That Chip Has Sailed

https://arxiv.org/abs/2411.10053 with Azalia Mirhoseini and Jeff Dean

AlphaChip was one of the first RL methods deployed to solve a real-world engineering problem, and it has been used to design superhuman chip layouts in three generations of TPUs, datacenter CPUs, and other chips across Alphabet. Its publication in Nature helped pioneer the field of AI for chip design, and this open-source method has been extended and built upon by external academics and chipmakers.?

Even so, a small group of detractors raised doubts about our work. Nature conducted a lengthy investigation and second peer review process, and found entirely in our favor, with the editors concluding “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)”. See Nature Addendum published at the conclusion of this process: https://www.nature.com/articles/s41586-024-08032-5

Despite this, a "meta-analysis" was published in the Nov 2024 issue of CACM repeating the same concerns that Nature had already found to be without merit. (Incidentally, the sole author is an employee at Synopsys, which sells commercial tools that compete with our free, open-source method.) The "meta-analysis" covers two (!) papers, neither of which was peer-reviewed.

1. The first is Cheng et al., which was an invited paper at ISPD 2023. Unfortunately, the authors did not run our method as described in Nature. For example, they did no pre-training, robbing our learning-based method of its ability to learn from prior experience (despite “pre-train” appearing 37 times in our Nature article). They also used 20x less compute and did not train to convergence, preventing our method from fully learning even on the chip design being placed. Nevertheless, in their main data table, this hamstrung version of our method still outperformed RePlAce, which was the state of the art when we published in Nature.

2. The second is an unpublished PDF with no author list, which was actually co-authored by Igor Markov, the sole author of the "meta-analysis" itself, though this was not disclosed anywhere in the “meta-analysis” (Markov refers to it as "[X] et al.", but when I reached out to [X], he did not affirm authorship, so we refer to it as Markov et al. [UPDATE (11/19/2024): The person Markov named as first author has asked not to be named in any discussion related to this matter, so I have removed his name from this post.]). This draft did not meet Google’s bar for publication – in 2022, it was reviewed by an independent committee at Google, which determined that “the claims and conclusions in the draft are not scientifically backed by the experiments” and “as the [AlphaChip] results on their original datasets were independently reproduced, this brought the [Markov et al.] RL results into question”. We provided the committee with one-line scripts that generated significantly better RL results than those reported in Markov et al., outperforming Markov’s “stronger” simulated annealing baseline. We still do not know how Markov and his collaborators produced the numbers in this draft.

At the end of the day, there is a physical reality in this field. Chips with layouts designed by AlphaChip have been manufactured and are running in data centers and devices all over the world. The time for speculating about whether our method could work or should work is over—it does work, and has already been used repeatedly in production to make real chips. The field of AI for chip design, which our method helped to pioneer, is alive and well. Our recent GDM blogpost gives more details on AlphaChip’s deployments, as well as external perspectives on the work (see quotes at the end): https://deepmind.google/discover/blog/how-alphachip-transformed-computer-chip-design/

For more details, please see "That Chip Has Sailed: A Critique of Unfounded Skepticism Around AI for Chip Design": https://arxiv.org/abs/2411.10053