Abstract
Traditional dynamic voltage drop (IR drop) analysis in integrated circuit design might be computationally expensive and oftentimes relies on late‑stage routing data. These traditional methods can lead to delays in design closure and may increase infrastructure costs due to high resource needs. This document discloses a two‑stage hybrid machine learning method for rapid IR hotspot identification. As a part of the method, an algorithm utilizes a gradient boosting model to capture a bulk IR drop distribution and triggers a neural network for high‑risk instances to predict peak hotspot values. The method incorporates physics‑grounded features (e.g., power switch density for gated designs, electrical stress proxies). Further, the algorithm applies a mass‑preserving mean neutralization algorithm to maintain global statistical integrity while highlighting local hotspots. The disclosed method enables early‑stage IR drop prediction during placement and clock tree synthesis, which achieves speed improvements compared to traditional sign‑off tools while maintaining high hotspot capture accuracy, facilitating proactive design fixes and reducing a risk of late‑stage design iterations.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Pandey, Praphulla and Gupta, Vaibhav, "Rapid IR Hotspot Identification via Hybrid Machine Learning", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10301