Effective Heterogeneous Graph Neural Network for Routing Congestion Prediction

Accurate prediction of routing congestion in the placement stage is critical in digital integrated circuit design. We propose an effective heterogeneous graph neural network named HeteroNet to predict global routing congestion from placement results. The structure of HeteroNet facilitates easy and fast adaptation to model a different global router, by using transfer learning. Experimental results show that HeteroNet models achieve higher predictive performance compared to a strong baseline model. When the global router changes, an existing HeteroNet model can be re-trained using only one circuit design, which is 5.9X faster compared with training from scratch.