AlphaChip RL chip placement

MAP-002 — Does RL macro placement beat existing methods?

A map of the dispute over Google's reinforcement-learning approach to chip macro placement, published in Nature in 2021 and later branded AlphaChip.

The short version: the central question has never been settled on shared evidence, and it structurally cannot be. The designs Google evaluated on are proprietary TPU blocks that have never been released. The critics' evaluations run on public benchmarks Google says are unrepresentative. Neither side has published results on the other side's test cases. Every subsequent round — Nature's investigation, the addendum, the rebuttals — has argued about methodology around that hole rather than closing it.

The original claim

Azalia Mirhoseini, Anna Goldie, Jeff Dean and 19 co-authors, then at Google Brain, published "A graph placement methodology for fast chip design" in Nature 594, 207–212 (June 2021). The 2020 arXiv preprint states the claim:

"in under 6 hours, our method can generate placements that are superhuman or comparable on modern accelerator netlists, whereas existing baselines require human experts in the loop and take several weeks."

— Chip Placement with Deep Reinforcement Learning, arXiv:2004.10746, abstract, final sentence.

The Nature version is widely quoted as: "In under six hours, our method automatically generates chip floorplans that are superior or comparable to those produced by humans in all key metrics, including power consumption, performance and chip area." I could not read the Nature abstract first-hand — see Notes on this run — so the preprint sentence above is the one I am standing behind, and I used the preprint deliberately because the journal version is behind a login wall.

The evaluation was on TPU blocks. Those netlists have never been released, in original or sanitized form. That single fact is the load-bearing element of everything below.

The critics

Satrajit Chatterjee and the internal team ("Stronger Baselines")

The first challenge came from inside Google. Satrajit Chatterjee, an engineering manager in Google's AI research unit, led a separate team that produced Stronger Baselines for Evaluating Deep Reinforcement Learning in Chip Placement (Bae, Yazdanbakhsh, Kim, Chatterjee, Woo, Markov; March 2022). Its technical claim: the Nature comparison was rigged by weak baselines — RePlAce was not run properly, and simulated annealing, the obvious classical competitor, was not run at all.

Google refused to authorise publication after review by an internal resolution committee; the manuscript was posted anonymously in March 2022. Chatterjee raised fraud concerns internally in October 2020, was disciplined in April 2021 for what Google called "uncredible claims of fraud", pressed again in February 2022, and was fired in March 2022. He sued under California whistleblower protections; Google failed to get the suit dismissed in July 2023.

This position is the least well-sourced on the map. I could locate no public URL for the Stronger Baselines manuscript itself, and every court document and news account of Chatterjee's own words sat on a domain this environment blocks. Everything above is second-hand, taken from Markov's meta-analysis (which cites court filings made "under penalty of perjury") and from search-engine summaries I could not open. Chatterjee has no first-person statement on this map.

Igor Markov — "The False Dawn"

Igor L. Markov, Distinguished Architect in AI & Innovation at Synopsys, formerly Google Search, Meta and the University of Michigan, published a meta-analysis: The False Dawn: Reevaluating Google's Reinforcement Learning for Chip Macro Placement (arXiv:2306.09633, v1 June 2023, v10 September 2024), later carried by Communications of the ACM (October 2024).

His performance claim, from the abstract:

"Our meta-analysis shows how two separate evaluations filled in the gaps and demonstrated that Google RL lags behind (i) human designers, (ii) a well-known algorithm (Simulated Annealing), and (iii) generally-available commercial software, while being slower; and in a 2023 open research contest, RL methods weren't in top 5."

His conclusion section puts it as: "prior methods outperform the methods of [1], [4] in quality and runtime" (§8.1, conclusion 3).

He makes a second, separable and much sharper claim about research integrity, also in the abstract:

"Crosschecked data indicate that the integrity of the Nature paper is substantially undermined owing to errors in conduct, analysis and reporting. Before publishing, Google rebuffed internal allegations of fraud, which still stand."

And a third, on reproducibility: "Google released no proprietary TPU chip design blocks used in [1] (nor sanitized equivalents), ruling out full external reproduction" (§5).

On the contest: 19 participants, 8 finalists, "Google did not take part," and "according to the participants' own descriptions, top six teams used traditional analytical optimization methods sans ML."

Stake. Markov works for Synopsys, which sells competing commercial place-and-route software — a direct competition stake. His paper carries the disclaimer "opinions given here are the author's and do not represent his employers." He also discloses that he co-authored Stronger Baselines, one of the two evaluations his meta-analysis relies on: "It is unclear why Google did not allow publishing [6] (coauthored by the author of this meta-analysis)." This matters because Google's rebuttal accuses him of not disclosing it. On the face of the arXiv v10 text, the disclosure is there.

Cheng, Kahng and the UCSD / TILOS group

Chung-Kuan Cheng, Andrew B. Kahng, Sayak Kundu, Yucheng Wang and Zhiang Wang (UC San Diego) ran the most substantial independent evaluation: Assessment of Reinforcement Learning for Macro Placement (ISPD 2023), then An Updated Assessment of Reinforcement Learning for Macro Placement (arXiv:2302.11014), with all data and flows stated to be in the public MacroPlacement repository — a claim I could not verify, because github.com is blocked here and I never opened that repository.

This is the single most important development since 2024, and it changes the shape of the argument: the updated assessment is now peer-reviewed and published in IEEE Transactions on CAD. The arXiv v3 comments field (revised 10 March 2026) reads "This paper has been accepted and published in IEEE Transactions on CAD". Google's central procedural attack — that the criticism was a "non-peer-reviewed invited paper" — no longer describes the current version of the work.

Their conclusion, §VIII:

"SA and human baselines remain superior to the latest AlphaChip, with statistically significant differences in proxy cost and postRouteOpt PPA metrics, using substantially fewer resources."

A separate and arguably more damaging claim concerns the objective function itself, §VI-C:

"The proxy cost function optimized in the Nature paper shows poor correlation with final post-route PPA metrics...underscoring a fundamental misalignment between the optimization target used in Nature and ultimate design objectives."

Crucially, the updated version was built to answer Google's objections on their own terms: eight V100 GPUs (matching Google's stated equivalence), sub-10nm TSMC 7nm test cases, 400 training iterations (double what Google suggested), and fine-tuning from Google's own pre-trained checkpoint released in August 2024 (their "CT-AC" configuration). Their stated aim: "to thoroughly and conclusively address the criticism in [13]" — [13] being That Chip Has Sailed.

Stake. Kahng was a peer reviewer of the Nature paper and wrote the accompanying Nature News & Views commentary — which he later retracted (21 September 2023), on the stated grounds that "new information about the methods used in the reported paper... has become available since publication, and that has changed the author's assessment of, and conclusions about, the paper's contributions." He also leads major open-source EDA efforts (OpenROAD, TILOS), which is a reputational and funding stake in the open-benchmark framing of the question. The updated paper's acknowledgements thank Cadence and Synopsys for policy changes enabling reproducibility, and Google engineers for clarifications; I found no grant sponsor named in the visible text.

Google's answer

Anna Goldie, Azalia Mirhoseini and Jeff Dean (Google DeepMind / Stanford / Google Research) replied with That Chip Has Sailed: A Critique of Unfounded Skepticism Around AI for Chip Design (arXiv:2411.10053, 15 November 2024).

Their positive claim, from the introduction:

"AlphaChip has also generated superhuman chip layouts used in three generations of TPU...and other chips across Alphabet, and been extended to new areas of chip design by external academics and chipmakers."

Named deployments: TPU v5e, v5p and Trillium, and Google Axion processors. They also state that AlphaChip "has been used in production on blocks with over 500 macros" — which is a direct answer to UCSD's scalability finding, and equally unverifiable from outside.

Their itemised attack on Cheng et al. (§2–3):

1. No pre-training. The critics ran the RL method from scratch, removing its defining ability to learn from prior designs. Google notes pre-training is mentioned 37 times in the Nature paper.
2. 20× fewer compute resources. 26 RL experience collectors versus 512, and 8 GPUs versus 16.
3. Not trained to convergence.
4. Unrepresentative test cases. 45nm and 12nm nodes, versus the sub-7nm processes AlphaChip targets.
5. Inappropriate comparison against commercial autoplacers; a "contrived" ablation of initial placement; a flawed proxy-cost correlation study; and an incorrect claim of validation by Google engineers.

On Markov (§3.4, §4): the allegations are unsupported by evidence and scientifically unbacked; an internal independent committee at Google reviewed and rejected them; and Google says he co-authored one of the "separate evaluations" without disclosing it. They further state Google supplied scripts producing RL results that beat the critics' "stronger" simulated annealing baseline.

On openness (§4.1): "AlphaChip is Fully Open-Source" — the Circuit Training release (January–July 2022, after export compliance review), plus the August 2024 pre-trained checkpoint. They add that an independent team replicated the Nature results, and turn the reproducibility charge around: Cheng et al., they say, will not share the synthesised netlists behind their own "open" test cases.

Markov's counter, which directly contradicts this, is that the release is incomplete: "source code for simulated annealing is still missing, and additionally the Nature results cannot be reproduced without proprietary training data and test data." UCSD's updated paper independently identifies the force-directed placement and proxy-cost code as remaining "blackbox elements".

Nature as a party

Nature is not a neutral venue here; it took a position. It attached an Editor's Note on 20 September 2023 — quoted verbatim in Markov's paper:

"Editor's Note: Readers are alerted that the performance claims in this article have been called into question. The Editors are investigating these concerns, and, if appropriate, editorial action will be taken once this investigation is complete."

The following day it retracted Kahng's News & Views. After a second peer review, it found for the authors. Google quotes the editors' finding:

"the best way forward is to publish an update to the paper in the form of an Addendum (not a 'Correction', as we have established that there is little that actually needs correcting)."

The Addendum was published 26 September 2024 (Nature 634, 8034); the Editor's Note was removed. The 2021 paper has never been retracted or corrected on its performance claims. An Author Correction was issued in April 2022 on an unrelated matter.

The commercial parties

Two commercial actors put statements on the record, and both cut in Google's direction.

MediaTek — an SVP announced in September 2024 that MediaTek had extended AlphaChip to accelerate development of its most advanced chips. I have this only as Google characterises it and as search results describe it; I could not read MediaTek's own announcement.

Synopsys — Shankar Krishnamoorthy, Head of Technology & Product Development, publicly distanced the company from its own employee, telling Jeff Dean that "Igor Markov's comments and writings do not represent Synopsys views or opinions in any way," and that Synopsys is aligned with Dean on the potential of RL for chip design, citing adoption of its DSO.ai products. This is the sharpest stake signal on the map, and it cuts both ways: it undercuts the "Markov speaks for a competitor" reading of his critique, while showing that the EDA vendor with the most to lose from RL commercially is publicly bullish on RL.

What was actually run, and what never was

Actually run and public:

• UCSD/TILOS on Ariane (133 macros), BlackParrot Quad-Core (220), MemPool Group (324), plus scaled Ariane variants at ×2 (266) and ×4 (532), across NanGate45, ASAP7 and GlobalFoundries 12LP. Configurations: CT-Scratch, CT-AC (fine-tuned from Google's August 2024 checkpoint), CT-Ours (own checkpoint), against simulated annealing, a commercial tool, and human placement, scored on true post-route PPA.
• Reported figures include: Ariane-ASAP7 routed wirelength — commercial 843,757 µm, SA 886,776 µm, CT-AC 1,014,938 µm, CT-Scratch 1,026,239 µm. BlackParrot-NG45 proxy cost — SA 0.878, CT-AC 0.970, commercial 1.024. BlackParrot-NG45 CPU-hours — commercial 2.64, SA 896, CT-Scratch 19,068, CT-AC 21,507. CT-Ariane-X4 (532 macros) diverged.
• MLCAD 2023 FPGA macro placement contest: 19 entrants, 8 finalists, top six used non-ML analytical methods (DREAMPlace/SimPL lineage). Google did not enter.

Never run, or never published:

• Google's method on the critics' public benchmarks, reported by Google. The central missing experiment. Google disputes the critics' runs of Google's method but has not published its own runs on those designs.
• Any external party on Google's TPU blocks. Impossible: never released.
• A head-to-head on shared test cases under agreed settings. Never happened, in five years.
• Google's internal 16-GPU baseline results, for direct comparison.
• Independent verification of the >500-macro production claim.
• The Nature paper's own simulated annealing baseline code.

What cannot be checked from public material

This is the part the map exists to state plainly, and it is not a shortfall of the search:

1. The TPU netlists are proprietary. Every headline number in the Nature paper rests on designs no one outside Google can obtain. The 2021 claim is, as published, unfalsifiable from outside.
2. The production deployment claims are unauditable. TPU v5e/v5p/Trillium, Axion, ">500 macros in production" — all internal. So is MediaTek's use.
3. Google's internal independent committee that reviewed and rejected Markov's allegations has published nothing. Neither has the earlier resolution committee that blocked Stronger Baselines.
4. Nature's post-publication review — the second referees, their reports, and the reasoning behind "little that actually needs correcting" — is not public in any form I could read.
5. The fraud allegations are contested assertions in a settled employment dispute, not adjudicated findings. Chatterjee's suit reportedly settled in July 2024; settlement terms are not public, and a settlement is not a ruling on the science.
6. Parts of the Circuit Training release remain closed by both critics' accounts — simulated annealing source, force-directed placement, proxy-cost internals — while Google states the release is complete. This specific factual disagreement is checkable by anyone who can read the repository, and I could not, because github.com is blocked in this environment.

Where it stands

The dispute has not converged; it has hardened into two internally consistent accounts.

Google's is: the method works, it shipped in real silicon three TPU generations running, the critics ran a crippled version of it, and Nature investigated and found for us.

The critics' is: on every design anyone outside Google can actually run, older and cheaper methods win — and the one experiment that would settle it is the one Google has never published.

Two things have shifted since Google's 2024 rebuttal, and both weaken its procedural footing rather than its substance. The UCSD assessment is now peer-reviewed in IEEE TCAD, so "non-peer-reviewed" no longer applies. And the updated assessment was run under Google's own stipulated conditions — its compute, its checkpoint, its node — and still reports SA and human baselines ahead. Google has not, as far as I can find, replied to the TCAD version.

The proxy-cost finding may outlast the whole argument. If the objective the RL agent optimises correlates poorly with post-route PPA, then which placer wins on proxy cost is close to beside the point — and that is a claim testable entirely on public designs, by anyone.