Direction-Aware Timing Modeling for Reinforcement-Learning-Based Clock Arrival Time Optimization in Power-Integrity-Aware Physical Design
Year: 2026
Type: article
Abstract: Dynamic IR-drop induced by large-scale synchronous register switching has become a critical power-integrity challenge in modern high-performance ICs. Highly balanced clock networks can trigger simultaneous switching, causing transient current surges, localized voltage drops, timing-margin loss, and potential functional failures. Existing clock-skew and Clock Arrival Time (CAT) optimization methods mitigate this issue by expanding CAT distributions to desynchronize register switching, but recent reinforcement-learning-based approaches often rely on aggregated TNS-only timing observations, which obscure the directional timing impact of CAT perturbations. This work proposes Direction-Aware Timing Modeling (DATM) for reinforcement-learning-based CAT optimization. DATM explicitly separates launch-side and capture-side timing behaviors, resolving the directional ambiguity of TNS-only formulations and enabling the agent to better infer how CAT adjustments affect timing. The proposed framework is evaluated on seven 28nm designs with diverse timing and structural characteristics. Results show that DATM achieves substantially larger CAT-distribution expansion than the evaluated baselines while maintaining stable setup timing. A comprehensive timing-surrogate audit quantifies model fidelity, and CAT-transfer analysis shows that physical implementation preserves CAT ordering while attenuating target amplitude. Industrial validation with ICC2 and RedHawk further demonstrates an average peak-current reduction of 50.57% and dynamic IR-drop hotspot reduction, with limited timing degradation and average total-area and total-power overheads of 9.33% and 8.63%, respectively. These results establish DATM as a practical EDA method for power-integrity-aware physical design.
DOI: 10.1145/3864457
Institutions: Hangzhou Dianzi University, Shanghai University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7236652213
