Research On Short Violation Prediction Method In Backend Design Based On Machine Learning

With the expansion of chip scale and shrinking of process nodes,the design rules of backend design become more and more complex,and the design cycle becomes longer and longer.Routing is the longest time-consuming phases in the back-end design process.If we can predict the quality of the design and find the design problems in advance,then the problematic design can no longer be routed,and the design can be modified in time,which can shorten the back-end design cycle,and avoid unnecessary resource consumption.Since short violation is one of the most common problems after routing,in this thesis,short violation one of the common problems in backend design is chosen,and a short violation prediction methods based on an actual engineering projects is designed to predict the distribution of short violations in advance at the placement stage.In order to achieve the goal of short violation prediction better,in this thesis,the chip is divided into partitions,the problem of predicting the short violation distribution under the global situation is transformed into predicting the possibility of short violation occurrence in each partition,and the problem is simplified to a binary classification problem of classifying the partitions by judging whether there are short violations in each partition.To solve the classification problem,supervised learning algorithm is used.By looking for the correlation between the information before routing(cell distribution,etc.)and the information after routing(short violation)in the physical design,the information is analyzed and a function is inferred,then the function will be used to predict short violation in new designs.In order to prepare the short violation data required for the training of machine learning models,following works are completed in this thesis: Multi groups of backend design for sub-modules in a DSP chip under advanced process is completed with EDA tools,data from different modules ensure the diversity and versatility of data sources;A flow of feature data extraction and processing is established based on scripting language,and it is integrated into the backend design flow to realize the automatic establishment of datasets;The class imbalance problem of short violation data is solved by combining the data re-sampling algorithm and learning model processing,and finally establishes multiple datasets of training datasets and test datasets.In order to obtain the best short violation prediction model,multiple kinds of classifiers are constructed in this thesis.There are mainly two types of learning models,including neural network models and decision tree-based ensemble learning models.Different datasets and parameter settings is used for different models,considering the problem of data class imbalance.Several short violation prediction models with an accuracy of more than 90% are achieved after training and testing.Among them,the Balanced Bagging model has the best performance with a balance accuracy of 91.04%,and a Matthews correlation coefficient of64.28%.It has an excellent prediction ability.In addition,the entire process of this thesis has high efficiency.For 700,000 gate-level modules,data extraction only takes 41 seconds,while training time is only 31 seconds,which is far less than the actual routing time.Experiments proved that the prediction method in this thesis has achieved better results than similar studies in terms of accuracy and efficiency.