TimeBoost: Accelerating timing analysis engines using XGBoost

Timing analysis is an essential verification method for the design and optimization of integrated circuits, while it also constitutes the cornerstone of the final signoff. However, the increasing complexity of modern designs, driven by aggressive technology scaling, necessitates more efficient timing engines. In this article, we introduce TimeBoost, a novel learning-based approach for performance optimization of gate-level timing analysis. TimeBoost employs GPU-accelerated XGBoost classifiers to evaluate stage complexity and select the optimal timing analysis method, exploiting the available runtime and accuracy trade-off. Unlike prior works that predict timing directly, our approach ensures reliable and accurate timing estimation by leveraging timing analysis methods as prediction targets. Moreover, our models utilize representative electrical, topological, and gate features for each stage, which can be extracted with minimal effort. We assess several Machine Learning (ML) models, including XGBoost, Artificial Neural Networks, and K-Nearest Neighbors, and explore their different hyperparameters. Experimental evaluation on one million stages demonstrates that our ML-based approach achieves a runtime improvement of up to 3.86×over our best available method, imposing only a 3%–4% overhead to timing analysis while maintaining high accuracy. Thus, the proposed models can be integrated into signoff tools or timing-driven iterative optimization flows to expedite timing closure.