Source-linked AI summary
GoalEvolve: From Handcrafted Algorithm Priors to Goal-Driven Evolution of Physical Design Algorithms
Haixu Liu, Lei Zhou, Yuhao Ren, Yumao Wu, Zhiang Wang
TL;DR
Physical design algorithm evolution must account for final full-flow QoR rather than stage-local gains. GoalEvolve directs source evolution using target gaps and checkpoint evidence, achieving better post-route QoR than default OpenROAD and Codex goal mode under matched budgets.
Problem
Existing physical design evolution must account for final multi-objective full-flow QoR because stage-local improvements may not persist through downstream interactions.
Method
GoalEvolve converts unmet QoR requirements into target gaps, diagnoses responsible stages, and guides bounded source experiments through Teacher–Student agents and persistent mechanism evidence.
Results
GoalEvolve improves post-route QoR over default OpenROAD and achieves better QoR than Codex goal mode under matched full-flow evaluation budgets.
Takeaways & Limitations
Goal-guided source evolution is a practical methodology for improving physical design algorithms across heterogeneous timing–power profiles.
Takeaways & Limitations
Evolution targets are constructed from preliminary commercial-reference-flow experiments and fixed before each run.
Abstract
from arXiv · showhide
Physical design algorithms operate within tightly coupled, multi-stage optimization flows, where stage-local gains may vanish or induce downstream degradation. Existing program-evolution frameworks often rely on stage-local objectives or undifferentiated multi-metric feedback, which neither guarantee better final results nor identify which unmet requirement should guide the next iteration. We present GoalEvolve, a goal-driven framework that makes physical design algorithm evolution accountable for the final quality of results (QoR) of the complete flow. Given a multi-objective QoR target region, GoalEvolve converts unmet requirements into normalized target gaps, identifies the dominant bottleneck, and uses stage-resolved checkpoint evidence to locate the responsible stage. An LLM-based Teacher then narrows the search to a relevant algorithmic decision and source region, while parallel Student agents implement and validate hypotheses through full-flow evaluation. Local effects, optimization debt, and downstream retention are retained as mechanism evidence for subsequent evolution. Across eight ASAP7 designs, GoalEvolve improves post-route TNS by 30.67% on average and reduces leakage and dynamic power by 21.18% and 9.42% versus default OpenROAD. Relative to commercial-tool goals, it closes 62.20% of the normalized power gap on power-dominant designs, surpasses the TNS goals on both timing-dominant designs, and closes 32.48% of the equal-weight timing-power gap on joint designs. Across all three designs evaluated against Codex goal mode under matched budgets, GoalEvolve further improves TNS by 26.46% while reducing leakage and dynamic power by 12.38% and 0.76%, respectively.
1 Introduction
Physical-design decisions interact across tightly coupled stages, so local improvements may disappear or worsen final QoR. GoalEvolve addresses this by steering algorithm evolution from explicit multi-objective final-QoR goals, normalized bottleneck gaps, and stage-resolved evidence.
- Motivation: Local stage improvements can fail to improve final QoR because downstream processing may introduce congestion, routing detours, or power costs.A stage-level decision is valuable only when its benefit persists through the complete flow while preserving design validity.
- Motivation: Algorithm-first development is limited when the desired QoR is explicit but the algorithmic path toward it remains unclear.Stage logs and automated evaluation motivate a shift from proposing mechanisms first toward using observed full-flow outcomes to determine what to address.
- GoalEvolve: GoalEvolve converts unmet multi-objective QoR requirements into normalized target gaps, uses the dominant gap to select the improvement target, and locates the responsible stage with checkpoint evidence.The framework treats final QoR as a control signal for deciding the next evolution step rather than merely as terminal fitness.
- Contributions: GoalEvolve provides a reusable methodology for deriving and revising optimization logic from explicit goals and full-flow evidence across designs.Its contributions formalize target-set reachability over feasible program variants and evaluate candidates by final full-flow QoR rather than stage-local gains.
2 Background and Related Work
Physical design algorithms have progressed through expert-led, task-specific optimization, while LLM-based systems introduce executable program evolution and repository-scale EDA agents. However, existing approaches still rely on predefined spaces, scaffolds, or staged evaluation mechanisms that constrain adaptation.
- Expert-Driven Physical Design: Physical design advances remain predominantly expert-led, spanning analytical models, differentiable operators, repair procedures, routing models, and learning-based sizing methods.Representative systems include RePlAce, DREAMPlace, OpenPhySyn, CUGR, RL-Sizer, and DAGSizer.
- Expert-Driven Physical Design: Task-specific training and predefined sizing spaces mean these optimization advances still require substantial backend design expertise for development and adaptation.This limitation is stated for RL-Sizer and DAGSizer and generalized to the broader advances discussed.
- LLM-Based Program Evolution: LLM-based program evolution closes the loop among search-space exploration, code generation, and executable evaluation through fixed scaffolds, retained program populations, and feedback-based parent selection.FunSearch searches implementations within a fixed scaffold and retains valid island-population variants, while AlphaEvolve edits code and selects parent programs using feedback.
- Repository-Scale EDA Agents: Repository-scale EDA agents extend program evolution to tool development and design-flow optimization by combining component evolution, repository graphs, literature-grounded planning, hard gates, and staged evaluation.Yao et al. evolve global-placement components using quality and diversity, while AuDoPEDA localizes OpenROAD edits and commits non-regressing patches.
3 Our Approach
GoalEvolve makes physical-design algorithm evolution goal-driven by converting unmet QoR requirements into targeted, stage- and source-level experiments. Its evidence loop preserves mechanism outcomes, including local effects, optimization debt, and downstream retention, to guide subsequent evolution.
- Framework overview: GoalEvolve couples an optimization loop that translates remaining target gaps into bounded source experiments with an evidence loop that preserves mechanism knowledge.The remaining goal gap controls which evidence to inspect and which source mechanism to revise.
- Goal-driven diagnosis: GoalEvolve normalizes design-specific QoR target violations and declares a candidate goal-compliant only when all selected QoR targets are satisfied.The formulation uses |TNS| for TNS, baseline values, target bounds, metric weights, and a positive epsilon; overall goal distance is zero exactly when all selected targets are satisfied.
- Goal-driven diagnosis: After each round, GoalEvolve identifies the dominant weighted QoR bottleneck and traces it through checkpoint trajectories to account for stage effects, optimization debt, and downstream loss.The diagnosis links the unmet target to checkpoint and object-level evidence, narrowing the next target to a responsible stage, mechanism, and source implementation.
- Teacher–Student evolution: The Teacher retrieves implementation constraints, published mechanisms, and historical outcomes before prioritizing hypotheses tied to expected stage responses, final QoR effects, and rejection criteria.OpenROAD Cards, Paper Cards, and historical evolutionary records support mechanism selection while avoiding ineffective proposals.
- Teacher–Student evolution: Four controlled Student branches evaluate hypotheses from a common parent: two Explorers try new ideas, an Enhancer strengthens promising mechanisms, and an Integrator combines compatible validated mechanisms.Each branch uses the same evaluation flow for attribution, while coding agents can diagnose implementation failures and revise patches within a fixed repair budget.
- Evidence and memory: The Evolutionary Program Database stores ideas, patches, evaluation records, and feedback statuses to balance exploration and exploitation while suppressing repeated negative proposals.GoalEvolve also retains feasible alternatives from distinct mechanism families and tracks local effects, optimization debt, and final effect retention across checkpoints.
4 Experiments
Across eight ASAP7 designs, GoalEvolve improves post-route QoR relative to OpenROAD and advances toward frozen commercial-tool targets. Controlled comparisons further examine runtime, cross-design transfer, matched-budget performance against Codex goal mode, and evolved JPEG mechanisms.
- Experimental setup: Experiments evaluate GoalEvolve on eight ASAP7 post-placement designs using OpenROAD commit 08f67ee5 and frozen post-route TNS, leakage, and dynamic-power targets.Targets are constructed externally from preliminary commercial-reference-flow experiments and fixed before evolution.
- QoR results: 30.67% average post-route TNS improvement accompanies 21.18% lower leakage and 9.42% lower dynamic power versus the OpenROAD baseline.Ariane, NVDLA-C, and NVDLA-A each improve TNS by at least 55.40%.
- Commercial-tool targets: 62.20% of the normalized total-power gap closes on power-dominant designs, while GoalEvolve surpasses commercial-tool TNS on both timing-dominant designs and closes 32.48% of the joint gap.On timing-dominant designs, average total power is 2.55% lower than the commercial-tool goal; trajectories show non-monotonic timing–power tradeoffs.
- Cross-design transfer: 46.72% average TNS improvement transfers across all seven non-source designs when the final AES-evolved OpenROAD source is replayed without further source evolution.Total power also decreases on six of the seven designs.
- Codex comparison: 26.46% average TNS improvement over Codex goal mode coincides with 12.38% lower leakage and 0.76% lower dynamic power under matched valid formal-evaluation budgets.GoalEvolve achieves better QoR in seven of nine timing–power comparisons.
- Evolved mechanisms: On JPEG, GoalEvolve evolves a stage-aware power-recovery policy combining early power reclaim, bounded leakage recovery with timing guards, and timing recovery with guarded power reduction.The leakage-recovery search caps the main budget at 4,200 moves and permits one or two tail searches of at most 512 candidates each.
5 Conclusion
GoalEvolve reframes physical design algorithm development as goal-directed source evolution under full-flow feedback, using remaining QoR gaps and checkpoint evidence to guide revisions. Experiments show improved post-route QoR over default OpenROAD and progress toward commercial-tool targets across heterogeneous timing–power profiles.
- Framework: GoalEvolve treats physical design algorithm development as goal-directed source evolution evaluated through full-flow feedback.This reframing accounts for interactions across the complete physical-design flow.
- Framework: Remaining QoR gaps identify which stage and source mechanism should be revised.The framework uses unmet quality requirements to focus subsequent algorithm evolution.
- Framework: Checkpoint evidence and mechanism memory prioritize effects that survive downstream interactions.These signals help distinguish local improvements that persist from those lost later in the flow.
- Results: Experiments show improved post-route QoR over default OpenROAD and substantial progress toward commercial-tool targets across heterogeneous timing–power profiles.The reported evaluation spans differing timing- and power-oriented design conditions.