MicroEvo cuts chip-search samples 10.6x

MicroEvo made microarchitecture design-space search 10.6 times more sample-efficient than NSGA-II and improved Pareto-front quality by as much as 36.2%, its authors report in an ICCAD 2026 paper with released code. The system couples off-the-shelf language models with Monte Carlo tree search and reusable optimization knowledge; the result remains a research comparison, not production-silicon proof. For chip teams assessing the economics of AI-assisted processor design, the number is a reason to pilot language models as candidate generators while keeping performance, power, and area simulation as the gate.