How Google used machine learning to dramatically improve chip design

3 years ago 358

Commentary: To get the most out of machine learning, it pays to avoid overthinking AI. Find out how Google engineers' were able to make a ML process take less than six hours instead of weeks.

Despite the hype, there's a lot that artificial intelligence (AI) and machine learning (ML) can't do. Consider the delay Tesla has had rolling out "full self driving" version 9. As founder Elon Musk tweeted on July 3, 2021: "Generalized self-driving is a hard problem, as it requires solving a large part of real-world AI. Didn't expect it to be so hard, but the difficulty is obvious in retrospect."

Actually, it was obvious in foresight, too.

SEE: Robotics in the enterprise (ZDNet/TechRepublic special feature) | Download the free PDF version (TechRepublic)

But what truly isn't obvious is the best place for an enterprise to place its ML bets, given how hype clouds the reality of where ML can shine. And yet some recent success by Google engineers with reinforcement learning and chip design points to principles that can guide any enterprise. So what did Google do?

Block by block

As written up in Nature, the Google engineers took a novel approach to "floorplanning":

Chip floorplanning is the engineering task of designing the physical layout of a computer chip. Despite five decades of research, chip floorplanning has defied automation, requiring months of intense effort by physical design engineers to produce manufacturable layouts....In under six hours, our method automatically generates chip floorplans that are superior or comparable to those produced by humans in all key metrics, including power consumption, performance and chip area. To achieve this, we pose chip floorplanning as a reinforcement learning problem, and develop an edge-based graph convolutional neural network architecture capable of learning rich and transferable representations of the chip.

So instead of weeks, the process took less than six hours. That's impressive, but as Andrew B. Kahng wrote in Nature "the most important revelation in Mirhoseini and colleagues' paper might be that the authors' floorplan solutions have been incorporated into the chip designs for Google's next-generation artificial-intelligence processors." In other words, this wasn't a science experiment–it's an AI-driven approach to chip design that is already paying dividends in production, and the methods are also being studied by other chip manufacturers to improve their own processes.

The approach the engineers took is instructive for any company hoping to get value from AI. Machines tend to trump people in areas like pattern-matching, where raw computational power is more important than creative insight (where humans excel). In this case, the engineers didn't come up with a clever algorithm and send it off to design chips; instead, they pre-trained their agent on a set of 10,000 chip floorplans. Using reinforcement learning, as detailed in Nature, the agent then "learns" from past success to project future success: "At any given step of floorplanning, the trained agent assesses the 'state' of the chip being developed, including the partial floorplan that it has constructed so far, and then applies its learnt strategy to identify the best 'action'–that is, where to place the next macro block."

For companies hoping to maximize their chances of AI success, a similar approach, with solid training data and clearly defined, somewhat constrained objectives is important. Or, as I've written before, the best hope for AI success is actually through ML, with "tightly define[d] projects [that] augment, not supplant, human actors."

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