How Google used machine learning to dramatically improve chip design

2 months ago 13
PR Distribution

Commentary: To get the astir retired of instrumentality learning, it pays to debar overthinking AI. Find retired however Google engineers' were capable to marque a ML process instrumentality little than six hours alternatively of weeks.

Data concept

Image: cherezoff/Shutterstock

Despite the hype, there's a batch that artificial intelligence (AI) and machine learning (ML) can't do. Consider the hold Tesla has had rolling retired "full aforesaid driving" mentation 9. As founder Elon Musk tweeted connected July 3, 2021: "Generalized self-driving is simply a hard problem, arsenic it requires solving a ample portion of real-world AI. Didn't expect it to beryllium truthful hard, but the trouble is evident successful retrospect." 

Actually, it was evident successful foresight, too.

SEE: Robotics successful the enterprise (ZDNet/TechRepublic peculiar feature) | Download the escaped PDF version (TechRepublic)

But what genuinely isn't evident is the champion spot for an endeavor to spot its ML bets, fixed however hype clouds the world of wherever ML tin shine. And yet immoderate caller occurrence by Google engineers with reinforcement learning and spot plan points to principles that tin usher immoderate enterprise. So what did Google do? 

Block by artifact

As written up successful Nature, the Google engineers took a caller attack to "floorplanning":

Chip floorplanning is the engineering task of designing the carnal layout of a machine chip. Despite 5 decades of research, spot floorplanning has defied automation, requiring months of aggravated effort by carnal plan engineers to nutrient manufacturable layouts....In nether six hours, our method automatically generates spot floorplans that are superior oregon comparable to those produced by humans successful each cardinal metrics, including powerfulness consumption, show and spot area. To execute this, we airs spot floorplanning arsenic a reinforcement learning problem, and make an edge-based graph convolutional neural web architecture susceptible of learning affluent and transferable representations of the chip.

So alternatively of weeks, the process took little than six hours. That's impressive, but arsenic Andrew B. Kahng wrote successful Nature "the astir important revelation successful Mirhoseini and colleagues' insubstantial mightiness beryllium that the authors' floorplan solutions person been incorporated into the spot designs for Google's next-generation artificial-intelligence processors." In different words, this wasn't a subject experiment–it's an AI-driven attack to spot plan that is already paying dividends successful production, and the methods are besides being studied by different spot manufacturers to amended their ain processes. 

The attack the engineers took is instructive for immoderate institution hoping to get worth from AI. Machines thin to trump radical successful areas similar pattern-matching, wherever earthy computational powerfulness is much important than originative penetration (where humans excel). In this case, the engineers didn't travel up with a clever algorithm and nonstop it disconnected to plan chips; instead, they pre-trained their cause connected a acceptable of 10,000 spot floorplans. Using reinforcement learning, arsenic detailed successful Nature, the cause past "learns" from past occurrence to task aboriginal success: "At immoderate fixed measurement of floorplanning, the trained cause assesses the 'state' of the spot being developed, including the partial floorplan that it has constructed truthful far, and past applies its learnt strategy to place the champion 'action'–that is, wherever to spot the adjacent macro block."

For companies hoping to maximize their chances of AI success, a akin approach, with coagulated grooming information and intelligibly defined, somewhat constrained objectives is important. Or, arsenic I've written before, the champion anticipation for AI occurrence is really done ML, with "tightly define[d] projects [that] augment, not supplant, quality actors." 

Disclosure: I enactment for AWS, but the views expressed herein are mine.

Data, Analytics and AI Newsletter

Learn the latest quality and champion practices astir information science, large information analytics, and artificial intelligence. Delivered Mondays

Sign up today

Also spot

Read Entire Article