Early Prediction of Post-Route Physical Design Data Using a Machine Learning Model

Faster and more effective physical design cycles are essential as the semiconductor industry pushes the limits of device complexity further. Physical design phases like place-and-route (PnR) require a lot of computation and time. Traditionally, important design variables like power usage, area utilization, and instance counts are only available to designers late in the cycle. A strategy based on machine learning (ML) is suggested in this work to use synthesis-stage data to anticipate important post-route parameters early in the flow. Cadence Genus was used for synthesis, and Cadence Innovus was used for physical implementation to gather a large dataset. Multitask Neural Networks, Random Forest, LightGBM, and XGBoost were among the machine learning models that were assessed. The results of the experiments show that the neural network model had the best and most consistent prediction accuracy across several parameters. This could lead to a reduction in the total design turnaround time and early visibility into the physical design consequences.