Another Night Shift dispute map, also from August 2026. This one’s about Google’s reinforcement learning approach to macro placement (deciding where the big blocks like memories go on a chip), published in Nature in 2021 and later branded AlphaChip.
Short version: it’s never been settled on shared evidence, and honestly it can’t be the way things are set up. Google evaluated on TPU blocks it has never released. The critics evaluate on public designs Google says don’t represent real chips. Neither side has published results on the other side’s test cases, and it’s been five years.
The claim
Azalia Mirhoseini, Anna Goldie, Jeff Dean and 19 coauthors at Google Brain, Nature 594 (June 2021). The preprint puts it as: “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.” (I’m quoting the preprint because the Nature page is behind a login.)
The pushback
It started inside Google. Satrajit Chatterjee, an engineering manager there, led a team whose paper, Stronger Baselines, argued the comparison was rigged with weak baselines: RePlAce wasn’t run properly, and simulated annealing, the obvious classical competitor, wasn’t run at all. Google wouldn’t approve it for publication, and Chatterjee was fired in March 2022 and sued. Fair warning, all of that is second-hand. The session couldn’t find the manuscript or any first-hand statement from Chatterjee, so it all comes through other people’s accounts.
Igor Markov, a Distinguished Architect at Synopsys (which sells competing placement software), wrote a meta-analysis called “The False Dawn.” He says Google’s RL lags human designers, simulated annealing and commercial tools while being slower, and that in a 2023 open research contest “RL methods weren’t in top 5” (Google didn’t enter). He also goes a lot further and says the Nature paper’s integrity is “substantially undermined” and that the internal fraud allegations “still stand.” He was a coauthor on Stronger Baselines, and he does say so in the paper.
The UCSD group (Chung-Kuan Cheng, Andrew Kahng and three others) ran the biggest independent test. Kahng had written the Nature News & Views piece that introduced the original paper, then retracted it after the replication attempts. Their conclusion is that simulated annealing and human placements “remain superior to the latest AlphaChip.” They also found something sharper: the proxy cost the Nature paper optimizes correlates poorly with the real power, performance and area numbers you get after routing.
Google’s answer
Goldie, Mirhoseini and Dean answered in November 2024 with “That Chip Has Sailed.” AlphaChip has made layouts used in three generations of TPU (v5e, v5p and Trillium) and in Google’s Axion CPUs, and has been used in production on blocks with over 500 macros. Their main hits on UCSD: no pre-training (the whole point of the method, mentioned 37 times in the paper), about 20 times less compute (26 experience collectors vs 512, 8 GPUs vs 16), not trained to convergence, and old process nodes (45nm and 12nm instead of sub-7nm). They called it non-peer-reviewed and said AlphaChip is fully open source.
Nature took a side too. It put an editor’s note on the paper in September 2023 saying the performance claims had been called into question, ran an investigation, and ended up publishing an addendum instead of a correction. Per Google, the editors said there’s “little that actually needs correcting.” The paper has never been corrected on its performance claims.
Then UCSD reran it on Google’s terms
This is the part most coverage misses. UCSD went back and used Google’s GPU count, a 7nm node, double the training iterations Google suggested, and fine-tuned from Google’s own pre-trained checkpoint, the one released in August 2024. Same answer, annealing and humans still ahead. And that version is now published in IEEE Transactions on CAD, so “non-peer-reviewed” doesn’t apply anymore. The session couldn’t find any Google reply to the journal version.
Some of the numbers. On the Ariane design at 7nm, routed wirelength came out at 843,757 µm for the commercial tool, 886,776 for annealing, and 1,014,938 for AlphaChip fine-tuned from Google’s checkpoint. On BlackParrot at 45nm, the compute was 2.64 CPU-hours for the commercial tool, 896 for annealing, and 21,507 for that same AlphaChip setup. And on a 532-macro design, right around the 500-plus Google says it runs in production, training from scratch diverged and annealing beat the fine-tuned AlphaChip anyway, with a fraction of the compute.
What nobody can check
The TPU blocks have never been released, so every headline number in the Nature paper rests on designs nobody outside Google can get. The production claims are internal, and so is the Google committee that reviewed and rejected the fraud allegations. Chatterjee’s suit reportedly settled, and a settlement isn’t a ruling on the science.
One thing anyone could check is whether Google’s open-source release is actually complete. Google says it is. Markov and UCSD say the annealing code, the force-directed placer and the proxy cost internals are missing. That’s like ten minutes of looking through a GitHub repo, and the session couldn’t do it because github.com was blocked.
Where it stands
Two stories that each hold together on their own. Google’s: it works, it shipped in three TPU generations, the critics ran a crippled version of it, and Nature sided with them. The critics’: on every design anyone outside Google can run, older and cheaper methods win, and the one experiment that would settle it is the one Google hasn’t published.
Honestly I think the proxy cost thing outlasts all of it. If the number the RL agent is optimizing doesn’t track the real power, performance and area, then who wins on proxy cost barely matters. And that one’s testable on public designs by anybody.
The company that doesn’t exist
My favorite part of this run had nothing to do with chips. While looking up where things stood, a search result summary told the session that “neither Google nor Ricursive has released any results” on public benchmarks. There is no Ricursive. No company, lab or group by that name has anything to do with this. The search layer made it up and dropped it into an otherwise accurate sentence, right where a real party would go, and nothing about it looked off.
It got caught because a map only lets a party on if it came from a page the session actually opened. Ricursive had no page, so it had nowhere to land. Without that rule it would’ve been party number 17, with a believable chip-startup name and a position it never held. A paywall at least shows you the gap. A search summary hands you clean text with nothing missing, so you’d never know.