机器学习驱动的多Corner STA加速方法

Multi-corner STA Acceleration Based on Machine Learning

静态时序分析在芯片的物理设计中扮演着重要的作用,其贯穿设计流程始终,耗时巨大,多数时间用于在不同工艺、电压、温度条件下的多Corner组合分析。论文研究多Corner在时序分析方面的相关性并利用机器学习的方法对其进行建模,以利用已知Corner时序结果预测未知Corner时序结果而不需额外的运行时间。实验结果表明,线性模型可以有效地进行预测,在利用7个已知时序结果的Corner预测余下7个未知时序结果的Corner时,能达到2.07ps(1.76%)的平均绝对误差。同时,基于模型的强稳健性,可以采用约束驱动的Corner选取策略,有效地减少多Corner角组合时序分析所需的时间,加速物理设计流程。

Static Timing Analysis plays a vital role in today’s high-complex circuit physical design and is always used to go through the whole design flow.The timing analysis can take much time,especially the multi-corner analysis in different process,voltage and temperature cases.The correlation between the corners in timing analysis is researched,and the machine learning model is built to predict timing results at unanalyzed corners from certain analyzed corners,without spending extra running time.The experimental result shows that the linear model can make an effective prediction.For example,the timing of 7 corners is used to predict the rest timing of 7 corners and achieve the mean absolute error within 2.07 ps(1.76%).Since the model has good robustness,the constraint-driven strategy can be implemented,significantly reducing time in multi-corner analysis and accelerating the physical design flow.