GoalEvolve: From Handcrafted Algorithm Priors to Goal-Driven Evolution of Physical Design Algorithms
| Source: arXiv AI
Tags: chip design, EDA, hardware architecture, LLM agents, OpenROAD, VLSI
GoalEvolve uses LLM agents to evolve chip physical design algorithms toward multi-objective quality-of-results targets, improving post-route timing by 30.7% and cutting leakage power by 21.2% versus default OpenROAD on ASAP7 benchmarks.
Details
GoalEvolve presents a framework for using LLMs to automatically improve physical design algorithms in chip design flows — the multi-stage optimization pipelines that convert HDL to a manufacturable layout. The problem it targets is subtle: existing program-evolution frameworks for EDA use stage-local objectives, so a change that improves one stage can degrade results downstream. GoalEvolve makes evolution accountable to the full flow's quality-of-results (QoR). Given a multi-objective QoR target region (timing, power, area), GoalEvolve identifies the worst-performing requirement gap, uses stage-resolved checkpoint evidence to pinpoint the responsible pipeline stage, then deploys an LLM-based Teacher agent to narrow down relevant algorithmic decisions and a source code region. Parallel Student agents implement and evaluate hypotheses through full-flow runs. Results across 8 ASAP7 benchmark designs: 30.67% average TNS improvement, 21.18% leakage power reduction, and 9.42% dynamic power reduction versus default OpenROAD. On power-dominant designs, GoalEvolve closes 62.20% of the gap to commercial tool targets. Compared to Codex goal mode under matched budgets, GoalEvolve improves TNS by a further 26.46%. This is one of the more concrete demonstrations of LLM-driven automated EDA improvement, with quantified results on an open-source toolchain (OpenROAD) that practitioners can replicate.