Research On Modeling Method For Statistical Delay Characteristics Of Standard Cells

With the development of integrated circuit,Electronic Design Automation(EDA)has become an important part of chip design.At the same time,the modular design based on the standard cells has gradually replaced the full-custom design based on the semiconductor components,and has become the main form of chip design,along with the expansion of the circuit scale.When the process level develops to 90 nm,the phenomenon of deviation between the actual value and the set value of the circuit parameters in the actual application process due to the influence of external conditions is called On Chip Variation(OCV).The influence of OCV on the electrical characteristics of the circuit especially delay gradually increases with the progress of the process,so the statistical delay of the circuit describing this deviation by statistical method.Statistical delay is not only one of the important attributes of the standard cells,but also the key and difficult point of the standard cell library design.This paper focuses on the study of statistical delay calculation method in the design of standard cell library:(1)This paper discussed the advantages and disadvantages of Monte Carlo method,and proposed a statistical delay calculation method based on Gaussian Process Regression(GPR).Comparing with Monte Carlo method,the proposed calculation method has faster computational speed.The calculation mean relative error is less than 2%and the calculation efficiency is improved by 10～80.(2)From the perspective of standard cell library design,this paper proposed the disadvantage of high resource consumption and long design cycle in the design of standard cell library under multiple PVT conditions.This paper proposed two classical machine learning algorithms which are Support Vector Regression(SVR)and Back Propagation Neural Network(BPNN)to improve the design efficiency of standard cell library.In the calculation of statistical delay under multiple PVT conditions,the model established a connection between voltage,temperature,input transition time and output load with statistical delay.It saves a lot of time when external condition changes.This paper used multiple metrics to evaluate the quality of the model and compared it with other advanced methods.The calculation error of the model is less than 10%,and the calculation efficiency is improved by about 5 orders of magnitude.Experiments show that the method proposed in this paper has fast modeling speed and better accuracy.It is a successful case of combining EDA tool with machine learning methods.Though the results have been verified with mature process datasets at 40 nm,28 nm,and 16 nm,it is still necessary to evaluate and test more advanced processes and attributes of standard cells in order to promote their application.