Agile Physical Design Of Chip Driven By Machine Learning

Physical design is a main stage of integrated circuit.The integration level of circuit keeps rising driven by Moore’s Law and has achieved billions of transistors per chip.The technology node also has entered the nanoscale and caused a lot of problems,which made the complexity of physical design rise rapidly.What’s more,physical design in advanced technology node needs domain knowledge and rich experience,almost all the designers have to do expensive iterations to satisfy the design goals.Traditional method increasing human and resource to accelerate physical design process can’t catch up the trend of physical design because of all above factors.Therefore,finding agile design methods that can make physical design iterate quickly has become an important way to deal with the current situation.Nowadays,machine learning has been widely used in various fields.So,applying machine learning techniques into integrated circuit physical design can learn the historical data of physical design to build empirical models,thus changing the traditional exploratory methods into predictive methods,which can make it possible to quickly explore large and complex design space to achieve the purpose of efficient convergence solution.This article examines machine learning-driven agile physics design methods and uses them in industrial designs,focusing on the following efforts and innovations:1.At present,there is no unified method of accelerating physical design by using machine learning.Our paper studies the characteristics of various stages of physical design and the application of machine learning.Then establishes the machine learning-driven chip agile design methodology.Firstly,we analyze the modeling space and summarize the important features of some stages of physical design.Then,we change the main elements of machine learning into the three stages of data acquisition,feature selection,model building and application.Particularly,the methodology provides details of data collection method,feature ranking and selection mechanism,model building and optimizing method,and the way of design space exploration.2.Considering that the quality estimation in the floorplan stage of physical design is a time-consuming process,our agile physical design methodology is used to predict the post-routing slack of memories from the early stage of the Floorplan.Firstly,we analyze the features in floorplan stage that affect post-routing timing results of Static Random-Access Memory(SRAM).Secondly,a model-based method is used to rank the important features and select the most important feature collection,which can not only give the designer optimization guidance,but also simplify the model and enhance the generalization.Then we use the stacking method to build the optimization model,reducing the floorplan process from 8 hours to 1 hour within the mean absolute error of23.03 ps,greatly reducing the iteration time of chip design.Finally,based on the validated model,a design space exploration is completed to explore 1849 design configurations.The model can improve the timing result over 75.5ps(117%)in average within 22.11 s,which is a great engineering application.3.Considering the multi corner analysis in the static timing analysis(STA: Static Timing Analysis)stage of physical design is a time-consuming process,this paper uses agile physical design methodology to use timing results of some corners to predict those of the remaining corners.Firstly,the correlations between multi corners are analyzed.Secondly,the relationship between corners is extracted by linear and nonlinear models.Then the strong linear relationship between corners is proved.Especially,an industrial design is used to evaluate the generalization of trained model and the effectiveness of Corner selection method,proving the applied prospect of them.Finally,based on the verified robustness of the model,the constraint-driven corner prediction selection strategy is proposed.Taking the constraint condition within a mean absolute error of 1ps as an example,the results of timing analysis of the remaining 7 corners can be predicted by the known 7corners,thus improving time efficiency of STA process by 2x.