In the reel I promised the paper and the rebuttal. There is more than one rebuttal, and Google answered back, so here is the whole paper trail. Read the first paper and the latest critique if you only have time for two.
The short version
- The claim (Google, 2021): a reinforcement learning agent places the big blocks on a chip (floorplanning) in under six hours, with results as good as or better than human experts. Google says it has used it on three generations of its TPU chips.
- The pushback (2023 onward): researchers at UC San Diego rebuilt the method and found that a well-tuned simulated annealing baseline, an algorithm from the 1980s, beat it while using a fraction of the compute. Others argued the original paper left out details needed to reproduce it.
- Google's answer (2024): the critics did not run the method as described. They skipped pre-training and used far less compute, so of course it did worse. Google also says Nature's own investigation found in its favour.
- Where it stands (Nov 2026): still open. The UC San Diego group updated their study in March 2026 and still found simulated annealing ahead. Google still points to its use inside real TPUs. Nobody outside Google has published a successful reproduction of the Nature results.
1. The original paper (the claim)
*Mirhoseini, Goldie et al., "A graph placement methodology for fast chip design", Nature 594, 207–212 (2021).*
- Paper: doi.org/10.1038/s41586-021-03544-w
- Earlier free preprint of the same work (2020): arXiv:2004.10746
- Code: github.com/google-research/circuit_training
What it says: treat floorplanning as a game. The agent places one macro (a big block such as a memory) at a time, gets a score at the end for wire length, congestion and density, and learns. After pre-training on earlier chips, it produces a placement for a new chip "in under six hours", which the paper compares with weeks of human effort.
2. Google's follow-up: the name AlphaChip (2024)
- Nature addendum, Sept 2024: doi.org/10.1038/s41586-024-08032-5. It names the method AlphaChip and adds method details, including that a commercial tool's placement was used as a starting point.
- DeepMind's blog post: How AlphaChip transformed computer chip design. It says AlphaChip was used in the last three TPU generations and releases a pre-trained checkpoint.
3. Rebuttal one: the UC San Diego reproduction
Cheng, Kahng, Kundu, Wang and Wang, "Assessment of Reinforcement Learning for Macro Placement", ISPD 2023. Updated as "An Updated Assessment of Reinforcement Learning for Macro Placement", IEEE Transactions on CAD (2025), with a new arXiv version in March 2026.
- Paper: doi.org/10.1145/3569052.3578926 (ISPD 2023)
- Free version, latest: arXiv:2302.11014 (v3, 10 March 2026)
- Their open code and benchmarks: github.com/TILOS-AI-Institute/MacroPlacement
What it says: they rebuilt the missing parts of Google's pipeline in the open, ran it on public benchmarks (including a 7 nm design), and measured the results with a commercial place-and-route tool. A stronger, multi-threaded simulated annealing baseline, and human designs, beat it on wire length and on Google's own proxy cost, with far less compute. The updated version also fine-tunes from the checkpoint Google released in 2024 and reaches the same conclusion.
This one matters for another reason: Andrew Kahng, one of its authors, had written Nature's own commentary praising the 2021 paper. In September 2023 he retracted that commentary, saying new information had changed his assessment.
4. Rebuttal two: Markov's meta-analysis
*Igor L. Markov, "Reevaluating Google's Reinforcement Learning for IC Macro Placement", Communications of the ACM 67(12), 2024.* The preprint is titled "The False Dawn".
- Article: doi.org/10.1145/3676845
- Free preprint: arXiv:2306.09633
What it says: it pulls together the reproduction attempts and argues the method trails human designers, simulated annealing and commercial tools, and that the paper withheld inputs needed to check it. It also mentions misconduct allegations made in a lawsuit by a former Google researcher, which was settled in 2024.
Read it knowing two things. Markov co-authored one of the papers his analysis reviews, which Google's reply points out. And in 2026 the ACM attached an expression of concern to the article while it investigates.
5. Google's reply
Goldie, Mirhoseini and Dean, "That Chip Has Sailed: A Critique of Unfounded Skepticism Around AI for Chip Design" (Nov 2024).
- Free: arXiv:2411.10053
What it says: the ISPD paper did not run the method as published. It skipped pre-training, so the agent could not learn from earlier chips, and it used far less compute (20 times fewer RL experience collectors and half as many GPUs). Google also argues that real use in production TPUs is the evidence that counts, and says Nature finished its own investigation in 2024 and found in Google's favour. Nature removed its editor's note in September 2024 without a public explanation.
How to read the fight
The two sides mostly disagree about what a fair test is.
| Question | Critics | |
|---|---|---|
| What baseline? | Human experts and the tools Google used internally | A strong, well-tuned simulated annealing, plus commercial placers |
| What chips? | Google's own TPU blocks, which outsiders cannot see | Public benchmarks anyone can rerun |
| Pre-training? | Essential, the critics left it out | Added in the 2026 update, still behind |
| Proof it works? | Used across three TPU generations | Not reproduced in public, and wins are measured on private data |
Both can be partly true. A method can be useful inside Google's flow and still not beat the best classical algorithm on a fair public test. That is why the reel ends on "we don't know yet".
If you want one more
- Retraction Watch's write-up of Nature's 2023 editor's note and the retracted commentary is a readable summary of the drama: retractionwatch.com.
- CACM's news piece on the 2024 updates: Updates spark uproar.