Google is using machine learning to help design its next generation of machine learning chips. The algorithm’s designs are “comparable or superior” to those created by humans, say Google’s engineers, but can be generated much, much faster. According to the tech giant, work that takes months for humans can be accomplished by AI in under six hours.
Google has been working on how to use machine learning to create chips
“Our method has been used in production to design the next generation of Google TPU,” write the authors of the paper, led by Google’s head of ML for Systems, Azalia Mirhoseini.
AI, in other words, is helping accelerate the future of AI development.
In the paper, Google’s engineers note that this work has “major implications” for the chip industry. It should allow companies to more quickly explore the possible architecture space for upcoming designs and more easily customize chips for specific workloads.
The specific task that Google’s algorithms tackled is known as “floorplanning.” This usually requires human designers who work with the aid of computer tools to find the optimal layout on a silicon die for a chip’s sub-systems. These components include things like CPUs, GPUs, and memory cores, which are connected together using tens of kilometers of minuscule wiring. Deciding where to place each component on a die affects the eventual speed and efficiency of the chip. And, given both the scale of chip manufacture and computational cycles, nanometer-changes in placement can end up having huge effects.
Google’s engineers note that designing floor plans takes “months of intense effort” for humans, but, from a machine learning perspective, there is a familiar way to tackle this problem: as a game.
AI has proven
Google’s engineers trained a reinforcement learning algorithm on a dataset of 10,000 chip floor plans of varying quality, some of which had been randomly generated. Each design was tagged with a specific “reward” function based on its success across different metrics like the length of wire required and power usage. The algorithm then used this data to distinguish between good and bad floor plans and generate its own designs in turn.
As we’ve seen when AI systems take on humans at board games, machines don’t necessarily think like humans and often arrive at unexpected solutions to familiar problems. When DeepMind’s AlphaGo played human champion Lee Sedol at Go, this dynamic led to the infamous “
Nothing quite so dramatic happened with Google’s chip-designing algorithm, but its floor plans nevertheless look quite different to those created by a human. Instead of neat rows of components laid out on the die, sub-systems look like they’ve almost been scattered across the silicon at random. An
This paper is noteworthy, particularly because its research is now being used commercially by Google. But it’s far from the only aspect of AI-assisted chip design. Google itself has explored using AI in other parts of the process like “