Source-linked AI summary
StateTune: Transforming LLM-Assisted EDA Flow Tuning into a Stateful, Closed-Loop Process
Kunlong Li, Shangshang Yao, Su Zheng, Lingli Wang
TL;DR
EDA flow tuning must search a large, coupled parameter space under expensive full evaluations, while prior LLM-assisted tuners mainly rely on transient context. StateTune introduces typed, evidence-gated persistent memory shared by candidate generation and runtime-aware EHVI promotion, achieving the best final hypervolume across six benchmark blocks. The study also reports ablation, poisoning, transfer, and reproducibility analyses supporting the memory design.
Problem
EDA flow tuning combines tightly coupled QoR-affecting parameters with expensive full evaluations, while prior LLM-assisted systems primarily use transient search context.
Method
StateTune uses typed, evidence-gated persistent optimization memory updated by evaluations and shared by candidate generation and runtime-aware EHVI promotion.
Results
StateTune achieves the best final hypervolume on every block in a six-block matrix, with persistent-memory removal reducing HV by 58.5%.
Takeaways & Limitations
Persistent memory is the primary contributor to StateTune’s reported performance, while EHVI promotion provides a secondary gain.
Takeaways & Limitations
The evaluation covers six blocks across two technology nodes and three designs; broader benchmarks and larger industrial SoC studies would strengthen external validity.
Abstract
from arXiv · showhide
EDA flow parameter tuning is critical for quality-of-results~(QoR), yet the parameter space is large, tightly coupled, and full evaluations are prohibitively expensive. Prior LLM-assisted tuners mainly use the LLM as an external proposer with transient working context; we instead present \textbf{StateTune}, which reformulates LLM-assisted EDA tuning as a closed-loop, state-carrying process. Its optimizer state is a typed, evidence-gated \emph{persistent optimization memory} that is updated by every evaluation and shared between candidate generation and budget allocation. On top of this optimizer state, an expected hypervolume improvement (EHVI)-guided, runtime-aware promotion policy ranks quick-stage candidates by expected Pareto frontier gain per unit of runtime cost. Evaluated on a Cadence industrial flow across six benchmark blocks (two technology nodes \(\times\) three designs), against five baselines including LLM+retrieval-augmented generation (RAG) and preference-based Bayesian optimization (BO) tuners, StateTune achieves the strongest final hypervolume on all six benchmark blocks, showing a stable improvement in frontier quality across the full matrix; it also matches or surpasses the strongest baselines on worst negative slack (WNS), area, and power across the same set. Ablation shows persistent memory is the largest contributor: removing it costs 58.5\% of the hypervolume. Dedicated analyses of evidence-gating sensitivity, memory poisoning, cross-design transfer, and three-seed reproducibility (CV\,\(<\)\,7\% on five of six blocks) further validate the memory design.
1 Introduction
EDA flow tuning is a budgeted multi-objective problem because many coupled parameters affect QoR while full evaluations are expensive. StateTune addresses this by carrying typed, evidence-gated memory through a closed loop that guides both proposal generation and budget allocation.
- Motivation: Many interacting knobs across floorplanning, placement, CTS, and routing can change timing, area, power, or run validity.Full RTL-to-route evaluations take hours, strictly limiting the tuning budget.
- Motivation: Prior numerical tuners leave failure modes, parameter interactions, and successful configuration rationales implicit and discard them after each run.Consequently, later iterations and new runs cannot benefit from those discovered insights.
- StateTune: StateTune reformulates LLM-assisted tuning as a closed-loop, state-carrying process with typed, evidence-gated persistent optimization memory updated by every evaluation.The shared memory supports both candidate generation and budget allocation.
- StateTune: EHVI-guided, runtime-aware promotion ranks candidates by expected Pareto frontier gain per unit of runtime risk.This policy directs expensive evaluation budget toward promising proposals.
- Evaluation: Across two technology nodes and three designs, StateTune records the best WNS, power, and hypervolume on all six benchmark blocks, with best or tied-best area on all six.The comparison includes qEHVI, Optuna-TPE, Random, RankTuner, and CROP.
2 Related Work
Prior work spans budgeted multi-objective optimization, structured LLM memory, and automated EDA tuning, but existing methods generally do not externalize tuning knowledge or share it with downstream budget allocation. StateTune combines typed, evidence-gated memory with a shared proposal-and-promotion workflow.
- Budgeted Multi-Objective Optimization: Multi-objective Bayesian optimization uses acquisition functions such as qEHVI to improve the Pareto frontier, while multi-fidelity methods exploit cheaper proxy evaluations.These approaches target sample efficiency under expensive evaluations.
- Structured Memory for LLM Agents: Structured-memory systems organize LLM knowledge through reflection, indexing, hierarchical storage, or retrieval, but persistent memory remains vulnerable to incorrect or outdated information.In optimization, early diagnoses based on insufficient data can create incorrect rules.
- Structured Memory for LLM Agents: StateTune types and gates memory artifacts, requiring hard rules to obtain corroborating full-valid observations before activation.This design addresses memory quality and poisoning risks.
- EDA Flow Tuning and LLM-Assisted Search: StateTune shares memory between candidate generation and downstream budget allocation, allowing rule or sensitivity updates to affect both proposal quality and promotion ranking.The cited comparison distinguishes this coupling from episodic reflection and note indexing.
- EDA Flow Tuning and LLM-Assisted Search: Earlier EDA tuners advance numerical search through multi-task BO, random embedding, preference-based BO, cross-stage transfer, and LLM-guided retrieval.They generally leave non-numerical failure modes, parameter interactions, and design heuristics unexternalized.
- EDA Flow Tuning and LLM-Assisted Search: Figure 1’s workflow connects knowledge, decision, and execution layers through persistent optimization memory.The decision layer generates and promotes candidates, while the execution layer schedules multi-fidelity evaluations and returns QoR metrics.
3 Method
StateTune uses a persistent, evidence-gated optimizer memory to connect knowledge, candidate generation, and budget allocation in a closed-loop multi-fidelity tuning process. EHVI-guided promotion then allocates expensive evaluations toward candidates with high expected frontier gain relative to runtime risk.
- 3.1 Problem Setting and Overview: StateTune optimizes WNS, area, and power under a runtime budget by using cheap quick-stage evaluations before selective full-stage runs.CTS is the quick-stage endpoint in the experiments, while routing provides the expensive full-stage evaluation.
- 3.2 Persistent Optimization Memory: The knowledge layer initializes and refines a design prior, failure summary, and sensitivity summary from constraints, RTL characteristics, documentation, and completed evaluations.Sensitivity analysis combines rank correlations with threshold-effect detection, while failure analysis categorizes failed runs and associated parameter combinations.
- 3.2 Persistent Optimization Memory: Persistent memory M_t stores typed artifacts and shared state that jointly condition candidate generation and EHVI-guided promotion.Its components include design prior, reusable rules, failure and sensitivity summaries, evaluation history, Pareto status, and a runtime model.
- 3.2 Persistent Optimization Memory: Hard rules require 12 corroborating full-valid observations before activation, limiting premature or poisoned knowledge from influencing later decisions.The number of active hard rules is capped at 8.
- 3.3 Knowledge-Guided Search: The decision layer selects among explore, exploit, diversify, repair, and promote modes using recent progress, frontier diversity, and failure patterns.Repair targets known failure modes, while diversify increases Pareto-frontier spread and exploit refines configurations near the current best.
- 3.3 Knowledge-Guided Search: The candidate buffer is refreshed after a new Pareto-optimal point or hypervolume improvement so subsequent proposals use the updated history and frontier.LLM calls occur when the buffer is depleted, keeping token consumption bounded.
- 3.4 EHVI-Guided, Runtime-Aware Promotion: EHVI promotion ranks candidates by expected Pareto-frontier gain per unit of runtime risk, directing budget toward cost-effective improvements.Across the benchmark set, promoted candidates averaged a 0.96 ± 0.08 fraction in the historical quick-stage top-20; removing promotion reduced the top-region hit rate from 34.29% to 8.57%.
4 Experimental Setup
The evaluation uses a shared execution and accounting framework, a configurable quick-stage cutoff, and defined multi-objective metrics for comparing tuning methods.
- Evaluation protocol: All methods share the same execution stack, quick-to-full protocol, and runtime accounting; candidate proposal strategies are the only varied component.LLM calls occur when the candidate buffer is depleted, typically every 4–8 iterations.
- Baselines: RankTuner and CROP are re-implemented within the evaluation framework because source code targeting the evaluated flow is unavailable.CROP uses the same RAG-EDA retrieval as StateTune, and all methods share identical execution, promotion, and budget-accounting pipelines.
- Quick-stage endpoint: CTS is used as the quick-stage endpoint rather than placement alone because it provides richer timing signals for quick-stage QoR estimates and promotion decisions.The quick-stage cutoff remains configurable.
- Evaluation metrics: Final hypervolume measures the dominated volume of the terminal Pareto frontier relative to a fixed reference point, while best WNS, area, and power are per-objective run extremes.Absolute hypervolume values are not directly comparable across tables because reference points are determined independently for comparison groups.
5 Results
StateTune delivers consistently stronger Pareto-frontier quality across the full six-block benchmark matrix, with persistent memory contributing most to performance and EHVI promotion providing an additional gain.
- StateTune attains the best WNS, power, and final HV on all six benchmark blocks, with best or tied-best area on all six.
- +33.7% over Optuna-TPE on ASAP-JPEG and positive final-HV advantages on the other five blocks establish consistent frontier gains across the matrix.
- CROP, despite combining an LLM with RAG, remains behind StateTune on all six blocks and can trail random search on final HV.
- CV remains below 7% on five of six blocks; ASAP7-JPEG has 15.4% CV, yet its worst single-seed HV still exceeds all baselines.
- Ablation Study: Removing persistent memory drops HV to 4.19 M (−58.5%) with zero Pareto hits, making it the largest single contributor.
- Ablation Study: Removing knowledge-agent guidance drops HV to 4.57 M (−54.8%), while removing RAG drops HV to 5.44 M (−46.1%) with zero Pareto hits.
- Ablation Study: Removing CoT drops HV to 6.92 M (−31.5%), whereas removing EHVI delays the first global Pareto hit from 10.29 h to 12.93 h and drops HV by 17.6%.
- Validation Analyses: HV peaks at 12.31 M with k=12, while ungated memory increases contradicted-rule rate to 17.4% and lowers full-valid rate to 2.31%.
6 Conclusion
StateTune recasts LLM-assisted EDA tuning as a state-carrying closed loop with persistent memory and EHVI-guided promotion. It achieves the best final hypervolume on every benchmark block, with persistent memory the primary contributor and broader validation left for future work.
- StateTune combines typed, evidence-gated persistent memory with EHVI-guided promotion across six benchmark blocks and five baselines.
- StateTune achieves the best final hypervolume on every benchmark block.
- Removing persistent memory reduces HV by 58.5%, while removing EHVI promotion reduces it by 17.6%.
- Future work includes larger industrial designs and richer memory organizations for scaling across more benchmark families.