A Novel Regression-Based Approach to Predict Latency in Networks on Chip

Abstract

Accurate performance evaluation is critical for optimising Networks on Chip (NoCs), but conventional methods often require extensive simulation time. This paper introduces a novel regression-based approach to predict NoC latency, enabling precise and efficient performance estimation. Our method innovatively leverages the $ {L_P} $ LP distance within the objective function to minimise prediction error and incorporates constraints relaxation to identify optimal regions within the design space. The proposed approach is validated through extensive experiments on XY routing as a deterministic algorithm and Duato as a fully adaptive algorithm, as well as on both regular and irregular network topologies with varying numbers of virtual channels. These evaluations demonstrate significant improvements in predictive accuracy, with up to 88% improvement for regular topologies and up to 12.8% improvement for irregular topologies, compared to conventional non-linear regression methods.