OhmNet: General Static IR Drop Estimation Neural Network Architecture

Static IR drop analysis is a crucial part in the signoff stage of the IC design flow. Conventional commercial tools are often time-consuming, hindering the acceleration of the industrial design cycle. We propose a novel OhmNet neural network architecture to provide quick and accurate predictions for the limited IR drop mitigation time budget in the IC design period. OhmNet incorporates circuit constraint into the model optimization process. This approach not only facilitates the convergence of the model but also enhances its interpretability. Furthermore, the implementation of the OhmN et architecture is modular. We develop four configurable modules that can be customized to meet specific accuracy or speed requirements, leading to the creation of the OhmNet-23 network. The experimental results show that the network is still sensitive to isolated, small-scale regions with high IR drop. Compared with the ground truth, the average correlation coefficient (CC) and the mean absolute error (MAE) of the estimation result are 92.92% and 2.14 mV, respectively, and the estimation speed is 28.98% faster than the method based on U-Net.