How Google used machine learning to dramatically improve chip design

Commentary: To get essentially the most out of machine studying, it pays to keep away from overthinking AI. Learn how Google engineers’ had been in a position to make a ML course of take lower than six hours as an alternative of weeks.

Data concept

Picture: cherezoff/Shutterstock

Regardless of the hype, there’s quite a bit that synthetic intelligence (AI) and machine studying (ML) cannot do. Think about the delay Tesla has had rolling out “full self driving” model 9. As founder Elon Musk tweeted on July 3, 2021: “Generalized self-driving is a tough drawback, because it requires fixing a big a part of real-world AI. Did not count on it to be so laborious, however the problem is apparent looking back.” 

Really, it was apparent in foresight, too.

SEE: Robotics within the enterprise (ZDNet/TechRepublic particular function) | Obtain the free PDF model (TechRepublic)

However what really is not apparent is the very best place for an enterprise to put its ML bets, given how hype clouds the truth of the place ML can shine. And but some latest success by Google engineers with reinforcement studying and chip design factors to ideas that may information any enterprise. So what did Google do? 

Block by block

As written up in Nature, the Google engineers took a novel strategy to “floorplanning”:

Chip floorplanning is the engineering job of designing the bodily structure of a pc chip. Regardless of 5 many years of analysis, chip floorplanning has defied automation, requiring months of intense effort by bodily design engineers to supply manufacturable layouts….In below six hours, our methodology mechanically generates chip floorplans which can be superior or akin to these produced by people in all key metrics, together with energy consumption, efficiency and chip space. To realize this, we pose chip floorplanning as a reinforcement studying drawback, and develop an edge-based graph convolutional neural community structure able to studying wealthy and transferable representations of the chip.

So as an alternative of weeks, the method took lower than six hours. That is spectacular, however as Andrew B. Kahng wrote in Nature “crucial revelation in Mirhoseini and colleagues’ paper is perhaps that the authors’ floorplan options have been integrated into the chip designs for Google’s next-generation artificial-intelligence processors.” In different phrases, this wasn’t a science experiment–it is an AI-driven strategy to chip design that’s already paying dividends in manufacturing, and the strategies are additionally being studied by different chip producers to enhance their very own processes. 

The strategy the engineers took is instructive for any firm hoping to get worth from AI. Machines are inclined to trump folks in areas like pattern-matching, the place uncooked computational energy is extra necessary than inventive perception (the place people excel). On this case, the engineers did not provide you with a intelligent algorithm and ship it off to design chips; as an alternative, they pre-trained their agent on a set of 10,000 chip floorplans. Utilizing reinforcement studying, as detailed in Nature, the agent then “learns” from previous success to undertaking future success: “At any given step of floorplanning, the educated agent assesses the ‘state’ of the chip being developed, together with the partial floorplan that it has constructed up to now, after which applies its learnt technique to determine the very best ‘motion’–that’s, the place to put the subsequent macro block.”

For firms hoping to maximise their possibilities of AI success, the same strategy, with stable coaching information and clearly outlined, considerably constrained targets is necessary. Or, as I’ve written earlier than, the very best hope for AI success is definitely by ML, with “tightly outline[d] tasks [that] increase, not supplant, human actors.” 

Disclosure: I work for AWS, however the views expressed herein are mine.

Additionally see

Recent Articles


Related Stories

Leave A Reply

Please enter your comment!
Please enter your name here

Stay on op - Ge the daily news in your inbox