Congestion-aware global routing using deep convolutional generative adversarial networks

The routing stage is one of the most time-consuming steps in System on Chip (SoC) physical design. For large designs, it can take days of effort to find a complete routing solution, and the result directly affects the circuit performance. In this paper, we present a routing strategy that decomposes global routing into three stages, with different objectives associated with each stage. This is in contrast to conventional approaches, which usually use a single global optimization objective for driving the entire process. Furthermore, we propose to use generative adversarial networks (GAN) to predict the congestion heatmap. This deep learning method has been used to successfully improve image recognition results. We adapt its use to global routing by converting data between the router and the image-based model. This model needs only placement and netlist information as input to make the forecast. Our GAN-based congestion estimator produces congestion heatmaps that show good fidelity with actual heatmaps produced by state-of-the-art global routers. Using this heatmap along with our modified routing flow, we achieve comparable global routing quality in terms of the total overflow and wirelength, but the runtime speedup on hard-to-route designs is significant.