Machine-learning circuit optimization using quantized prediction functions

Abstract

An EDA tool trains a machine-learning optimization tool using quantized optimization solution (training) data generated by conventional optimization tools. Each training data entry includes an input vector and an associated output vector that have quantized component values respectively determined by associated operating characteristics of initial (non-optimal) and corresponding replacement (optimized) circuit portions, where each initial circuit portion is identified and replaced by the corresponding replacement circuit portion during optimization of an associated target IC design. The stored training data entries are used by the machine-learning optimization tool to generate an efficient (e.g., piecewise-linear) prediction function. During optimization of later IC designs, an input vector is generated for each selected initial circuit portion, a predicted output vector is generated for the input vector using the prediction function, and then the predicted output vector is used to identify a replacement circuit portion for the corresponding selected initial circuit portion using best-fit methodologies.