Routing Congestion Prediction with Machine Learning in Physical Synthesis
TimeMonday, July 11th1:30pm - 1:45pm PDT
Location2008, Level 2
DescriptionAccurate prediction of the final routing congestion early in the Physical Synthesis process can help significantly improve overall quality of results. Such capability may have numerous usages such as fast evaluation of floor-planning decisions during design planning or exploration; dynamically adapting later stages of the physical synthesis process to mitigate any expected high congestion; and abandoning long running jobs with very high expected congestion, saving TAT and computing resources. In addition, some Physical Synthesis processes exhibit non-determinism or run-to-run variation, leading designers to make simultaneous runs and select the best run. In such cases, early congestion prediction may be used to select and focus on a limited set of jobs with lower expected congestion, weeding out higher congestion runs & saving computing resources. In this work, we propose a novel machine learning based method that, after the clock tree synthesis step, predicts expected final congestion at the end of the Physical Synthesis process. We performed extensive data cleaning, analyzed a rich set of features, used various feature selection techniques, and explored multiple models using a dataset for high performance processor designs on advanced technology nodes. Our experimental results indicate significantly low error rates, especially with respect to classifying designs by various degrees of congestion. While testing various models, XGBoost has outperformed other models and is able to predict a run’s final routing congestion value using only data from steps after the Clock Tree Synthesis step.