ARS-Flow 2.0: An enhanced design space exploration flow for accelerator-rich system based on active learning

Abstract

Surrogate model-based design space exploration (DSE) is the mainstream method to search for optimal microarchitecture designs. However, building accurate models for accelerator-rich systems within limited samples is challenging due to their high dimensionality. Additionally, these models often fall into local optima or have difficulty converging. To address these issues, we propose a DSE flow based on active learning, called ARS-Flow. This method features particle-swarm-optimized Gaussian process regression modeling (PSOGPR), a multiobjective genetic algorithm with self-adaptive hyperparameter control (SAMOGA), and a Pareto-region-oriented stochastic resampling method (PRSRS). Using the gem5-SALAM system for evaluation, the proposed method can build more accurate models and find better microarchitecture designs with acceptable runtime costs.