Using Machine Learning to Predict Operating Frequency During Placement in FPGA Designs

Abstract

Circuit placement is an NP-hard problem and is considered to be one of the most challenging steps in the FPGA design flow. The goal of this paper is to explore how machine-learning regression models can be used during placement to predict the maximum frequency of operation. Each model uses static features from the circuit netlist, and dynamic features from the current placement, as input. Results obtained using standard benchmarks indicate that ensemble based machine learning models are capable of accurately predicting the maximum frequency of operation with an average error of 1.72%.