Abstract
Routability in physical design has hit a bottleneck, because congestion estimated by conventional routers does not cope well with modern sophisticated routing parameters. Several techniques have recently been developed to predict routability information through a supervised-learning based mechanism. However, features extracted by such methods are rather primitive for representing actual physical properties. Furthermore, the lack of global information leads to worsened global routing performance. In this paper, we propose a supervised-learning regression model able to capture accurate global routing behaviors, through which a congestion prediction model is trained to improve the global routing. Experimental results show that, in contrast with conventional global-routing based congestion estimation, our predictor is at least 9.33 × faster in execution, while maintaining an accurate prediction. Moreover, by integrating our model into the router, not only a better routing topology is achieved, but also and superior quality of performance is observed.