Google's claims of super-human AI chip layout back under the microscope | Hacker News Reader

Google's claims of super-human AI chip layout back under the microscope

theregister.com

27 pointsLinuxBender10 comments

Looking at the last few years of Google's work in the AI space it seems very clear to me that the company urgently needs to unchain itself from academia and start building and launching real products again. There are entire divisions at the company spending billions of dollars and doing great work only to spend their day arguing about research papers and peer reviews. Meanwhile smarter and more nimble companies are taking this same research and successfully commercializing it without giving a shit about what a random UCSD professor or some "AI ethicist" feels about the whole thing.

I agree. Corporate labs that fail to ship ultimately die a slow death.

Shipping is a different game though. It takes a culture of pragmatism that many academic groups don't have and don't want to have. I am curious to see how fast Google can pivot.

uh, Andrew Kahng is not a "random UCSD professor". And its not just "arguing about research papers and peer reviews" its called Science. What we used to do before fake news.

There has been plenty of breakthrough science happening at Google for the last 25 years. The problems they have solved in information retrieval, distributed computing, data encoding, video compression and delivery, internet infrastructure, web scaling, big data processing, graph algorithms and lots more are nothing short of revolutionary, and they did it all with a clear (and very successful) commercial focus. More recently while the same great research is still happening, the focus has completely shifted away from shipping. The research is important, the thousand+ years worth of academic baggage and its associated politics is not.

Before capital-S science we had dilettantes that did it for pure joy and curiosity -- not some misguided sense of true belief and absolutist, self-serving non-sense.

Andrew Kahng has credentials, numerous published papers, and held various positions that other people deem to be prestigious -- but what has the man done?

To me, academia and Kahng look like middle-managers endlessly jerking each other off, while other people are out actually accomplishing things.

What has this man done? There are probably only a handful of people who have done more to keep Moore's law going than Dr. Kahng. Without the research he and his many students have done, we wouldn't have these wonderful thousand-core chips my friend.

I think this is overindexing on Dr. Kahng (and his accomplishments) and missing the bigger argument, which I'll restate as: if this research has merit, Google should apply it to their silicon design. Then they can have a paper about how ML helped their world class engineering team create an even better chip, save design time, etc. If the research translates to strong business results, it really doesn't matter what an external researcher thinks about it.

While it's possible to slap a label of ML on a mediocre product for marketing purposes, a strong and experienced engineering team generally won't use a new tool if it doesn't make their workflow substantially better.

You're completely right that if it has merit, Google should apply it to their silicon design.

As it turns out, they do! This method has been used in TPU for multiple generations, including the most recent.

I'm curious if there is actually something wrong with writing papers with the goal of winning cloud customers (one of the complaints in the Satrajit lawsuit implies such). I've always worked from the belief that if a cloud company wants to invest in improved software or hardware that benefits me as a customer, that's a good thing that would make me consider them as a business partner.

Statement from the authors of the super-human AI chip method: https://drive.google.com/file/d/1jWUw6rUDcc7fuHu_iGeVDUkBxNJ...

The concerns presented in the Register are way overblown.

Reply on news.ycombinator.com