GNN-based TICER for RC Reduction on Large-scale Interconnect

Time Constant Equilibration Reduction (TICER) is a famous RC reduction method. Traditional TICER performs the floating-point operations repeatedly in time constants calculation, which is time-consuming to deal with large-scale interconnect. This paper converts the node elimination speed-up problem into the classification issue by abstracting the RC network as a graph structure. The deep anomaly detection method is adopted to address the data imbalance among network nodes. The proposed GNN-based TICER is validated on benchmark circuits with a size from 100 to 2 million nodes. Experimental results show that our method achieves 7-926X speedup compared to traditional TICER at 3000 to 2 million nodes with less than 1.54% error.