LLMs in Chip Physical Design: What Works in 2026 and What Still Does Not

Large language models have moved from RTL code generation into physical design. Research systems now write tool scripts, tune place-and-route flows, analyze timing reports, refine macro placement, and propose ECO actions. Commercial EDA vendors have also introduced copilots and agentic workflow platforms.

The maturity of these technologies varies widely. Reinforcement-learning-based design-space exploration is already used in production. LLM copilots are productized for knowledge retrieval and script generation. Autonomous agents that can reliably complete a commercial RTL-to-GDS flow remain experimental.

This article reviews the state of the field as of July 2026, with a focus on implementation, timing closure, ECO, and signoff.

Executive Summary

The practical dividing line is straightforward: AI can already accelerate bounded tasks with explicit metrics and strong tool feedback. It is not yet a dependable replacement for the engineer responsible for full-flow closure and signoff.

Reinforcement Learning Is Not an LLM, but It Is the Production Baseline

The most mature AI technology in digital implementation is not generative AI. It is reinforcement-learning-driven design-space exploration.

Tools such as Cadence Cerebrus and Synopsys DSO.ai search large combinations of implementation settings, including floorplan choices, synthesis and placement effort, clock optimization, and other flow parameters. The objective is measurable PPA improvement rather than natural-language interaction.

The reported deployment scale is substantial:

• Cadence reported in its first-quarter 2025 results that Cerebrus had supported more than 1,000 tapeouts. Earlier published customer results included a MediaTek deployment that reported a 5% area reduction and more than 6% power reduction on an advanced-node SoC block.
• Synopsys announced 100 commercial DSO.ai tapeouts in early 2023 and reported more than 700 by its fiscal fourth quarter of 2024. Published customer examples include up to 5% die-size reduction and double-digit PPA gains.
• Synopsys positions PrimeClosure as an AI-driven signoff ECO system. Vendor-published results report fewer ECO iterations and improved timing, including combined timing and voltage-drop optimization with RedHawk-SC.

These figures are vendor-reported and are not directly comparable across tools or designs. They nevertheless show a different level of maturity from most LLM-agent research: the optimization systems have been used on commercial tapeouts, while public LLM-agent evaluations are still dominated by open-source flows and limited test designs.

What LLM Agents Add to the Physical-Design Loop

An LLM agent operates at a different layer. Instead of optimizing only a numeric parameter vector, it can interpret a goal, select tools, generate or modify scripts, read logs and reports, and decide what to try next.

A simplified tool-using loop is:

1. Parse the design objective and current constraints.
2. Generate a tool command or modify a flow configuration.
3. Execute the implementation stage.
4. Extract timing, area, power, congestion, and error data.
5. Diagnose the result and select the next action.
6. Stop when the target is met or the experiment budget is exhausted.

This interface is attractive because physical design already produces large volumes of text and structured reports. It is also risky because every action changes the state consumed by later stages. A syntactically valid command is not evidence that the design remains legal, routable, or signoff-clean.

Script Generation and Flow Tuning: Strong Results on OpenROAD

Most public physical-design-agent research uses OpenROAD or OpenLane. Open-source tools make it possible to publish data, execute benchmarks reproducibly, and inspect failures without commercial license or disclosure restrictions.

ORFS-agent

ORFS-agent uses an LLM to modify OpenROAD-flow-scripts configuration and SDC inputs, execute the flow, read the results, and iteratively optimize the next run. It requires no model fine-tuning.

The updated study evaluates six designs across SKY130HD and ASAP7. Its reasoning-model backends reached results comparable to or better than the OR-AutoTuner Bayesian-optimization baseline while using 40% fewer iterations. The paper also shows that model choice materially changes final quality and that retrieval can improve early convergence without necessarily improving the final endpoint.

The important result is not that an LLM always beats numerical optimization. It is that a tool-using model can navigate a mixed space of configuration, constraints, reports, and natural-language objectives with a relatively small experiment budget.

OpenROAD Agent

OpenROAD Agent integrates a Qwen2.5-Coder model with OpenROAD and uses tool feedback to detect and correct generated scripts. The authors report 94% accuracy on their script-generation tasks and released the model, training data, and evaluation infrastructure.

The result demonstrates the value of execution feedback. Documentation retrieval alone does not prevent hallucinated APIs; the agent must run the command, interpret the error, and repair the script. The 94% figure applies to the paper's script-generation benchmark, not to complete physical-design closure.

ChatEDA

ChatEDA was an earlier attempt to use a fine-tuned LLM for task planning and Tcl generation across an OpenLane RTL-to-GDS flow. It showed that domain adaptation and task decomposition can outperform a general model on the evaluated planning tasks.

ChatEDA established an important architectural pattern: a planner should decompose the workflow, while execution and feedback mechanisms constrain the generated actions. Later systems add stronger retrieval, self-correction, and explicit optimization loops to the same basic structure.

Independent Benchmarks Show the Reliability Gap

High task-level accuracy can hide end-to-end failure. Two 2026 evaluations make that distinction explicit.

PDAGENT-BENCH contains 353 physical-design problems covering foundational knowledge, report comprehension, root-cause analysis, script generation, and full-flow implementation. Across 11 models, the authors found that conceptual performance was much stronger than tool-centric execution. Innovus script-generation accuracy was 42.2%, and long-horizon multi-stage reasoning remained a major weakness.

Can AI Agents Really Complete RTL-to-GDS? evaluates multiple agents and foundation models on a commercial-tool PicoRV32 flow. The study reports three practical findings:

• Domain-specific skills improve understanding of individual subtasks but do not guarantee full-flow completion.
• Agents making similar design progress can differ by as much as 141 times in token return on investment.
• Low-level interface mismatches, including version- or mode-dependent Tcl behavior, are a major failure source.

These benchmarks measure a different capability from script synthesis. Physical-design autonomy requires persistent state, stage-aware recovery, version-correct tool interaction, controlled execution, and a signoff-quality completion criterion.

Macro Placement Is the Main Multimodal Research Frontier

Macro placement is a natural target for multimodal agents because text-only models do not directly represent geometry. Recent systems combine structured design data, visual feedback, conventional placers, and multi-agent refinement.

CHIP-MAP

CHIP-MAP converts LEF, library, and constraint information into structured representations and uses specialized analysis to guide iterative macro-placement repair. On its evaluated designs, the authors report an average repair of 61.6% of total negative slack and area reduction of up to 1.5%.

The framework is notable because it does not ask the LLM to invent coordinates from prose alone. Connectivity weighting and standard-cell usability scoring provide explicit physical signals for refinement.

VeoPlace

VeoPlace uses a vision-language model as a high-level guide for an existing placement algorithm. The model proposes placement subregions or constraints, while evolutionary search evaluates the resulting candidates.

An earlier version appeared at an ICML 2025 workshop. The expanded 2026 preprint reports state-of-the-art results against prior learning-based methods on 9 of 10 open-source benchmarks, with peak wirelength reductions above 32%, and reports gains on all eight evaluated DREAMPlace benchmarks.

MAGE

MAGE encodes floorplanning principles such as boundary placement, macro stacking, pin-aware orientation, pocket minimization, and whitespace quality. Ten specialized agents operate across a six-phase pipeline, with visual audits and tournament-style candidate refinement.

The authors report strong post-route timing improvements over academic and anonymized commercial baselines. They also disclose a significant cost boundary: a full tournament averages about ten hours and roughly $310 per design, with downstream P&R runs dominating runtime.

These systems suggest that the most effective role for an LLM or VLM is not replacing the geometric optimizer. It is supplying higher-level structure, interpreting feedback, and coordinating specialized tools.

Timing Closure, ECO, and DRC Are Better-Bounded Agent Tasks

Closure work is a promising application because it has explicit objectives and rich feedback. Timing reports identify violating paths, ECO tools expose bounded actions, and every iteration produces measurable deltas.

Open-LLM-ECO, described in the paper as a retrieve-schedule-reflect framework, uses RAG to ground optimization strategies and Pareto feedback to schedule actions through an ECO tool. Across eight ASAP7 designs, the authors report 10% greater timing improvement than a reinforcement-learning baseline, more than 4x speedup, and comparable quality to their human-expert baseline. These are research results on a limited benchmark set, not evidence of general signoff autonomy.

EvoDRC decomposes a layout into bounded regions, assigns repair agents, checks connectivity and impact, and evolves reusable repair skills from traceable outcomes. On seven designs from the DAC26 DRC Benchmark, the authors report a 73.5% overall reduction in design-rule violations relative to the reported baseline.

Log interpretation is even more accessible because the input is already textual. A UC Berkeley technical report combined rule-based filtering with GPT-4o for synthesis-log analysis. It reported 100% issue detection and an 89% correct-fix rate on its evaluation, while a multi-agent pipeline resolved an average of 93.7% of synthesis issues across six small Verilog designs.

The strongest near-term pattern is therefore bounded autonomy: the agent diagnoses and proposes or executes a limited repair, while deterministic tools verify legality, timing, connectivity, and regression impact.

Commercial Products Are Moving from Copilots to Agents

Commercial EDA development follows three overlapping layers:

1. Optimization AI searches implementation and verification spaces.
2. Generative AI copilots answer product questions, summarize reports, and generate scripts.
3. Agentic systems plan and execute multi-step workflows with human checkpoints.

Cadence reported more than 1,000 Cerebrus tapeouts by the first quarter of 2025. Cadence has also introduced natural-language assistance in its digital design portfolio and is positioning agentic AI as the next layer above established optimization engines.

Synopsys expanded Synopsys.ai Copilot in 2025. In vendor-published early-access results, documentation search and script generation were reduced from hours to minutes, PrimeTime script generation was reported as 10 to 20 times faster, and formal-verification users reported a 35% productivity improvement. These figures describe selected customer workflows and should not be interpreted as full-flow speedups.

Synopsys describes AgentEngineer as an orchestrated multi-agent framework with increasing autonomy and human-in-the-loop checkpoints. The product direction is consistent with the independent benchmark evidence: bounded workflows can be automated first, while high-risk decisions and completion approval remain with engineers.

Why Autonomous RTL-to-GDS Is Still Unreliable

1. Public Benchmarks Are Not Representative of Full Industrial Closure

OpenROAD enables reproducibility but does not reproduce every commercial-tool behavior, advanced-node rule, proprietary design methodology, or large-SoC integration constraint. Results on a small open design cannot be directly extrapolated to a production block.

2. Command Correctness Is Version- and Context-Dependent

Commercial Tcl interfaces vary by tool, release, application mode, loaded design state, and licensed features. A command that exists may still be illegal at the current stage or may silently change the wrong object set.

3. PPA Is a Coupled, Delayed Objective

An apparent timing improvement after placement can create routing congestion, power-integrity risk, hold violations, or DRC failures later. Local reward signals do not reliably predict signoff quality.

4. Errors Compound Across Long Workflows

RTL-to-GDS contains many stateful stages. A small constraint mistake can survive synthesis and placement before appearing as a late timing or implementation failure. Recovery may require revisiting an earlier stage rather than retrying the latest command.

5. CTS and PDN Remain Underrepresented

Clock-tree synthesis and power-delivery planning require geometry, topology, electrical constraints, and methodology-specific judgment. Public generative-agent work is much thinner in these areas than in scripting, placement, or report analysis.

6. Vendor Metrics and Independent Metrics Measure Different Things

Vendor case studies typically report selected productivity or PPA improvements under customer-specific conditions. Academic papers often report benchmark-relative QoR. Independent agent benchmarks measure completion, reliability, and cost. These numbers should not be combined into a single headline speedup.

Practical Adoption Model for Physical-Design Teams

The safest deployments share several properties:

• The agent operates in an isolated, reproducible run directory.
• Tool versions, libraries, constraints, and design inputs are pinned.
• Commands are retrieved from version-correct documentation or validated with tool help.
• Every action and metric is recorded.
• Deterministic checks decide whether a change is accepted.
• The agent has explicit limits on runtime, compute, licenses, and iteration count.
• A human approves irreversible ECO, release, and signoff decisions.

Outlook

Physical-design AI in 2026 is best understood as a layered system rather than a single autonomous designer.

Reinforcement-learning optimization is already in production. Copilots are useful for knowledge access, report interpretation, and script drafting. Tool-using and multimodal agents show credible progress in flow tuning, macro placement, ECO scheduling, and DRC repair. Independent benchmarks still show weak long-horizon reliability, particularly when commercial tools, stateful flows, and version-specific interfaces are involved.

The near-term engineering opportunity is not unattended RTL-to-GDS. It is controlled automation of expensive, repetitive closure loops, with conventional EDA engines providing ground truth and engineers retaining responsibility for methodology and signoff.

References

1. Cadence Q1 2025 financial results
2. Cadence Cerebrus MediaTek results
3. Synopsys DSO.ai reaches 100 commercial tapeouts
4. Synopsys fiscal Q4 2024 prepared remarks
5. ORFS-agent
6. OpenROAD Agent
7. ChatEDA
8. PDAGENT-BENCH
9. Can AI Agents Really Complete RTL-to-GDS?
10. CHIP-MAP
11. VeoPlace
12. MAGE
13. Open-LLM-ECO
14. EvoDRC
15. LLM-Powered EDA Log Analysis for Effective Design Debugging
16. Synopsys expands Synopsys.ai Copilot capabilities
17. Synopsys AgentEngineer overview