YogurtNet: Enhanced machine learning approach for voltage drop prediction

Abstract

Abstract The aim of this study is to design and implement YogurtNet, a machine learning based static voltage drop (IR Drop) prediction system for SoC power networks. The system employs class image processing techniques, unsupervised learning clustering methods and shallow neural network techniques. Important algorithmic components include the Word2Vec algorithm for implementing clustering, which maps instance names into a name coordinate system; the Pix2pix algorithm for transforming 17 channels of raw data into two channels of predicted data; and an in-house developed shallow neural network, YogurtPyramid, which is used to further approximate and optimize the prediction results. Our model is trained on the names, coordinates, power consumption, and resistance of the back-end layout component instances of three mainstream open-source SoCs (RISCY, Zero-riscy, and RISCY-FPU), and successfully predicts the IR Drop data of each on-chip instance. In terms of model execution results, the average execution time is 69.94 seconds and the average MAE value is 5.2784, demonstrating the high accuracy prediction capability of the model.