#What does it mean to design a chip?
A chip in a phone or an AI server contains an enormous number of electronic switches and metal connections on a tiny piece of material. Designers first describe in code and diagrams what it should compute and in what order it should process signals. At this stage the factory does not yet have a layout it can manufacture.
The team must place circuit blocks, route the metal connections, and check that signals arrive on time and that the layout meets the factory’s rules. Electronic design automation (EDA) software supports this work. AI can help search the many options and identify promising candidates more quickly.
Decide what the chip will do
→ Describe the circuit
→ Place the components on the chip
→ Route the metal connections
→ Check speed, power, and manufacturability
→ Manufacture the chip
#Where does AI help?
#1. Finding candidates in a large design space
The positions of large components and the settings of routing tools affect the outcome. A setting that improves speed may increase power consumption or die area. Design teams try to balance higher performance, lower power, and smaller area. Together these measures are called PPA (performance, power, and area).
AI can compare hundreds or thousands of settings and suggest the next one to try. Its contribution is choosing a more informative experiment from earlier results.
#2. Making design reports easier to interpret
After a check, reports flag issues such as congested routing, a late signal, or metal spacing below the factory minimum. An AI agent can read these reports and plan which settings to change. Tools such as OpenROAD MCP let an assistant run OpenROAD and read back its reports and measurements.
A suggested change still has to pass EDA checks and final signoff before production. The AI proposes candidates. Verification tools determine which candidates meet the requirements.
#3. Reusing experience from earlier designs
A placement strategy learned on one design can provide a starting point for another. This is transfer learning. For example, an arrangement that reduced routing congestion may suggest a first candidate for a related chip. Differences in design size, process, or cell library can undermine that transfer, so the new design still needs its own checks.
A glossary for first-time readers
| Term | Meaning |
|---|---|
| RTL | A register-transfer-level description of what the chip will do |
| EDA | Tools that turn a circuit description into a manufacturable layout and check the result |
| PPA | The balance between performance, power, and area |
| Timing | Checking that signals arrive by their deadlines |
| Congestion | Too many wires competing for the available routing space |
| DRC | Design-rule checks for constraints such as minimum width and spacing |
| Signoff | The final checks required to approve a design for fabrication |
#Examples from chip design
#Google AlphaChip learns where to place large blocks
circuit_training learns where to place large circuit blocks through repeated trials. Good positions can shorten connections and reduce congestion. The agent places a block, observes the result, and chooses the next position. It learns from the outcome through reinforcement learning (RL).
#Synopsys DSO.ai and Cadence Cerebrus search tool settings
Commercial systems often search the settings of existing design tools. Synopsys DSO.ai and Cadence Cerebrus are two examples. Designers set goals and constraints. The software runs and compares candidates for the team to review.
#OpenROAD provides an open environment for experiments
OpenROAD makes many stages of chip design available in open tools. Its AutoTuner can compare settings without someone manually changing each one. OpenROAD MCP gives an assistant access to run results and reports.
#Can AI design a chip on its own?
Current systems are especially useful for reducing repetitive experiments and identifying promising directions among many candidates. Production approval still relies on verification tools that check electrical behavior, physical effects, and manufacturing rules.
When a report says that AI “designed a chip,” look for the particular task it accelerated. It might be placement, parameter search, report interpretation, or experiment planning.
#A guide for further research and experiments
The remaining sections provide reference material for readers planning a review or experiment. A first reading can end here.
Distinguishing different claims of self-improvement
The basic loop is change a setting → run EDA → inspect performance, power, area, and rule violations → choose the next setting. Choosing a better candidate from earlier results is adaptive optimization. Learning the choice rule from rewards is RL. Applying experience to a different design is transfer learning.
| Stage | Term used in this review | Meaning |
|---|---|---|
| 0 | Automation | Execute a predefined script or flow. |
| 1 | Parameter search | Sample settings with a grid or random search. |
| 2 | Adaptive optimization | Use previous results to choose a more promising candidate. |
| 3 | Reinforcement learning | Update the choice policy using rewards from actions. |
| 4 | Transfer learning | Use experience from one design to initialize another. |
| 5 | Continual learning | Retain and reuse experience across projects over time. |
| 6 | Agentic EDA | An agent plans, calls tools, and reads reports. |
| 7 | Self-verifying agentic EDA | EDA and physical checks accept or reject each agent action. |
| 8 | Recursive improvement | Experiments improve the search strategy, objective, or tool sequence itself. |
This framework distinguishes what the public evidence demonstrates. The stages do not rank products. Public results have yet to establish a general solution for long-term memory or recursive improvement.
#How is the next candidate selected?
| Method | What is learned | Feedback | Sample efficiency | Generalization | EDA use | Public implementation |
|---|---|---|---|---|---|---|
| Grid / random | No learning | QoR for each trial | Low | None | Sanity checks for a few settings | OpenROAD AutoTuner sweep |
| Bayesian optimization / Gaussian process | Surrogate for the objective | A small number of expensive EDA trials | High | Depends on feature design | Flow tuning, HPO | Ax/Optuna-based AutoTuner |
| Evolutionary / genetic / CMA-ES | Candidate population or distribution | QoR per generation | Medium | Weak to moderate | Nondifferentiable or mixed discrete settings | Nevergrad, Optuna |
| Simulated annealing | Acceptance probability for a move | Change in objective | Medium | None | Placement or combinatorial baseline | Research combining cyclic RL and SA |
| Contextual bandit | Action value for each context | Rapid reward | High | Depends on context quality | Repeated recipe selection | BO + bandit search |
| RL | Policy or value function | Trajectory reward | Usually low | Pretraining may help | Placement, sequential optimization | circuit_training |
| GNN + RL | Netlist or graph representation and policy | Physical QoR | High training cost | Potential for graph transfer | Macro placement, HLS DSE | AlphaChip-related research |
| Surrogate model | Fast QoR predictor | Historical EDA labels | Very high | Vulnerable to distribution shift | Filter expensive runs | CircuitNet-based research |
| Differentiable / gradient placement | Position and density gradients | Analytic surrogate | High | Tool- and assumption-dependent | Global placement | DREAMPlace |
| Meta / transfer / continual learning | Initialization, policy, or memory | Results across designs | Potentially high over time | Central research challenge | Cross-design warm start | Research stage |
| LLM / multi-agent | Planning, tool sequence, and explanations | Reports and tool output | Depends on tool cost | Depends on prompt and tool schema | Flow orchestration | OpenROAD MCP, vendor agents |
A Gaussian process is one common surrogate used in Bayesian optimization. Evolutionary algorithms, including genetic algorithms and CMA-ES, use different ways of changing a candidate population or its sampling distribution. Treating every method simply as “AI” obscures differences in sample efficiency, failure modes, and reproduction costs.
#A closer look at representative systems
#AlphaChip and Google circuit_training: RL for physical placement
circuit_training is Google Research’s open framework for chip floorplanning with distributed deep RL. It represents a large circuit as a graph and chooses a position for each macro in sequence. The state includes macros already placed and their connections. Actions correspond to candidate locations for the next macro. Rewards use physical proxies such as wirelength, congestion, and density. A fast proxy screens candidates before evaluation through physical implementation.
Here, improvement means that the policy uses rewards from earlier placement trajectories to update future decisions (level 3). The repository also reports that pretraining and fine-tuning help placement speed and quality. Reusing an initial policy on a different block is a level 4 claim. The authors’ evidence does not establish an advantage across every company, process design kit (PDK), and block. Read the original study, the independent evaluation, and the 2024 addendum together, paying attention to reproduction conditions and protocols. Nature paper · Public critical evaluation · 2024 addendum
#Synopsys DSO.ai and commercial design-space optimization
DSO.ai is described as using reinforcement learning to explore settings across an implementation flow. It repeatedly evaluates choices in tools such as Fusion Compiler and IC Compiler II. Synopsys reports commercial tapeouts and PPA results. Attribute improvement factors and tapeout counts to the vendor or customer that reported them. Establishing general performance requires comparable conditions and independent benchmarks. Synopsys announcement of 100 commercial tapeouts
#Cadence Cerebrus expands from exploration to multi-block orchestration
Cerebrus began as Intelligent Chip Explorer in 2021, offering AI-based exploration across physical implementation. AI Studio subsequently emphasized subsystems, multiple blocks and users, and transfer. This direction adds elements of level 6 orchestration to levels 2–4 design-space exploration. Public claims leave open whether knowledge accumulates across projects without regressions. Designs still require signoff. Cadence multi-block white paper
#Siemens Fuse and Solido: what self-verification means
In 2026 Siemens announced the Fuse EDA AI Agent and a self-verifying architecture. The agent plans and executes steps, and deterministic EDA engines and physics-based analyses check the results. This design comes closest to level 7 in the framework. Treat phrases such as “learns from every project” as vendor architecture or roadmap claims and check the public benchmarks and which parts of the model actually update.
#Reproducible design experiments with OpenROAD
OpenROAD and OpenROAD-flow-scripts (ORFS) provide an open RTL-to-GDSII flow. AutoTuner uses JSON to specify parameter ranges and Ray Tune to select grid or random search, PBT, HyperOpt/TPE, AxSearch, Optuna with CMA-ES, or Nevergrad evolutionary search. PPA coefficients and METRICS2.1 results supply objectives or rewards. This makes it a useful level 1–2 baseline even without RL.
The 2026 OpenROAD MCP interface lets an AI client interact with OpenROAD/ORFS sessions, command history, metrics, and report images. One possible loop is LLM agent → MCP → ORFS/OpenROAD → metrics and reports → analysis → revised configuration → rerun. MCP is a protocol for connecting tools. Learning, transfer, and safe optimization require additional design. A level 7 setup also needs clear permissions for changing settings, an allowlist for commands, an experiment database, and deterministic signoff.
#Open projects for optimization and reproduction
The directory below covers optimization, EDA machine learning, and automation, alongside the design flows, PDKs, and verification tools needed to evaluate them. It includes established infrastructure as well as work published in 2024–26.
| Role | Project | Why it matters |
|---|---|---|
| RL placement | circuit_training | RL framework in the AlphaChip family |
| RTL-to-GDS | OpenROAD · ORFS | Open physical-design flow for evaluating results |
| Tuning | AutoTuner | Ray-based DSE baseline |
| agent tool bridge | OpenROAD MCP | tool calls, metrics, reports |
| open flow | LibreLane · OpenLane | reproducible ASIC flow |
| differentiable placement | DREAMPlace · AutoDMP | GPU placement / design automation research |
| ML data | CircuitNet | congestion·DRC·IR-drop dataset |
| implementation | iEDA | open EDA implementation |
| analog automation | ALIGN | SPICE netlist to analog layout |
| architecture/RTL | Chipyard · OpenFASoC | design generation/evaluation input |
| compiler/simulation | Yosys · Verilator · Cocotb | synthesis·simulation regression |
| FPGA / layout | nextpnr · VTR · Coriolis | alternative CAD experiments |
| HDLs / generators | Amaranth · Magma · PyMTL3 · LiteX | parameterized design sources |
| PDK / reproducibility | SkyWater PDK · Hammer · Fault | open PDK·flow orchestration·test |
This directory links directly to at least 26 projects. Reproducing a paper takes more than locating its repository. Record the commit, tool version, PDK license, design input, seed, and run logs.
#The manufacturing side of optimization
Optimization in the fab runs on different timescales from optimization during design.
equipment sensor → process measurement / metrology → AI model
→ recipe recommendation or bounded change → wafer processing
→ inspection · yield → model update
Five distinct tasks arise here.
| Task | Question | Example |
|---|---|---|
| Process control / APC | How much should the current process conditions be corrected? | Overlay or critical-dimension feedback |
| Virtual metrology | Can a value be estimated without a direct measurement every time? | Estimating film thickness from sensors |
| Run-to-run control | How should the last wafer or lot change the next recipe? | Drift correction |
| Predictive maintenance / anomaly detection | Can equipment degradation be detected early? | Chamber health or a pump anomaly |
| Digital twin / autonomous process optimization | Can recipes be explored within safe limits? | Bounded experiments with a yield check |
In its 2025 annual report, ASML describes holistic lithography that links computational lithography with wafer metrology and inspection to predict and control the process window. Applied Materials’ AIx/ExtractAI, Lam’s Equipment Intelligence/Fabtex, and KLA’s inspection and data analytics fall under the same broad trend. They generally operate within a customer’s recipe, data, and safety constraints. A more precise description is a closed loop with bounded actions, validated by measurement, inspection, and yield. Lam Equipment Intelligence · KLA process control
#Maturity matrix for this review
| System | Automation | Adaptive | RL | Transfer | Continual | Agentic | Self-verification | Recursive |
|---|---|---|---|---|---|---|---|---|
| Grid/script flow | YES | NO | NO | NO | NO | NO | PARTIAL | NO |
| OpenROAD AutoTuner | YES | YES | NO | NO | NO | NO | PARTIAL | NO |
| Open AlphaChip framework | YES | YES | YES | PARTIAL | UNKNOWN | NO | PARTIAL | NO |
| Synopsys DSO.ai | YES | YES | YES | PARTIAL | UNKNOWN | PARTIAL | PARTIAL | UNKNOWN |
| Cadence Cerebrus / AI Studio | YES | YES | YES | PARTIAL | UNKNOWN | PARTIAL | PARTIAL | UNKNOWN |
| Siemens Fuse agent architecture | YES | PARTIAL | UNKNOWN | UNKNOWN | vendor claim | YES | YES | UNKNOWN |
| OpenROAD MCP + a custom agent | YES | Implementation-dependent | Implementation-dependent | Implementation-dependent | Implementation-dependent | YES | Implementation-dependent | NO by default |
| Fab APC / process control | YES | YES | Some systems | Limited | Limited | Some systems | Measurement gate | No general evidence |
UNKNOWN indicates that public materials do not establish what gets updated, how much is retained, or how the system performs across projects. Vendor announcements document capabilities and intentions. Independent comparisons answer a different question about measured performance.
#A roadmap for reproducible experiments
Start with a small closed loop in which failures can be traced to a specific step.
| Stage | Setup | Goal or metric | Risk | Research contribution |
|---|---|---|---|---|
| 1. BO/AutoTuner baseline | RTL → Yosys → ORFS/OpenROAD → public PDK → Ray Tune | PPA, WNS, DRC, runtime, cost | Search-space leakage, seed variance | A publishable baseline |
| 2. RL optimizer | Graph or state, action mask, reward, replay, physical evaluation | Constraint-satisfying Pareto QoR | Sparse rewards, simulator or proxy gap, compute cost | Test whether a policy beats the baseline |
| 3. Self-verifying agent | LLM agent → OpenROAD MCP → ORFS → experiment DB → signoff gate | Valid action rate, recovery, regression-free QoR | Hallucinated commands, unsafe changes | Independent planning and verification |
In stage 1, fix the PDK, RTL, constraints, seeds, and trial budget, and save every log. In stage 2, compare the RL reward with actual signoff metrics. In stage 3, the agent may interpret results, but EDA signoff determines whether a proposed change is approved. Documenting compute, data, objectives, metrics, and risks consistently helps distinguish algorithmic gains from effects of benchmark selection.
#A minimal experiment database
Append a record for every trial with design_id, rtl_commit, pdk_version, flow_version, config, seed, action, metrics, constraint_status, reports, runtime, compute_cost, reviewer_decision. These data make transfer and continual learning testable. Evidence for memory includes measuring gains on new designs alongside possible regressions on earlier designs. Saving a conversation alone cannot answer either question.
#Questions to ask when reviewing claims
When a system is said to improve chip design on its own, ask the following questions.
- What improved: a policy, hyperparameter, script, physical design, model weight, workflow, memory, or recipe?
- Did feedback come from a proxy, a full EDA flow, deterministic signoff, or a real wafer and measured yield?
- Was learning online or offline? Did a person approve the action?
- Did performance improve on one benchmark, or did it generalize to new designs, blocks, or PDKs?
- Was knowledge retained? How did the authors measure catastrophic regression?
- Which validator rejects bad decisions: timing, DRC, IR drop, electromigration, formal verification, or a physics-based check?
- Can the system change candidate settings only, or can it also change the search strategy, objective, or tool sequence?
An automation or search tool can be valuable even when the evidence leaves these questions open. A claim of recursive self-improvement requires testing them. Building closed loops in which those tests are possible is a practical next step for AI-for-EDA research.
#Official sources and further reading
- Google circuit_training · AlphaChip in Nature · Independent critical evaluation
- Synopsys DSO.ai · Cadence Cerebrus · Siemens Fuse
- OpenROAD · ORFS AutoTuner · OpenROAD MCP
- ASML 2025 Annual Report · Applied Materials technology portfolio · Lam introduction to etch · KLA Annual Report
Interpreting the sources. The paper, official repositories, and documentation provide the primary evidence. Product pages and press releases document vendor claims and features. Performance, transfer, and autonomy claims are attributed where independent verification is lacking. This article provides neither investment advice nor tapeout assurance.