Graph Neural Network Based Initialization for Timing Driven Placement

Authors: Ziyi Ju, Yunqi He, Ping Yu, Hai Zhou, Jia Wang, Fang Yan

Venue: ACM Transactions on Design Automation of Electronic Systems (TODAES), Apr. 2026

Download: ACM Digital Library

Placement quality depends heavily on the starting point handed to the analytical placer, yet conventional initializations ignore timing structure in the netlist. This work uses graph neural networks that learn cell- and net-level embeddings directly from the netlist to produce timing-aware initial placements. Placement initialization is framed as a node-level prediction task on the netlist graph; the learned seed warm-starts the placer, cutting placement iterations and improving post-placement timing over conventional initialization while improving convergence of the analytical placer.