数据支持训练模型,智能化分析决策,生成高性能的预期芯片,代表厂商有Cadence、Solido等。在晶圆厂中,机器学习技术被应用在芯片的计量和检测中,以查明芯片中的缺陷。然而,该技术的详细实现被禁止访问,其技术逻辑在内部实现,细节设计在机器学习实验室中完成,所以无法给予特定评价,目前存在的问题包括:芯片数据特征的理解处于较低水平,导致芯片数据的有效利用率不足;技术实现角度固化,并不是自动利用数 据,而是把实验室训练好的模型安装进EDA,这便会导致功能的灵活度较差,而且约束较多,不利于发展。
发明内容
有鉴于此,本申请实施例的目的在于提出一种芯片辅助设计方法、装置、设备及非易失性存储介质,通过特征工程上设计复式加权结构,能够增强对芯片数据特征的理解程度,提高芯片数据的有效利用率,输出更可靠的结果,基于机器学习的芯片辅助设计思路,灵活性更加,约束较少,且成本较低,技术角度能够适用多变的应用场景。
基于上述目的,本申请实施例的一方面提供了一种芯片辅助设计方法,包括以下步骤:获取原始芯片数据,并对上述原始芯片数据进行预处理以得到预处理后的芯片数据;基于芯片经验对上述预处理后的芯片数据进行探索性分析以得到第一特征化数据,基于统计检验对上述预处理后的芯片数据进行过滤法分析以得到第二特征化数据,基于模型方向对上述预处理后的芯片数据进行嵌入法分析以得到第三特征化数据;对上述第一特征化数据、上述第二特征化数据和上述第三特征化数据进行加权处理并进行分析以得到特征化数据集;选择不同的算法基于上述特征化数据集建立不同的预测模型,并对上述预测模型进行训练和评估以得到最优预测模型;以及基于上述最优预测模型对芯片数据的数据特征进行分析以得到数据特征的重要性排序,基于上述重要性排序并且通过上述最优预测模型进行芯片辅助设计。
在一些实施方式中,获取原始芯片数据包括:通过EDA工具获取原始芯片数据;或通过测试数据获取原始芯片数据;或通过服务器数据获取原始芯片数据。
在一些实施方式中,对上述原始芯片数据进行预处理以得到预处理后的芯片数据包括:对上述原始芯片数据进行缺失值处理;和/或对上述原始芯片数据进行重复值处理;和/或对上述原始芯片数据进行异常值处理;和/或对上述原始芯片数据进行字符数据编码处理;和/或对上述原始芯片数据进行无量纲化处理。
在一些实施方式中,基于芯片经验对上述预处理后的芯片数据进行探索性分析以得到第一特征化数据包括:通过直方图分析离散芯片特征的趋势变化;通过核密度曲线分析连续芯片特征的趋势变化;通过计算交叉表分析芯片特征相较于标签的趋势情况;通过分组箱线图分析搭配特征相较于标签的趋势情况;通过相关系数和热力图分析特征与特征、特征与标签之间的相关性;通过主成分分析法分析创造特征。
在一些实施方式中,基于统计检验对上述预处理后的芯片数据进行过滤法分析以得到第二特征化数据包括:基于方差过滤和/或相关性过滤对上述预处理后的芯片数据进行过滤以得到第二特征化数据。
在一些实施方式中,对上述第一特征化数据、上述第二特征化数据和上述第三特征化数据进行加权处理并进行分析以得到特征化数据集包括:利用统计分析软件对加权处理后的上述第一特征化数据、上述第二特征化数据和上述第三特征化数据进行分析。
在一些实施方式中,选择不同的算法基于上述特征化数据集建立不同的预测模型包括:基于上述特征化数据集建立芯片品质预测模型;和/或基于上述特征化数据集建立芯片性能评分模型。
在一些实施方式中,对上述预测模型进行训练和评估以得到最优预测模型包括:通过K次交叉验证、学习曲线、网格搜索中的至少一种对上述预测模型进行训练。
在一些实施方式中,对上述预测模型进行训练和评估以得到最优预测模型包括:通过混淆矩阵、准确率、精确度、召回率、f1分数、roc曲线中的至少一种对上述芯片品质预测模型进行评估。
在一些实施方式中,对上述预测模型进行训练和评估以得到最优预测模型包括:通过均方误差对上述芯片性能评分模型进行评估。
在一些实施方式中,方法包括以下步骤:将上述重要性排序作为反馈信息进行反馈,以对芯片设计进行调整。
在一些实施方式中,方法还包括:获取上述新制芯片的芯片数据,并基于上述最优预测模型对上述新制芯片的芯片数据进行分析。
在一些实施方式中,上述获取的原始芯片数据包括以下至少之一:数据库的结构化数据、文件日志的半结构化数据、音频、视频数据的非结构化数据。
在一些实施方式中,上述对上述原始芯片数据进行缺失值处理包括以下之一:删除上述原始芯片数据中的缺失值;利用0填充上述原始芯片数据中的缺失值;利用同一特征下的均值填充上述原始芯片数据中的缺失值;利用其余特征指标预测上述原始芯片数据中的缺失值。
在一些实施方式中,上述对上述原始芯片数据进行重复值处理包括:删除上述原始芯片数据中的重复值。
在一些实施方式中,上述对上述原始芯片数据进行异常值处理包括:删除上述原始芯片数据中的异常值。
在一些实施方式中,上述基于模型方向对上述预处理后的芯片数据进行嵌入法分析以得到第三特征化数据包括:将所有芯片特征导入到模型中,经过训练调优后,利用模型属性评估上述芯片特征的重要性,通过预先设定的阈值,对所述所有芯片特征进行筛选,得到特征 子集,其中,所述第三特征化数据包括所述特征子集,所述预处理后的芯片数据为所述所有芯片特征。
本申请实施例的另一方面,还提供了一种芯片辅助设计装置,包括:第一模块,被配置为获取原始芯片数据,并对上述原始芯片数据进行预处理以得到预处理后的芯片数据;第二模块,被配置为基于芯片经验对上述预处理后的芯片数据进行探索性分析以得到第一特征化数据,基于统计检验对上述预处理后的芯片数据进行过滤法分析以得到第二特征化数据,基于模型方向对上述预处理后的芯片数据进行嵌入法分析以得到第三特征化数据;第三模块,被配置为对上述第一特征化数据、上述第二特征化数据和上述第三特征化数据进行加权处理并进行分析以得到特征化数据集;第四模块,被配置为选择不同的算法基于上述特征化数据集建立不同的预测模型,并对上述预测模型进行训练和评估以得到最优预测模型;以及第五模块,被配置为基于上述最优预测模型对芯片数据的数据特征进行分析以得到数据特征的重要性排序,基于上述重要性排序并且通过上述最优预测模型进行芯片辅助设计。
本申请实施例的再一方面,还提供了一种计算机设备,包括:至少一个处理器;以及存储器,存储器存储有可在处理器上运行的计算机指令,指令由处理器执行时实现上述方法的步骤。
本申请实施例的再一方面,还提供了一种非易失性存储介质,非易失性存储介质存储有被处理器执行时实现如上方法步骤的计算机程序。
本申请实施例至少具有以下有益技术效果:特征工程上设计复式加权结构,能够增强对芯片数据特征的理解程度,提高芯片数据的有效利用率,输出更可靠的结果,此外,基于机器学习的芯片辅助设计思路,灵活性更加,约束较少,且成本较低,技术角度能够适用多变的应用场景。
附图说明
为了更清楚地说明本申请实施例或现有技术中的技术方案,下面将对实施例或现有技术描述中所需要使用的附图作简单地介绍,显而易见地,下面描述中的附图仅仅是本申请的一些实施例,对于本领域普通技术人员来讲,在不付出创造性劳动的前提下,还可以根据这些附图获得其他的实施例。
图1为本申请实施例提供的芯片辅助设计方法的实施例的示意图;
图2为本申请实施例提供的芯片辅助设计装置的实施例的示意图;
图3为本申请实施例提供的计算机设备的实施例的示意图;
图4为本申请实施例提供的非易失性存储介质的实施例的示意图。
具体实施方式
为使本申请实施例的目的、技术方案和优点更加清楚明白,以下结合可选实施例,并参照附图,对本申请实施例进行详细说明。
需要说明的是,本申请实施例中所有使用“第一”和“第二”的表述均是为了区分两个相同名称非相同的实体或者非相同的参量,可见“第一”“第二”仅为了表述的方便,不应理解为对本申请实施例的限定,后续实施例对此不再一一说明。
基于上述目的,本申请实施例的第一个方面,提出了芯片辅助设计方法的实施例。图1示出的是本申请提供的芯片辅助设计方法的实施例的示意图。如图1所示,本申请实施例的芯片辅助设计方法包括如下步骤:
001、获取原始芯片数据,并对原始芯片数据进行预处理以得到预处理后的芯片数据;
002、基于芯片经验对预处理后的芯片数据进行探索性分析以得到第一特征化数据,基于统计检验对预处理后的芯片数据进行过滤法分析以得到第二特征化数据,基于模型方向对预处理后的芯片数据进行嵌入法分析以得到第三特征化数据;
003、对第一特征化数据、第二特征化数据和第三特征化数据进行加权处理并进行分析以得到特征化数据集;
004、选择不同的算法基于特征化数据集建立不同的预测模型,并对预测模型进行训练和评估以得到最优预测模型;以及
005、基于最优预测模型对芯片数据的数据特征进行分析以得到数据特征的重要性排序,基于重要性排序并且通过上述最优预测模型进行芯片辅助设计。
在本实施例中,在特征工程中,通过复式加权结构,探索性分析、过滤法、嵌入法,再通过SPSS(Statistical Product Service Solutions,统计产品与服务解决方案)专业统计分析的方式可以解决芯片数据特征理解度过低的困境,提高芯片数据的有效利用率。开放式利用芯片数据更具发展,技术角度分支多,灵活性高,且约束较少,无论是对于芯片的品质预测,或是对于芯片的性能评分,我们都给出了自己的解决方案。基于机器学习的芯片应用不仅仅局限于此,可能涉及到更复杂的应用场景,这取决于项目的需求。
在本申请的一些实施例中,获取原始芯片数据包括:通过EDA工具获取原始芯片数据;或通过测试数据获取原始芯片数据;或通过服务器数据获取原始芯片数据。
在本实施例中,假定芯片数据完整存在于数据库中并标记;芯片数据可来源于EDA工具 中,如仿真数据、变化性度量,或来源于芯片的测试数据、或来源于服务器数据等,包括但不限于电流、电压、温度、频率、数据流量、工艺库参数、IP(Internet Protocol,互联网协议)选型等。
收集的数据包括:数据库的结构化数据、文件日志的半结构化数据、音频、视频数据的非结构化数据。其中文件日志通过Flume(一种分布式的海量日志采集、聚合和传输的系统)收集,音频、视频通过Kafka(一种开源流处理平台)消息队列收集。
在本申请的一些实施例中,对原始芯片数据进行预处理以得到预处理后的芯片数据包括:对原始芯片数据进行缺失值处理;和/或对原始芯片数据进行重复值处理;和/或对原始芯片数据进行异常值处理;和/或对原始芯片数据进行字符数据编码处理;和/或对原始芯片数据进行无量纲化处理。
在本实施例中,对原始复杂的芯片数据清洗,使其规范并可以使用数学语言描述,预处理模块涉及缺失值、重复值、异常值、字符数据编码、无量纲化等处理方案。
在本实施例中,对于缺失值处理包括:直接删除;0填充,利用0填充缺失值;均值填充,利用该特征下的均值填充缺失值;预测填充,利用其余特征指标预测缺失值。
作为一种示例,本申请的各个实施例中的缺失值用于指示芯片特征所对应的数值中出现丢失的数值。
在本实施例中,对于重复值处理包括:直接删除去重。
在本实施例中,对于异常值处理包括:既定规则,按照物理规律,删除超异常的items(项);3σ原则,按照统计学经验,若芯片数据服从正态分布,数值分布在(μ-3σ,μ+3σ)中的概率为0.9974,即P(|x-μ|>3σ)≤0.003,距离在均值μ的3σ外的芯片数据可认定为异常值(统计上,样本量大于30可假定数据服从正态分布,即使非正态分布,也基本满足3σ原则)。其中,超异常的items(项)为超过异常阈值的item(项),即超出(μ-3σ,μ+3σ)。在一个可选的示例中,“超异常”和“异常”可理解为同一含义,“删除超异常item(项)”可被理解为“删除异常值”。
在本实施例中,对原始芯片数据进行字符数据编码处理包括:直接编码,适用于分类变量;独热编码,适用于名义变量;分箱编码,根据需要,可基于距离分箱、分位数分箱、熵值分箱等;word embedding(词嵌入),利用神经网络编码。
在本实施例中,对原始芯片数据进行无量纲化处理包括:归一化,公式如下:
Figure PCTCN2022141691-appb-000001
一般值域缩放至[0,1];
标准化公式如下:
Figure PCTCN2022141691-appb-000002
芯片数据经过预处理之后,可理解转化为高质量数据,有利于提高结果的准确性。
在本申请的一些实施例中,基于芯片经验对预处理后的芯片数据进行探索性分析以得到第一特征化数据包括:通过直方图分析离散芯片特征的趋势变化;通过核密度曲线分析连续芯片特征的趋势变化;通过计算交叉表分析芯片特征相较于标签的趋势情况;通过分组箱线图分析搭配特征相较于标签的趋势情况;通过相关系数和热力图分析特征与特征、特征与标签之间的相关性;通过主成分分析法分析创造特征。
在本实施例中,包括但不限于以下分析方法,可以充分借鉴《统计学基础》:
直方图,分析离散芯片特征的趋势变化;
核密度曲线,分析连续芯片特征的趋势变化;
计算交叉表,分析芯片特征相较于标签的趋势情况;
分组箱线图,分析搭配特征相较于标签的趋势情况;
相关系数和热力图,分析特征与特征,特征与标签之间的相关性;
创造特征,主成分分析法、科学经验等;
芯片设计工程师根据工程经验、探索性分析结果等,选择出主要芯片数据特征,获取特征化数据T1。
在本申请的一些实施例中,基于统计检验对预处理后的芯片数据进行过滤法分析以得到第二特征化数据包括:基于方差过滤和/或相关性过滤对预处理后的芯片数据进行过滤以得到第二特征化数据。
在本申请的一些实施例中,对上述第一特征化数据、上述第二特征化数据和上述第三特征化数据进行加权处理并进行分析以得到特征化数据集包括:利用统计分析软件对加权处理后的上述第一特征化数据、上述第二特征化数据和上述第三特征化数据进行分析。
在本实施例中,过滤法主要考虑方差、相关性等进行特征筛取,可设定阈值控制芯片特征的留余量。包括:
方差过率,过滤掉方差低于阈值Threshold的芯片特征,一般认为此类特征无意义;
相关性过率,过滤掉与标签相关性较低的芯片特征,一般认为此类特征对标签影响很 小,包括:卡方检验,考虑芯片特征与标签之间的独立性;F检验,考虑芯片特征与标签间的线性关系;互信息法,考虑数芯片征与标签间的线性关系与非线性关系;
获取过滤法后的特征化数据T2。
在本申请的一些实施例中,基于模型方向对预处理后的芯片数据进行嵌入法分析以得到第三特征化数据。利用模型本身挑选芯片特征,工业界中常用的特征选择方式,可以有效提高模型得分。嵌入法对模型有限制条件,一般来讲,模型应具有coef_属性、feature_importances_属性、L1或L2惩罚项等其中之一。
嵌入法将所有芯片特征导入到模型(如随机森林)中,经过训练调优后,利用模型属性feature_importances_评估特征的重要性,设定阈值,得到数据特征子集(或称为,特征子集),即获取特征化数据T3。嵌入法一般存在模型要求,即存在feature_importances_属性或惩罚项等,建模后可对特征量化理解,设定阈值,根据阈值保留重要特征,得到特征子集。在一种可选的示例中,上述“将所有芯片特征导入到模型中”可以但不限于被理解为:将预处理后的芯片数据导入到模型中,而后利用模型筛选特征,构造特征子集。在一种可选的示例中,上述预处理后的芯片数据可以为上述所有芯片特征。在一种可选的示例中,随机森林模型是一种示例,还可以采用其他的模型,例如,逻辑回归模型等等。可以理解的是,对于上述使用的模型(例如,随机森林模型或逻辑回归模型),都可预先设定用于筛选特征的阈值。
在本实施例中,将三个角度获取到的数据特征按照出现次数进行统计,并赋予1,2,3的权重,将加权后的数据导入统计分析软件进行分析。在可选的实施例中,将加权后的数据导入SPSS(如果内存不够,可以采样)专业分析软件,做系统的统计分析,如方差齐性检验、相关性分析、主成分分析、因子分析、回归分析、聚类、方差扩大因子法(去除多重共线性)等,得到最终的特征化数据集T。利用SPSS分析的目的是为了得到关于芯片数据的最终的特征化数据集T,最终的特征化数据集T是通过加权芯片的特征化数据T1、T2、T3并经分析得到的,属于数据的再利用,例如以下上述的在相关性分析和主成分分析中使用的情况。
在本实施例中,以相关性分析为例,保留高权重特征,并将高权重的关联特征次数累加,收集排名前n的弱数据特征,加工处理,划到最终集;
在本实施例中,以主成分分析为例,主要将原始高维空间数据投影到低维数据空间,可将权重3的数据特征值域增至3倍,权重2的数据特征值域增值2倍,降维操作时投影会向该类数据特征偏斜,提高关注度;
在本申请的一些实施例中,选择不同的算法基于特征化数据集建立不同的预测模型包 括:基于特征化数据集建立芯片品质预测模型;和/或基于特征化数据集建立芯片性能评分模型。对上述预测模型进行训练和评估以得到最优预测模型包括:通过K次交叉验证、学习曲线、网格搜索中的至少一种对上述预测模型进行训练。
在本实施例中,算法建模是模型选择、训练优化的过程。基于特征化数据T建立芯片品质预测模型①、基于特征化数据T建立芯片性能预测模型②;
在本实施例中,芯片品质预测模型①属于分类问题,假定只判断一块芯片的异常与否,即二分类模型,考虑逻辑回归、随机森林、支持向量机、朴素贝叶斯等分类算法;假如采用随机森林,树模型的Bagging(Bootstrap aggregating,引导聚集算法)集成模型,其优点是可处理任意数据类型,擅长处理高维数据,模型泛化能力强。详细而言,把输入的芯片数据设为x,芯片是否异常设定为目标ya,ya=1异常芯片,ya=0正常芯片。二分类模型ya=f(x),f作为随机森林函数,能够把芯片数据x映射到目标ya中,f的确定是以熵增来确定的。
在本实施例中,芯片性能评分模型②属于回归问题,即对一块芯片的性能给出预测得分,考虑线性回归模型等。假如采用线性回归模型,基本的数学模型,其优点是简单且可解释性强;详细而言,把输入的芯片数据设为x,芯片的性能得分设为目标yb,yb的大小代表芯片的性能,yb越大代表芯片的性能越优,yb越小代表芯片的性能越差;线性回归模型yb=g(x),g作为线性回归函数,能够把芯片数据x映射到目标yb中,g确定是对芯片数据的拟合,最小二乘法即可。
芯片品质预测模型①或芯片性能预测模型②,都需要训练过程,可选择K次交叉验证、学习曲线、网格搜索等技巧,提高模型的得分,包括:
K次交叉验证,将数据集分成K份,依次将其中K-1份作为训练集,剩下1份作为测试集,重复K次,观察训练结果;
学习曲线,分析模型拟合情况;
网格搜索,通过穷举参数、组合参数进行交叉验证优化模型的一种方法。
在通过上述方法对模型进行训练后,在本申请的一些实施例中,可以通过混淆矩阵、准确率、精确度、召回率、f1分数、roc曲线中的至少一种对芯片品质预测模型进行评估,或者通过均方误差对芯片性能评分模型进行评估。完成评估后,将对模型进行分析,详情将在后续可选实施例中进行说明。
在完成分析后,在本申请的一些实施例中,方法,可以包括以下步骤:将上述重要性排序作为反馈信息进行反馈,以对芯片设计进行调整。可以将重要性排序反馈给芯片设计工程 师,芯片设计工程师在收到该反馈信息后可以根据不同情况对芯片进行调整。在一个实施例中,芯片设计工程师可以根据上述反馈信息设计新制芯片,获取新制芯片数据,并基于最优预测模型对新制芯片的芯片数据进行分析。例如,当芯片设计工程师利用反馈信息设计了一款新制芯片,可将新制芯片的数据导入到模型中,进行数据仿真验证。
下面根据可选实施例进行阐述本申请实施例的可选实施方式。针对现有背景技术存在的不足,本申请实施例提供基于机器学习的开放式芯片辅助设计方法,特别地,特征工程上设计复式加权结构,能够增强对芯片数据特征的理解程度,提高芯片数据的有效利用率,输出更可靠的结果。此外,基于机器学习的开放式芯片辅助设计思路,灵活性更佳,约束较少,且成本较低,技术角度能够适用多变的应用场景。详细如下:
基于机器学习的开放式芯片辅助设计方法,包括芯片设计数据、数据预处理模块、特征工程模块、算法建模模块、评估模块、分析等。
芯片数据,假定完整存在于数据库中并标记。数据来源于EDA工具,如仿真数据、变化性度量,或来源于芯片的测试数据、或来源于服务器数据等;
数据预处理模块,对原始复杂的芯片数据清洗,使其规范并可以使用数学语言描述。预处理模块涉及缺失值、重复值、异常值、字符数据编码、无量纲化等处理方案;
特征工程模块,设计复式加权结构创新,抽取、创造具有价值的芯片数据,为后续算法建模提供特征化数据。以三个角度出发,开展数据特征工程,以芯片经验方向做探索性分析获取特征化数据T1,以统计检验方向做过滤法分析获取特征化数据T2,以模型方向做嵌入法分析获取特征化数据T3,将三个角度得到的数据特征按照出现次数统计,并赋予1,2,3的权重,将加权后的数据导入SPSS(如果内存不够,可以采样)专业分析软件,做系统的统计分析,如方差齐性检验、主成分分析、因子分析、回归分析、聚类、方差扩大因子法(去除多重共线性)等,得到完善的特征化数据T;
算法建模模块,基于特征化数据T建立芯片品质预测模型①或基于特征化数据T建立芯片性能预测模型②。芯片品质预测模型①,属于分类问题,假定只判断一块芯片的异常与否,即二分类模型,考虑逻辑回归、随机森林、支持向量机、朴素贝叶斯等分类算法;芯片性能预测模型②,属于回归问题,即对一块芯片的性能给出预测得分,考虑线性回归模型等。无论芯片品质预测模型①或芯片性能预测模型②,都需要训练过程,可选择K次交叉验证、学习曲线、网格搜索等技巧,提高模型的得分;
评估模块,评估芯片数据在不同算法下的预测效果,基于评估结果选择最优的算法。评估的指标取决于模型的种类,对于芯片品质预测模型①,可以采用混淆矩阵 (confusion_matrix)、准确率、精确度、召回率、f1score、roc曲线(Receiver Operating Characteristic Curve,接收者操作特征曲线)等衡量模型效果;对于芯片性能预测模型②,可以采用均方误差MSE等衡量模型的效果;包括以下分类评价指标:
混淆矩阵(confusion_matrix),二分类问题的多维衡量指标体系,1代表异常芯片,0代表正常芯片。在混淆矩阵中,真实值在前,预测值在后,如下表所示:
Figure PCTCN2022141691-appb-000003
准确率(Accuracy),所有预测正确的样本占总样本量,公式如下:
Figure PCTCN2022141691-appb-000004
精确度(Precision),所有预测为异常类中真正异常类占比,公式如下:
Figure PCTCN2022141691-appb-000005
召回率(Recall),所有预测的异常类占总体异常类比,捕捉异常类,公式如下:
Figure PCTCN2022141691-appb-000006
f1score,基于Precision和Recall的综合统计量,公式如下:
Figure PCTCN2022141691-appb-000007
roc曲线,以假正率为横坐标、召回率为纵坐标的一条曲线,一般情况,选择模型的准确率作为评价指标;当面对芯片数据不均衡,会考虑f1score(f1分数)。
回归评价指标包括均方误差(MSE,Mean Square Error),公式如下:
Figure PCTCN2022141691-appb-000008
回归评价指标也可使用均方根误差(RMSE,Root Mean Square Error)、平均绝对误差(MAE,Mean Absolute Error)、判定系数(Coefficient of Determination)等。
分析,通过建立的算法模型探索芯片数据潜在的物理规律。对于芯片品质预测模型①,若采用逻辑回归模型,可以使用系数衡量芯片特征的重要性,即系数的绝对值越大,该芯片特征影响芯片的品质因素越深,如果系数为正,说明该芯片特征具有积极作用。若为负,说明该芯片特征具有消极作用。若采用随机森林模型,可以使用属性feature_importances_衡量芯片特征的重要性,其值越大,说明影响因素越大,通过归纳总结可以学到影响芯片品质的特征;对于芯片性能预测模型②,MSE越小意味着模型的准确性越高,该得分可以描述对芯片性能预测的可信度。
当我们分析芯片数据的潜在物理规律后,可以向芯片设计工程师反馈信息,若是品质分类模型,可以根据归纳的芯片特征重要性逐步调整芯片设计细节,如通过改变部分电路结构、工艺参数、布线等,调整该特征或部分特征下的数据,使其由异常芯片转化为正常芯片,提高效率;若是性能预测模型,可以通过数学优化方法,计算芯片输入特征数据使其性能最大化,即芯片的最优设计,便可基于此设计出高性能芯片,即完成基于机器学习的辅助芯片设计。
需要特别指出的是,上述芯片辅助设计方法的各个实施例中的各个步骤均可以相互交叉、替换、增加、删减,因此,这些合理的排列组合变换之于芯片辅助设计方法也应当属于本申请实施例的保护范围,并且不应将本申请实施例的保护范围局限在实施例之上。
基于上述目的,本申请实施例的第二个方面,提出了一种芯片辅助设计装置。图2示出的是本申请实施例提供的芯片辅助设计装置的实施例的示意图。如图2所示,本申请实施例的芯片辅助设计装置包括如下模块:第一模块011,被配置为获取原始芯片数据,并对原始芯片数据进行预处理以得到预处理后的芯片数据;第二模块012,被配置为基于芯片经验对预处理后的芯片数据进行探索性分析以得到第一特征化数据,基于统计检验对预处理后的芯片数据进行过滤法分析以得到第二特征化数据,基于模型方向对预处理后的芯片数据进行嵌入法分析以得到第三特征化数据;第三模块013,被配置为对第一特征化数据、第二特征化数据和第三特征化数据进行加权处理并进行分析以得到特征化数据集;第四模块014,被配置为选择不同的算法基于特征化数据集建立不同的预测模型,并对预测模型进行训练和评估以得到最优预测模型;以及第五模块015,被配置为基于最优预测模型对芯片数据的数据特征进行分析以得到数据特征的重要性排序,基于重要性排序并且通过上述最优预测模型进行芯片辅助设计。
基于上述目的,本申请实施例的第三个方面,提出了一种计算机设备。图3示出的是本申请实施例提供的计算机设备的实施例的示意图。如图3所示,本申请实施例的计算机设备包括如下装置:至少一个处理器021;以及存储器022,存储器022存储有可在处理器上运行的计算机指令023,指令由处理器执行时实现以上方法的步骤。
本申请实施例还提供了一种非易失性存储介质。图4示出的是本申请实施例提供的非易失性存储介质的实施例的示意图。如图4所示,非易失性存储介质031存储有被处理器执行时执行如上方法的计算机程序032。
最后需要说明的是,本领域普通技术人员可以理解实现上述实施例方法中的全部或部分流程,可以通过计算机程序来指令相关硬件来完成,芯片辅助设计方法的程序可存储于一计算机可读取存储介质中,该程序在执行时,可包括如上述各方法的实施例的流程。其中,程序的存储介质可为磁碟、光盘、只读存储记忆体(ROM(Read-Only Memory,只读存储器)或随机存储记忆体(RAM(Random Access Memory,随机存取存储器)等。上述计算机程序的实施例,可以达到与之对应的前述任意方法实施例相同或者相类似的效果。
此外,根据本申请实施例公开的方法还可以被实现为由处理器执行的计算机程序,该计算机程序可以存储在非易失性存储介质中。在该计算机程序被处理器执行时,执行本申请实施例公开的方法中限定的上述功能。
此外,上述方法步骤以及系统单元也可以利用控制器以及用于存储使得控制器实现上述步骤或单元功能的计算机程序的非易失性存储介质实现。
本领域技术人员还将明白的是,结合这里的公开所描述的各种示例性逻辑块、模块、电路和算法步骤可以被实现为电子硬件、计算机软件或两者的组合。为了清楚地说明硬件和软件的这种可互换性,已经就各种示意性组件、方块、模块、电路和步骤的功能对其进行了一般性的描述。这种功能是被实现为软件还是被实现为硬件取决于详细应用以及施加给整个系统的设计约束。本领域技术人员可以针对每种应用以各种方式来实现的功能,但是这种实现决定不应被解释为导致脱离本申请实施例公开的范围。
在一个或多个示例性设计中,功能可以在硬件、软件、固件或其任意组合中实现。如果在软件中实现,则可以将功能作为一个或多个指令或代码存储在非易失性存储介质上或通过非易失性存储介质来传送。非易失性存储介质包括计算机存储介质和通信介质,该通信介质包括有助于将计算机程序从一个位置传送到另一个位置的任何介质。存储介质可以是能够被通用或专用计算机访问的任何可用介质。作为例子而非限制性的,该非易失性存储介质可以包括RAM、ROM、EEPROM(Electrically Erasable Programmable read only memory,带电 可擦可编程只读存储器)、CD-ROM(Compact Disc Read-Only Memory,只读光盘)或其它光盘存储设备、磁盘存储设备或其它磁性存储设备,或者是可以用于携带或存储形式为指令或数据结构的所需程序代码并且能够被通用或专用计算机或者通用或专用处理器访问的任何其它介质。此外,任何连接都可以适当地称为非易失性存储介质。例如,如果使用同轴线缆、光纤线缆、双绞线、数字用户线路(DSL,Digital Subscriber Line)或诸如红外线、无线电和微波的无线技术来从网站、服务器或其它远程源发送软件,则上述同轴线缆、光纤线缆、双绞线、DSL或诸如红外线、无线电和微波的无线技术均包括在介质的定义。如这里所使用的,磁盘和光盘包括压缩盘(CD)、激光盘、光盘、数字多功能盘(DVD,Digital Video Disc)、软盘、蓝光盘,其中磁盘通常磁性地再现数据,而光盘利用激光光学地再现数据。上述内容的组合也应当包括在非易失性存储介质的范围内。
以上是本申请公开的示例性实施例,但是应当注意,在不背离权利要求限定的本申请实施例公开的范围的前提下,可以进行多种改变和修改。根据这里描述的公开实施例的方法权利要求的功能、步骤和/或动作不需以任何特定顺序执行。此外,尽管本申请实施例公开的元素可以以个体形式描述或要求,但除非明确限制为单数,也可以理解为多个。
应当理解的是,在本文中使用的,除非上下文清楚地支持例外情况,单数形式“一个”旨在也包括复数形式。还应当理解的是,在本文中使用的“和/或”是指包括一个或者一个以上相关联地列出的项目的任意和所有可能组合。
上述本申请实施例公开实施例序号仅仅为了描述,不代表实施例的优劣。
本领域普通技术人员可以理解实现上述实施例的全部或部分步骤可以通过硬件来完成,也可以通过程序来指令相关的硬件完成,程序可以存储于一种非易失性存储介质中,上述提到的存储介质可以是只读存储器,磁盘或光盘等。
所属领域的普通技术人员应当理解:以上任何实施例的讨论仅为示例性的,并非旨在暗示本申请实施例公开的范围(包括权利要求)被限于这些例子;在本申请实施例的思路下,以上实施例或者不同实施例中的技术特征之间也可以进行组合,并存在如上的本申请实施例的不同方面的许多其它变化,为了简明它们没有在细节中提供。因此,凡在本申请实施例的精神和原则之内,所做的任何省略、修改、等同替换、改进等,均应包含在本申请实施例的保护范围之内。

Claims (20)

  1. 一种芯片辅助设计方法,其特征在于,包括以下步骤:
    获取原始芯片数据,并对所述原始芯片数据进行预处理以得到预处理后的芯片数据;
    基于芯片经验对所述预处理后的芯片数据进行探索性分析以得到第一特征化数据,基于统计检验对所述预处理后的芯片数据进行过滤法分析以得到第二特征化数据,基于模型方向对所述预处理后的芯片数据进行嵌入法分析以得到第三特征化数据;
    对所述第一特征化数据、所述第二特征化数据和所述第三特征化数据进行加权处理并进行分析以得到特征化数据集;
    选择不同的算法基于所述特征化数据集建立不同的预测模型,并对所述预测模型进行训练和评估以得到最优预测模型;以及
    基于所述最优预测模型对芯片数据的数据特征进行分析以得到数据特征的重要性排序,基于所述重要性排序并且通过所述最优预测模型进行芯片辅助设计。
  2. 根据权利要求1所述的芯片辅助设计方法,其特征在于,获取原始芯片数据包括:
    通过EDA工具获取原始芯片数据;或
    通过测试数据获取原始芯片数据;或
    通过服务器数据获取原始芯片数据。
  3. 根据权利要求1所述的芯片辅助设计方法,其特征在于,对所述原始芯片数据进行预处理以得到预处理后的芯片数据包括:
    对所述原始芯片数据进行缺失值处理;和/或
    对所述原始芯片数据进行重复值处理;和/或
    对所述原始芯片数据进行异常值处理;和/或
    对所述原始芯片数据进行字符数据编码处理;和/或
    对所述原始芯片数据进行无量纲化处理。
  4. 根据权利要求1所述的芯片辅助设计方法,其特征在于,基于芯片经验对所述预处理后的芯片数据进行探索性分析以得到第一特征化数据包括:
    通过直方图分析离散芯片特征的趋势变化;
    通过核密度曲线分析连续芯片特征的趋势变化;
    通过计算交叉表分析芯片特征相较于标签的趋势情况;
    通过分组箱线图分析搭配特征相较于标签的趋势情况;
    通过相关系数和热力图分析特征与特征、特征与标签之间的相关性;
    通过主成分分析法分析创造特征。
  5. 根据权利要求1所述的芯片辅助设计方法,其特征在于,基于统计检验对所述预处理后的芯片数据进行过滤法分析以得到第二特征化数据包括:
    基于方差过滤和/或相关性过滤对所述预处理后的芯片数据进行过滤以得到第二特征化数据。
  6. 根据权利要求1所述的芯片辅助设计方法,其特征在于,对所述第一特征化数据、所述第二特征化数据和所述第三特征化数据进行加权处理并进行分析以得到特征化数据集包括:
    利用统计分析软件对加权处理后的所述第一特征化数据、所述第二特征化数据和所述第三特征化数据进行分析。
  7. 根据权利要求1所述的芯片辅助设计方法,其特征在于,选择不同的算法基于所述特征化数据集建立不同的预测模型包括:
    基于所述特征化数据集建立芯片品质预测模型;和/或
    基于所述特征化数据集建立芯片性能评分模型。
  8. 根据权利要求1所述的芯片辅助设计方法,其特征在于,对所述预测模型进行训练和评估以得到最优预测模型包括:
    通过K次交叉验证、学习曲线、网格搜索中的至少一种对所述预测模型进行训练。
  9. 根据权利要求7所述的芯片辅助设计方法,其特征在于,对所述预测模型进行训练和评估以得到最优预测模型包括:
    通过混淆矩阵、准确率、精确度、召回率、f1分数、roc曲线中的至少一种对所述芯片品质预测模型进行评估。
  10. 根据权利要求7所述的芯片辅助设计方法,其特征在于,对所述预测模型进行训练和评估以得到最优预测模型包括:
    通过均方误差对所述芯片性能评分模型进行评估。
  11. 根据权利要求1所述的芯片辅助设计方法,其特征在于,包括以下步骤:
    将所述重要性排序作为反馈信息进行反馈,以对芯片设计进行调整。
  12. 根据权利要求11所述的芯片辅助设计方法,其特征在于,还包括:
    根据所述反馈信息设计新制芯片,
    获取所述新制芯片的芯片数据,并基于所述最优预测模型对所述新制芯片的芯片数据进行分析。
  13. 根据权利要求1所述的芯片辅助设计方法,其特征在于,所述获取的原始芯片数据包括以下至少之一:数据库的结构化数据、文件日志的半结构化数据、音频、视频数据的非结构化数据。
  14. 根据权利要求3所述的芯片辅助设计方法,其特征在于,所述对所述原始芯片数据进行缺失值处理包括以下之一:
    删除所述原始芯片数据中的缺失值;
    利用0填充所述原始芯片数据中的缺失值;
    利用同一特征下的均值填充所述原始芯片数据中的缺失值;
    利用其余特征指标预测所述原始芯片数据中的缺失值。
  15. 根据权利要求3所述的芯片辅助设计方法,其特征在于,所述对所述原始芯片数据进行重复值处理包括:
    删除所述原始芯片数据中的重复值。
  16. 根据权利要求3所述的芯片辅助设计方法,其特征在于,所述对所述原始芯片数据进行异常值处理包括:
    删除所述原始芯片数据中的异常值。
  17. 根据权利要求1所述的芯片辅助设计方法,其特征在于,所述基于模型方向对所述预处理后的芯片数据进行嵌入法分析以得到第三特征化数据包括:
    将所有芯片特征导入到模型中,经过训练调优后,利用模型属性评估所述芯片特征的重要性,通过预先设定的阈值,对所述所有芯片特征进行筛选,得到特征子集,其中,所述第三特征化数据包括所述特征子集,所述预处理后的芯片数据为所述所有芯片特征。
  18. 一种芯片辅助设计装置,其特征在于,包括:
    第一模块,被配置为获取原始芯片数据,并对所述原始芯片数据进行预处理以得到预处理后的芯片数据;
    第二模块,被配置为基于芯片经验对所述预处理后的芯片数据进行探索性分析以得到第一特征化数据,基于统计检验对所述预处理后的芯片数据进行过滤法分析以得到第二特征化数据,基于模型方向对所述预处理后的芯片数据进行嵌入法分析以得到第三特征化数据;
    第三模块,被配置为对所述第一特征化数据、所述第二特征化数据和所述第三特征 化数据进行加权处理并进行分析以得到特征化数据集;
    第四模块,被配置为选择不同的算法基于所述特征化数据集建立不同的预测模型,并对所述预测模型进行训练和评估以得到最优预测模型;以及
    第五模块,被配置为基于所述最优预测模型对芯片数据的数据特征进行分析以得到数据特征的重要性排序,基于所述重要性排序并且通过所述最优预测模型进行芯片辅助设计。
  19. 一种计算机设备,其特征在于,包括:
    至少一个处理器;以及
    存储器,上述存储器存储有可在上述处理器上运行的计算机指令,上述指令由上述处理器执行时实现权利要求1-17任意一项上述方法的步骤。
  20. 一种非易失性存储介质,所述非易失性存储介质存储有计算机程序,其特征在于,上述计算机程序被处理器执行时实现权利要求1-17任意一项上述方法的步骤。
PCT/CN2022/141691 2022-05-24 2022-12-23 一种芯片辅助设计方法、装置、设备及非易失性存储介质 WO2023226423A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210569483.1 2022-05-24
CN202210569483.1A CN114722746B (zh) 2022-05-24 2022-05-24 一种芯片辅助设计方法、装置、设备及可读介质

Publications (1)

Publication Number Publication Date
WO2023226423A1 true WO2023226423A1 (zh) 2023-11-30

Family

ID=82232161

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/141691 WO2023226423A1 (zh) 2022-05-24 2022-12-23 一种芯片辅助设计方法、装置、设备及非易失性存储介质

Country Status (2)

Country Link
CN (1) CN114722746B (zh)
WO (1) WO2023226423A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117571742A (zh) * 2024-01-12 2024-02-20 贵州大学 基于人工智能实现芯片质检的方法及装置

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114722746B (zh) * 2022-05-24 2022-11-01 苏州浪潮智能科技有限公司 一种芯片辅助设计方法、装置、设备及可读介质
CN116187235A (zh) * 2023-04-25 2023-05-30 九识(苏州)智能科技有限公司 一种基于数学建模设计芯片架构的方法及系统

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103164567A (zh) * 2012-12-04 2013-06-19 天津蓝海微科技有限公司 一种根据流片数据拟合晶圆参数的方法
US20190363880A1 (en) * 2019-07-10 2019-11-28 Lg Electronics Inc. System on chip, method for operating the system on chip and electronic device including the system on chip
US10628546B1 (en) * 2018-06-29 2020-04-21 Cadence Design Systems, Inc. Method and system for automatically extracting layout design patterns for custom layout design reuse through interactive recommendations
US20220004900A1 (en) * 2018-11-26 2022-01-06 Agency For Science, Technology And Research Method and system for predicting performance in electronic design based on machine learning
CN114520694A (zh) * 2022-04-21 2022-05-20 苏州浪潮智能科技有限公司 一种计算芯片、系统及数据处理方法
CN114722746A (zh) * 2022-05-24 2022-07-08 苏州浪潮智能科技有限公司 一种芯片辅助设计方法、装置、设备及可读介质

Family Cites Families (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2005072054A2 (en) * 2004-01-27 2005-08-11 Chipvision Design Systems Ag Predictable design of low power systems by pre-implementation estimation and optimization
CN101276322A (zh) * 2008-05-23 2008-10-01 首都师范大学 一种具有公式发现功能的嵌入式智能芯片
CN111079937A (zh) * 2019-11-15 2020-04-28 苏州金智渠信息技术有限公司 一种快速建模的方法
CN111339720B (zh) * 2020-02-28 2023-08-29 深圳市赤狐软件技术有限公司 基于大数据采集的芯片信息生成方法、装置及计算机设备
CN112163394B (zh) * 2020-09-28 2023-05-12 海光信息技术股份有限公司 一种cpu芯片设计方法、装置及电子设备
CN113204938B (zh) * 2021-04-06 2022-08-12 普赛微科技(杭州)有限公司 一种集成电路的时延特性改进方法、装置及存储介质
CN113204936B (zh) * 2021-07-02 2021-09-17 苏州贝克微电子有限公司 一种自动添加环境稳定系统的芯片设计方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103164567A (zh) * 2012-12-04 2013-06-19 天津蓝海微科技有限公司 一种根据流片数据拟合晶圆参数的方法
US10628546B1 (en) * 2018-06-29 2020-04-21 Cadence Design Systems, Inc. Method and system for automatically extracting layout design patterns for custom layout design reuse through interactive recommendations
US20220004900A1 (en) * 2018-11-26 2022-01-06 Agency For Science, Technology And Research Method and system for predicting performance in electronic design based on machine learning
US20190363880A1 (en) * 2019-07-10 2019-11-28 Lg Electronics Inc. System on chip, method for operating the system on chip and electronic device including the system on chip
CN114520694A (zh) * 2022-04-21 2022-05-20 苏州浪潮智能科技有限公司 一种计算芯片、系统及数据处理方法
CN114722746A (zh) * 2022-05-24 2022-07-08 苏州浪潮智能科技有限公司 一种芯片辅助设计方法、装置、设备及可读介质

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117571742A (zh) * 2024-01-12 2024-02-20 贵州大学 基于人工智能实现芯片质检的方法及装置
CN117571742B (zh) * 2024-01-12 2024-04-05 贵州大学 基于人工智能实现芯片质检的方法及装置

Also Published As

Publication number Publication date
CN114722746B (zh) 2022-11-01
CN114722746A (zh) 2022-07-08

Similar Documents

Publication Publication Date Title
WO2023226423A1 (zh) 一种芯片辅助设计方法、装置、设备及非易失性存储介质
CN111612039B (zh) 异常用户识别的方法及装置、存储介质、电子设备
CN114444986B (zh) 产品分析方法及系统及装置及介质
US20180082215A1 (en) Information processing apparatus and information processing method
CN107168995B (zh) 一种数据处理方法及服务器
CN112767106B (zh) 自动化审计方法、系统、计算机可读存储介质及审计设备
Xu et al. Hybrid feature selection for wafer acceptance test parameters in semiconductor manufacturing
CN117236278B (zh) 一种基于数字孪生技术的芯片生产仿真方法及系统
CN112434024A (zh) 面向关系型数据库的数据字典生成方法、装置、设备及介质
CN112487146A (zh) 一种法律案件争议焦点获取方法、装置以及计算机设备
CN115630433A (zh) 一种基于机器学习和bim技术的建筑设计方法及系统
CN117131449A (zh) 面向数据治理的具有传播学习能力的异常识别方法及系统
Goyle et al. Dataassist: A machine learning approach to data cleaning and preparation
US11227288B1 (en) Systems and methods for integration of disparate data feeds for unified data monitoring
Revindasari et al. Traceability between business process and software component using Probabilistic Latent Semantic Analysis
KR101456187B1 (ko) 복합 평가 요소에 기초한 특허 평가 방법
Srinivas et al. A Comprehensive Survey of Big Data in the Age of AI
CN114139482A (zh) 一种基于深度度量学习的eda电路失效分析方法
CN113254428A (zh) 一种基于决策树的缺失数据填充方法及系统
WO2024103436A1 (zh) 设备缺陷数据规则库构建方法及设备缺陷关联性分析方法
CN117194963B (zh) 工业fdc质量根因分析方法、设备及存储介质
CN116451771B (zh) 图像分类卷积神经网络压缩方法及芯粒器件数据分配方法
CN117828539B (zh) 数据智能融合分析系统及方法
US20240104436A1 (en) Chained feature synthesis and dimensional reduction
Xu et al. ML-Accelerated Yield Analysis Framework Using Regularization for Sparsity in High-Sigma and High-Dimensional Scenarios

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22943608

Country of ref document: EP

Kind code of ref document: A1