Talk:AlphaChip

Request to update AlphaChip controversy bullet with investigation outcome

Hi, I represent Anna Goldie and AlphaChip and am requesting an addition to the disambiguation list to include information about Nature's investigation outcome.

Requested changes:

Addition of Nature addendum to AlphaChip disambiguation

Current text:

AlphaChip may refer to: * DEC Alpha, a 64-bit RISC instruction set architecture developed by Digital Equipment Corporation, originally named "Alpha AXP", as well as DEC CPUs using this architecture * AlphaChip (company), a Russian semiconductor design and R&D company established in 1992 * AlphaChip controversy, a scientific integrity dispute regarding Google DeepMind's AI chip design system published in Nature in 2021

Proposed addition (new bullet after the last item):

• Following an 18-month investigation into the controversy, Nature published an addendum in September 2024 upholding the original paper's methodology and findings

Rationale: The current disambiguation page mentions the controversy but not its resolution. Since Nature's investigation outcome is significant and directly related to the controversy bullet point, it provides important context for readers navigating to information about AlphaChip. Ag77777 (talk) 15:53, 30 October 2025 (UTC)

Not done: This is a disambiguation page, which is the primary reason for this being really too much material. As it stood, the disambiguation page had rather long descriptions, which I've taken the liberty of trimming to be closer to how we typically handle disambiguation as well as removing links to articles other than those being disambiguated. This particular page is not the proper venue to litigate or relitigate how this should be presented. Removing the Nature reference should in its own way bring the page back into balance. Sammi Brie (she/her · t · c) 04:08, 31 October 2025 (UTC)

References

Editing for clarity and neutrality

This article suddenly was expanded by based on Wikipedia:Articles for deletion/AlphaChip (controversy) (3rd nomination) and as a result had an undue focus and tone based on the controversy. I've now made a series of edits that attempted to remove some POV-ridden SYNTH, and also to provide simpler navigation and structure. --ZimZalaBim^talk 20:00, 17 July 2026 (UTC)

Agree with many of the simplifications, but disagree about the removal of the statement that other companies have not adopted this method. This is an exact parallel to the argument made by Ricursive and quoted in the paper. They claim it's used around the world, implying lots of smart people choose to use it, so it must be good. The exact opposite argument is that there are lots of smart people who could adopt it but don't, so the benefits must not be obviously superior. Either both should be included or neither should. In my personal opinion both provide information that might be useful to the reader. LouScheffer (talk) 20:56, 17 July 2026 (UTC)

Lead and Background sections

Hi. I would like to propose changes to the article that improve it, especially concerning the level of detail and issues of WP:DUEWEIGHT. However, to begin, I am suggesting a few relatively minor changes to the Lead and Background sections, as follows:

• Please replace the current lead paragraph with a more accurate version, as follows. (You can see the placement of the references in the code.)

−

'''AlphaChip''' is a deep [[reinforcement learning]] method for automated [[chip floorplanning]]. It was developed at Google and is now a portion of the offerings of the spinoff [[Ricursive]]. The basic ideas were introduced in a 2021 paper, which describes an approach to macro placement, a stage of [[chip floorplanning]]. It is based on [[reinforcement learning]] (RL), a [[machine learning method]] in which a system iteratively improves its decisions by optimizing performance-based reward signals.

+

'''AlphaChip''' is an [[open-source]] deep [[reinforcement learning]] method for automated [[chip floorplanning]]. It was developed at [[Google Brain]] and has been used in multiple generations of Google's [[Tensor Processing Unit]]. The method was published in a 2021 ''[[Nature]]'' paper, which describes an approach to macro placement, a stage of [[chip floorplanning]]. It is based on [[reinforcement learning]] (RL), a [[machine learning method]] in which a system iteratively improves its decisions by optimizing performance-based reward signals.

• Please change the phrasing of that last sentence so it is more informative and less "value laden," and remove the link. How about something like this:

Some researchers expressed skepticism over the 2021 paper's claims, relating to methodology, reproducibility and scientific integrity.

• Please replace the second paragraph of the "Background" section with the following, for clarity and accuracy:

−

[[Placement (electronic design automation)|Macro placement]], the step performed by AlphaChip, is the portion of chip design that determines the locations of large circuit components (macros). The number of macros per circuit typically ranges from several to thousands.

+

AlphaChip performs [[Placement (electronic design automation)|Macro placement]], which determines where large circuit components (macros) will be laid out on the chip. Macro placement had resisted automation for over fifty years, and was previously performed by human experts over the course of weeks or months. The number of macros per circuit typically ranges from several to thousands.

Thanks. LovinLifer (talk) 21:12, 24 August 2026 (UTC)

A few comments on these proposed changes. Some are fine, but others seem POV-centric in favor of Ricursive. I've made some of them but have concerns about others.

• Open source seems too much detail for the first sentence. It's covered later for those who care. The total removal of the reference to Ricursive seems odd, as AlphaChip is cited over and over by Ricursive as the inspiration for their company (14 times alone in one interview), and Ricursive was founded by the very same developers.

• For the statement about skepticism, the biggest skepticism (IMO) is over the quality of results. This should be mentioned first and foremost. And the link is helpful to anyone interested in the controversy.

• The statement of macro placement resisting automation, in my mind, is just plain wrong. Note that in the comparative results quoted later in the article, human design can't beat methods that are in many cases decades old. A reference to back up the statement would be very interesting. It also was one of the points not well compared in the orginal article. LouScheffer (talk) 22:04, 25 August 2026 (UTC)

Hi LouScheffer. Thanks for your edits and thoughts. Please allow me to respond to each of your comments:

• It is common for "open-source" to be mentioned in the lead of articles discussing such projects. Since it is a major component that distinguishes this method from other, related methods, it is important that this descriptor be in the lead. See these articles as examples of the use of "open-source" in the lead: Django (web framework), Node.js, and Bootstrap (front-end framework). Please add the phrase "open-source" to the first line in the lead.

• Ricursive is not a spinoff of Google or any other organization. Nor does Google own Ricursive. A spinoff is an entity formed by a parent organization, and usually at least partially owned by that group after the spinoff. That is not the case with Ricursive. (See Corporate spin-off.) Please replace the word "spinoff" with the word "company."

• The sentence that links to the Controversy section uses POV-laden language. The following sentence better reflects the skepticism about AlphaChip among some researchers in a more precise and less value-laden way. Please change the last sentence (second paragraph) of the lead to:

"Some researchers have expressed skepticism about the 2021 paper's claims."

• Please replace the first sentence of the second paragraph in the Background section with the following two sentences. The first sentence reads more smoothly and is more accurate, and the second sentence (rewritten to avoid confusion over the phrase "resisted automation") conveys the fact that, despite much progress, automation in macro placement has been a challenge for many years.

"AlphaChip performs macro placement, which determines where large circuit components (macros) will be laid out on the chip. Automation for macro placement, which was traditionally performed manually by human experts over the course of weeks or months, has been a challenge for researchers since the 1970s.

• There should not be a wikilink to the Controversy section, or any other section, in the lead. The lead should strictly be a summary of the main points of the article.

• Please change "Google" to "Google Brain" in the second sentence of the first paragraph of the lead, and add a wikilink. Since "Google Brain" is more accurate and has its own Wikipedia article, it would be worthwhile to add.

References

Thanks again. LovinLifer (talk) 09:32, 10 September 2026 (UTC)

Started working on these comments.

• Agree this is not a spin-off in the sense that the parent maintains a stake. The more correct term here is spin-out, which I have linked.
• Linking in the lede to other sections is actually the preferred style. From Wikipedia:How to create and manage a good lead section#"Lead to body" (L2B) links

The lead in this essay uses clickable L2B links, a type of internal referencing making the use of the usual external references unnecessary in this lead. They point to the section where regular references would normally be found and show exactly which wordings in the lead are derived from which sections in the body. Click them to jump to the relevant section from which wording in the lead is derived.

If the use of L2B links became part of the Manual of style, we could probably eliminate the use of references in most leads, and provide good documentation for where more precise information can be found in the body of the article. The connection between lead content and article content would become very precise.

• Changed Google to Google Brain as suggested.
• Changed first sentence in second paragraph of background as you suggested. Not sure about second sentence since it implies designers were doing it manually before AlphaChip. I do not doubt this was true at Google Brain, but suspect it is not true in general, given the alternatives in the table later in the article. Needs more thought, IMO.

Will address remaining comments as soon as I can.

LouScheffer (talk) 00:49, 11 September 2026 (UTC)

• Added open source. While technically true, this seems a little misleading. When most people see open source, they think they can then verify the results claimed in the paper by themselves. But in this case you can't do that as the training data is not available. The later release of pre-trained models helps with this somewhat, but to my knowledge independent replication of the results in the paper are still not possible.

LouScheffer (talk) 16:00, 16 September 2026 (UTC)

• Modified the controversy link, reducing from 4 issues to 2. In my opinion, the most important controversy, in the long term, is whether the algorithm is really an improvement over other, existing, state of the art techniques. The academic integrity dispute is also important, but in the long run the fate of the approach will depend on the comparative quality, no matter how good, or poor, or honest, the initial report.

LouScheffer (talk) 16:17, 16 September 2026 (UTC)

Nature paper and Controversy sections

Thank you LouScheffer for your previous edits. I would like to suggest edits to the "Nature paper" and "Controversy" sections.

• In the last sentence of the "2021 Nature paper" section, please add "TPU" before the word "blocks"; please also add "an open-source RISC-V CPU" after "blocks" for clarity and accuracy. The sentence should read:

The paper reported results on five TPU blocks, and an open-source RISC-V CPU, and described the approach as generalizable across chip designs.

• After the above sentence, please add the following sentence to add relevant additional information:

Professor David Pan at University of Austin later reported results on additional open-source designs.

• Please change the title of the section from "Controversy" to "Reception and adoption," which is a more neutral and accurate description of this section. WP:CSECTION discourages the use of "Controversy" as a section title.

• I still do not agree with your wording for the sentence about how the "claims of the 2021 paper have engendered considerable controversy..." The most concise, neutral, and accurate description of the dispute is the following, or something similar: "Some researchers have expressed skepticism about the 2021 paper's claims."

• Please remove the opening sentence and the following paragraph from what is currently the "Controversy" section. Since the article is no longer called "AlphaChip controversy," it does not need to supply a definition of that phrase nor does it need a paragraph-long commentary on the controversy with no supporting references, entering the territory of WP:OR.

• Please replace both paragraphs of the "Internal dispute at Google" section, which is WP:UNDUE, with the following paragraph, which describes the events of 2019 and beyond more succinctly and objectively, based on reliable sources.

In 2019, Satrajit Chatterjee, a Google engineer, attempted to take control of the AlphaChip project, and when he was rebuffed, he began a campaign against the leads of that project. After repeatedly attempting to shut down the project, in 2022, he drafted Stronger Baselines, arguing that simulated annealing outperformed AlphaChip’s RL approach. A Google committee, chaired by Jon Orwant, ruled that Stronger Baselines was not fit for publication, determining that "the claims and conclusions in the draft are not scientifically backed by the experiments." The AlphaChip authors said that they "provided the committee with one-line scripts that generated significantly better RL results than those reported in Stronger Baselines, outperforming their 'stronger' simulated annealing baseline" and that they "still do not know how the Stronger Baselines authors produced the numbers in their paper." In March 2022, Chatterjee's employment was terminated "with cause"; he then sued Google, alleging fraud and scientific misconduct. In May 2024, Google and Chatterjee reached a settlement.

References

Thank you, LovinLifer (talk) 11:47, 21 September 2026 (UTC)

I've been on (and am still on) vacation. But here are replies to a few points.

Including the RISC-V example, and the settling of the court case make perfect sense. I've included them.

The exact comparisons that have been done between AlphaChip and other placement techniques needs more discussion, including the initial RISC-V core and the Pan comparison. I'll work on this.

The court case discussion is hard to keep neutral. In your suggestion, for example, every sentence tilts in the AlphaChip's favor, whereas at least some of the existing text is tilted the other way.

I've tried to make this more neutral.

I think the word "controversy" is entirely appropriate here. Wikipedia guidelines state

Sections or article titles should generally not include the word "controversies".

Note the use of the work generally. They then go on to state

Criticisms and controversies are two distinct concepts, and they should not be commingled. Criticisms are specific appraisals or assessments, whereas controversies are protracted public disputes, with opposing opinions rather than universal disapproval. Thus, sections such as "Criticisms and controversies" are generally inappropriate.

This is absolutely a protracted public dispute (with 10s of external references) and hence explicitly a controversy. As examples, the IEEE uses "bitter disagreement" and the New York Times uses "controversy", both in the headlines of their articles. Likewise, lawsuits are public disputes over differing opinions. Calling this any less than a controversy is whitewashing, in my opinion. Of course others are welcome to comment. LouScheffer (talk) 01:04, 1 October 2026 (UTC)

Latest suggestions

Hi, LouScheffer. Thank you for taking the time to respond to, and sometimes implement, my suggestions.

• "Both parties then publicly accused each other of bad faith, personal animosity, and technical errors."

I checked the NYTimes and Wired references and found no support for this statement. This is an opinion about the event and is not supported in reliable sources. Please remove the statement.

• "The paper also evaluated the method on fewer benchmarks (five) than is common in the field, and showed mixed results across different evaluation goals."

I checked the named sources "TCAD2025" and "MarkovCACM2024," which you supplied to support this statement. The first source does not support it, and the second source has a publisher's "Expression of Concern" that explicitly recommends against citing it.

EXPRESSION OF CONCERN: An investigation is underway regarding the content and transparency of disclosure for this article. As of February 27, 2026, ACM recommends that this article not be cited in the literature until ACM's investigation has concluded and a final decision has been made.

Please remove that sentence.

• "An additional paper also compared directly to only one external placer (DREAMPlace), out of many potential competing placers, and only older examples with few macro blocks."

This statement is an opinion about what are many versus few blocks, and not backed up by a source. In addition, DREAMPlace was widely considered to be a very strong placer, and was co-produced by the author of the paper you cited.

• The following statement is incorrect:

"Additional criticism noted the claimed six-hour runtime bound per circuit example did not account for pre-training. In the described experiments, RL policies were trained on twenty circuit blocks and then evaluated on five additional blocks, but the reported runtime reflected only the evaluation phase."

The pre-training time of 48 hours is listed in the paper, and they then used matched computing resources for the other tools. Please see the Appendix.

Please make the appropriate corrections to these statements. Thanks again. LovinLifer (talk) 03:58, 12 October 2026 (UTC) LovinLifer (talk) 03:58, 12 October 2026 (UTC)

Replies

Started working on these suggestions. First, 5 is definitely fewer examples than typical. It's less (often much less) than half the number of examples from the leading academic placers of the last 20 years. I added their example counts in a footnote so the user can decide for themselves (the numbers were taken from the cited academic papers.) And the results being mixed is directly supported by the table above, which is straight from the 2021 Nature paper. LouScheffer (talk) 21:22, 12 October 2026 (UTC)

For the 'few' vs 'many' macro blocks, I provided numeric support, with a reference. Note that Pan paper referenced used only the 'few' macro blocks when comparing RL to DreamPlace. There are other examples in the paper, with almost a thousand macro blocks, but they are used for a different purpose - comparing RL with force directed cleanup to RL with DreamPlace cleanup. And DreamPlace may be a strong placer, but it's only one, and it's well known that different placers have different strengths, and differing results on different examples. That's why academic place papers compare to a variety of existing techniques. (LouScheffer (talk) 21:58, 12 October 2026 (UTC)

As far as ""Both parties then publicly accused each other of bad faith, personal animosity, and technical errors.", this is well supported by the references provided. From the NYT article alone

Ms. Goldie said that Dr. Chatterjee had asked to manage their project in 2019 and that they had declined. When he later criticized it, she said, he could not substantiate his complaints and ignored the evidence they presented in response.

That's the definition of bad faith. On the Chatterjee side, he accused Google of having a subordinate of Jeff Dean review his paper. That's also bad faith.

For the 'personal animosity' statement, we have these two quotes from the NYT article. First from the AlphaChip side:

“Sat Chatterjee has waged a campaign of misinformation against me and Azalia for over two years now,” Ms. Goldie said in a written statement.

And from the Chatterjee side:

Laurie M. Burgess, Dr. Chatterjee’s lawyer, said it was disappointing that “certain authors of the Nature paper are trying to shut down scientific discussion by defaming and attacking Dr. Chatterjee for simply seeking scientific transparency.”

Finally, the technical disagreements are well and publicly documented, in the paper 'Stronger Baselines' and the reply "That Chip Has Sailed: A Critique of Unfounded Skepticism Around AI for Chip Design".

Overall the sentence is well supported by reliable sources and stands as written. Other neutral parties are welcome to weigh in, of course. LouScheffer (talk) 22:40, 12 October 2026 (UTC)

"Additional criticism noted the claimed six-hour runtime bound per circuit example did not account for pre-training. In the described experiments, RL policies were trained on twenty circuit blocks and then evaluated on five additional blocks, but the reported runtime reflected only the evaluation phase."

This statement appears correct as written. Clearly the 6 hour evaluation time does not include the 48 hours of training time. However, I will change the text to make this clearer. LouScheffer (talk) 23:03, 12 October 2026 (UTC)

Hi, LouScheffer. Again, thanks for your engagement. The main problem remains that these assertions and characterizations are WP:OR and WP:SYNTH, asserting that a paper used fewer benchmarks than is common, framing a paper's comparison to DREAMPlace as "only one external placer" to present it as a point of weakness, speculating without basis that the motivation behind certain statements was "personal animosity," etc.

The second problem is the extensive citing (14 times) of a paper that has received an expression of concern, including as the source for the "pre-training" criticism. My understanding is that such papers should not be cited on Wikipedia.

I agree with you that this discussion would benefit from the involvement of other neutral parties, as I'm not sure we are making progress at this point. I will see what I can do to bring in other impartial voices. Thanks, LovinLifer (talk) 03:02, 14 October 2026 (UTC)

First, 'only five benchmarks.' This is verging on technical debate, but the Nature paper says:

To address the chip floorplanning problem, we developed an RL method capable of generalizing across chips—meaning that it can learn from experience to become both better and faster at placing new chips—allowing chip designers to be assisted by artificial agents with more experience than any human could ever gain.

Training placement policies that generalize across chips is extremely challenging, because it requires learning to optimize the placement of all possible chip netlists onto all possible canvases.

Claiming the method generalizes is a strong claim. Placing five designs, four from the same family (TPUs), is not very strong evidence for this claim, given their own statement about the enormity of the search space. See this chip image website for a small fraction of what has been designed. That's why the academic papers consider a wider set of designs, and why I think the word 'only' is justified. But maybe a stronger reference can be found.

Similarly, for the Pan paper comparing against only DreamPlace, this goes back to what evidence suffices to show a method is superior. In general, you need to compare against all comers, not just one in particular. While DreamPlace is a strong placer, it is certainly not the best in every case. If you can provide a reference showing DreamPlace is the best placer in almost all cases, then comparing only the DreamPlace would be justified. But such a reference does not exist, to my knowledge.

In particular, when comparing placers, the norm is to campare with a variety of alternative placers. For example, aPlace compared against 4 existing placers, ePlace compared against 10, NTUplace3 compared against 7, and RePlAce against 10 (indirectly - for each benchmark, it compared of the best of 10 other placers on that particular benchmark). DreamPlace only compared against rePlAce, as it was intended to be a faster version of RePlAce.

Objections to the comparisons cited share a common subtext. AlphaChip could do what all academic placers do, and provide results on a wider variety of designs and compared to a wider variety of existing placers. This would settle this particular dispute once and for all. Many researchers (myself included) have asked for this comparison, but it has not been forthcoming. This makes people skeptical, since if the claims are true it would surely be in their interest to show that unambiguously. This is not just my view. From the paper "More Details, but Not Enough",

There is no dearth of recognized benchmarks in design now, though, and Google's reluctance to use those benchmarks in the paper trouble Madden.

As far as 'personal animosity', the article is careful not to speculate on motivation (which would indeed be WP:OR). Instead it notes that each side HAS ACCUSED THE OTHER of bad faith and personal animosity. This is (a) independent of the real motivations, which are unknown, and (b) very well documented by neutral and reliable sources (see the NYT article).