Predicted Congestion Using a Density-based Fast Neural Network Algorithm in Global Routing

As the feature size of devices decreases and the number of transistor interconnect exceeds billions, the runtime of global router becomes a problem in very large-scale integrated circuit design. The congestion prediction in the global routing, which can cost huge time, is one of the most important and challenging problems. In this paper, we propose a density and pins peaks-based fast neural network algorithm (NNA) to predict congestion map. In order to save runtime, traditional prediction methods are replaced by a machine learning method with local density and pins as features. Furthermore, to improve the performance of proposed method, the predictive identification method to specify order of the rip-up and reroute for congested regions is also proposed. Compared with state-of-the-art router, NTHU-Route 2.0, on ISPD08 benchmarks, our method greatly improves the speed of predicting routing congestion information by 1094% in test cases, and saves runtime of the global router.