With the continuous advancement of semiconductor technology, the explosive increase in the number of corners to be covered during chip design has imposed severe challenges on timing analysis resource and time cost. Recent studies on machine learning-based cross-corner timing prediction have emerged as a promising paradigm for accelerating multi-corner timing analysis. Those existing approaches predominantly rely on precomputed multi-corner timing analysis results or process technology parameters for model training, but suffer from limited industrial adoption due to the large number of training corners, the complex pre-training flow, and the opaque nature of the black-box architecture. In case of this, we propose an XGBoost-based cross-corner timing analysis framework that requires only a single-corner static timing analysis report as input to achieve high-precision timing prediction with an average mean absolute error of approximately 1.604 ps across corners. The proposed methodology has been validated through commercial projects based on different deep nanometer process nodes (7 nm and below), demonstrating robust performance stability. In the task of critical path identification and optimization, this proposed method achieves near-100% coverage with acceptable increase in additional STA overhead. Furthermore, we employ SHapley Additive exPlanations analysis to interpret the XGBoost model, quantifying key contributors to cross-corner timing variations. This not only enhances the explainability of the black-box model but also provides insights for design optimization and robustness improvement.
Ouyang et al. (Wed,) studied this question.