基于RNN的标准单元延时预测方法

打开文本图片集

DOI：10.13705/j. issn.1671-6841.2023213

A Standard Cell Delay Prediction Method Based on RNN

YOU Huiqing1，2.3， HUANG Pengcheng1，2， ZHAO Zhenyu1，2，WANG Bin 1，2 ， XIANG Lingyun³ （1. College of Computer Science， National University of Defense Technology， Changsha 410o73，China; 2. Key Laboratory of Advanced Microprocessor Chips and Systems， National University of Defense Technology， Changsha 410073， China ; 3. School of Computer and Communication Engineering， Changsha University of Science and Technology，Changsha ，China）

Abstract： During the iterative optimization timing processfrom the post-routing to the sign-of stage，a significant time-cost issue was incurred due to the repetitive execution of static timing analysis. Therefore，a standard cellfeature extraction algorithm was devised and the standard cell delay prediction problem was modeled. Utilizing the recurrent neural network （RNN）as the foundation，the cell-delay prediction model （C-DPM）was constructed to delve into the nonlinear mapping relationship between standard cellcharacteristics and delay，facilitating rapid prediction of standard cell delay. To assess the delay prediction performance of C-DPM for diferent design modules under various process，voltage，and temperature conditions， experiments were conducted on six different design modules with sub- 30nm process. The experimental results revealed that the maximum average absolute error in delay prediction for C-DPM ranged from 0.519 ps to 1.310ps ，while the minimum average absolute error in delay prediction ranged from O.38O ps to 1.016 ps.This demonstrated that C-DPM could trade off minimal error for a reduction in time overhead，thereby accelerating the efficiency of physical design.

Key words：recurrent neural network； static timing analysis；machine learning；standard cell； delayprediction

0 引言

随着摩尔定律的失效和晶体管尺寸的逐渐缩小，集成电路的物理设计变得越来越复杂。（剩余9597字）