MicroEvo: LLM-Guided Framework for Microarchitecture Design Space Exploration

A new framework called MicroEvo has been launched to tackle the difficulties associated with exploring microarchitecture design spaces. Described in a paper available on arXiv (2608.06183), this innovative approach combines large language models (LLMs) with Monte Carlo Tree Search (MCTS) to enhance multi-objective microarchitecture designs. It seeks to address the shortcomings of current methods that engage in blind searches, neglecting microarchitectural dependencies and failing to learn from iterative processes, which can lead to inefficient evaluations and poor Pareto convergence. MicroEvo features LLM-driven evolutionary operators, a Pareto-aware tree policy, an active knowledge accumulation mechanism, and state-aware directives. Experiments reveal that MicroEvo can enhance Pareto-front quality by as much as 36.2% compared to NSGA-II, showcasing its promise for efficient design exploration within constrained simulation budgets.