Determine Routing Order of Register File Signals using Machine Learning

Abstract

Every new Register File (RF) project during enablement of routing, there is a need for a common order for trunk and detailed routing of various RF signals (nets) that gives us the best routing. Majority of the enablement time for detailed routing is spent in coming up with these settings, which are mostly repeated in other project/ process, but not fully applicable because of different constraints. Since we know majority of the features of nets the priority is decided based on it, we plan to train a Machine Learning (ML) Model to give us priority of nets to be trunked/ detail routed in place of existing custom routing settings. Current enablement time for routing alone is 3 to 4 months. Given that we see new scenarios every project, we might have to still keep training the model with new cases with manual analysis and setting. Therefore, the target is to reduce the time by as much as possible by training a robust model (for different Machine Learning Algorithm) tuned for multiple parameters, of which ANN showed the best results with 78% test accuracy which is poor but in reality, showed very good results on the test cases.