Machine Learning Framework for Power Delivery Network Modelling

Traditional approach to power delivery network (PDN) design requires significant amount of time and iterations. In this work, we propose a machine learning-based approach to PDN design which significantly speeds up new PDN option evaluation. The quality metrics we use are package and printed circuit board (PCB) inductances and resistances and die bump voltage droop. The proposed AI architecture combines neural networks and regressor trees to predict inductance and resistance, and employs recurrent-neural-networks to predict voltage droop as a function of time. The proposed approach reduces PDN option evaluation time from weeks to minutes with an average prediction accuracy of 94%.