Machine-Learning-Based Early-Stage Timing Prediction in SoC Physical Design

Timing closure is essential in SoC physical design. In this paper, machine learning models for timing prediction after floorplan are established. In the models, the features are selected and abstracted by analyzing these parameters from gate-level netlist, constraint files, and floorplan files. The classic machine learning algorithms, such as neural network, support vector machine (SVM), and ensemble machine learning, are explored. The corresponding regression models are applied to predict the timing of SoC. The testcase constructed by open source IP core is used to verify the proposed idea. The results show that the hybrid ensemble learning model has the best prediction performance among various learning models evaluated in this paper.