Skip to content
hw.dev
hw.dev/signal/goalevolve-llm-physical-design-algorithm-evolution-openroad-2026
SignalarXiv

GoalEvolve Replaces Handcrafted EDA Heuristics with LLM Agents That Evolve OpenROAD Toward Full-Flow QoR

LLM agents now rewrite OpenROAD's physical design algorithms in response to specific post-route QoR failures, closing 30.67% of TNS shortfall and 21% of leakage gap versus default flow -- making open-source P&R a compounding system rather than a fixed tool.

#eda#ai-hardware#tools
Read Original

Physical design algorithms in OpenROAD have always been handcrafted: humans encode heuristics that optimize a stage-local objective, ship them, and leave the remaining quality gap for the next tape-out cycle. GoalEvolve replaces that with LLM agents that evolve the algorithms themselves toward full-flow quality-of-results targets. When post-route timing does not close, the system identifies which stage incurred the debt, generates candidate algorithm rewrites, and validates them against the complete flow. On eight ASAP7 benchmark designs, the result is 30.67% better post-route TNS and 21.18% lower leakage versus default OpenROAD, with 9.42% dynamic power reduction included.

The gap GoalEvolve closes is between stage-local and flow-level optimization. Traditional EDA placement and routing algorithms cannot account for how gains at one stage compound or wash out downstream -- a placer that minimizes wirelength can create routing congestion that costs timing at post-route. GoalEvolve converts unmet QoR targets into normalized gaps, pinpoints the responsible stage via checkpoint evidence, and feeds that diagnosis to an LLM Teacher that constrains the search to a relevant algorithmic decision and source region. Parallel Student agents implement and validate hypotheses against the full flow. Mechanism evidence -- local effects, optimization debt, downstream retention -- accumulates across iterations so the system does not repeat debt it already incurred.

The practical consequence is that OpenROAD's heuristics are no longer fixed by design. A team running open-source physical design can now close harder designs by evolving the tool itself rather than paying for a commercial P&R license that embeds better heuristics behind a GUI. Open-source physical design was already winning on access; GoalEvolve gives it a compounding improvement loop. The commercial EDA vendors whose P&R differentiation lives in proprietary algorithm quality now have to explain why theirs improve faster than an agent framework running against an open-source flow.

Back to Signal