Routability-driven Global Routing with 3D Congestion Estimation Using a Customized Neural Network

In global routing, one factor that affects routability is the routing congestion, which happens when the number of wires and vias in a region exceeds its capacity. Such congestion may cause Design Rule Violations (DRVs) or incorrect routing solutions that ultimately lead to design failure. To improve routability, we propose a global routing congestion estimation algorithm based on a Convolutional Neural Network (CNN). This algorithm estimates the severity of congestion of designs in the global routing phase based on a design’s placement information only. Placement features are extracted and fed into the proposed network which produces congestion estimation results. The predicted congestion is taken as an input to our proposed UBC-GR, a modified global router based on the state-of-the-art CU-GR. In comparison with CU-GR, this work achieved an average reduction of 15% in routing channel overflow and 3% in the number of vias, without increasing the total wire length. Moreover, UBC-GR produced a routing solution for a previously unroutable design using CU-GR.