一种基于机器学习的众工艺角延迟预测方法

在不同工艺角下，关键路径呈现显著差异，因此需要进行大量的静态时序分析，从而导致时序分析运行时间较长。与此同时，随着工艺尺寸的缩小，静态时序分析的精度问题变得不容忽视。本文提出一种基于机器学习的适用于众工艺角下的延迟预测方法，考虑工艺、电压和温度对时序的影响，利用基于自注意力Transformer模型对关键路径进行全局聚合编码，预测众工艺角下关键路径的统计延迟。在EPFL基准电路下进行验证，结果表明该方法的平均绝对误差范围为5.8%～9.4%,有良好的预测性能，可以提高时序分析的准确度和效率，进而缩短数字电路设计周期和设计成本。

A Delay Prediction Method of Various PVT Conditions Based on Machine Learning

The critical path exhibits significant variations under different process corners, necessitating extensive static timing analysis. Consequently, timing analysis needs long execution times, and the accuracy of static timing analysis becomes increasingly important as process sizes continue to shrink. A machine learning-based delay prediction method specifically designed for different process corners was proposed in this paper. By considering the impact of Process, Voltage and Temperature (PVT) on timing, the critical path was globally aggregated and encoded by using a self-attention transformer model to predict the statistical delay across various PVT conditions. The verifications using the EPFL reference circuit demonstrate that the average absolute error of the proposed method ranges from 5.8% to 9.4%, showcasing its strong prediction performance. This approach can enhance the accuracy and efficiency of timing analysis, leading to shorter design cycles and reduced costs in digital circuit design.