Google uses machine learning to help design the next generation of machine learning chips. The design of the algorithm “Comparable to or superior” to human-made designs, Google engineers say, but can be built much faster. According to the tech giant, a month-long task for humans can be accomplished by AI in less than six hours.
Google has been working on how to use machine learning to build chips for years. But this latest effort is described in a journal article this week. nature — It appears to be the first time the research has been applied to a commercial product: Google̵7;s upcoming TPU (tensor processing unit) chip, which is optimized for AI computing.
“Our method was used in manufacturing to design the next-generation Google TPU,” the authors of the article, led by Azalia Mirhoseini, head of ML for Systems at Google.
In other words, AI is accelerating the future of AI development.
In this report, Google engineers noted that the work was “significant” for the chip industry. should help companies Explore possible architectural areas for designs to happen faster. and easier to customize the chip for specific workloads
Editorial in nature This research is called He noted that such work could help offset the predicted end of Moore’s Law. This is a chip design axiom from the 1970s that states that the number of transistors on a chip doubles every two years. AI doesn’t have to solve the physical problem of squeezing more transistors on a chip. But it can help find other routes to optimize at the same rate.
The specific task that Google’s algorithm handles is called This often requires a human designer working with the help of computer tools to find the optimal layout on silicon molds for the chip subsystem. These components include things like the CPU, GPU, and memory cores. which are connected together using small wiring tens of kilometers Deciding where to place each component on the mold ultimately affects the speed and efficiency of the chip. And with both the size of the chip production and the computation cycle. The change in position at the nanometer scale can have a big impact.
Google engineers noted that the floor plan design uses “Months of hard effort” for humans, but from a machine learning point of view There is a familiar way to fix this: it’s a game.
Time and again AI has proven to be superior to humans in board games like Chess and Go, and Google engineers have noticed that floor plans are similar to those challenges. instead of a board game Do you have a silicone mold? instead of a piece like a knight and a gangster You have components like CPU and GPU. The task is to find out. “Winning Conditions” of each board In chess that might be checkmate in chip design It’s computational efficiency.
Google engineers trained reinforcement learning algorithms in a dataset of 10,000 chip floormaps of varying quality. some of which are randomly generated Each design is tagged with a specific “reward” function based on the achievement of various metrics such as required wire length and power consumption. The algorithm then uses this information to differentiate between good and bad floor plans. and create my own design in return.
As we have seen when AI systems deal with humans in board games. Machines don’t have to think like humans. And it often comes to unexpected solutions to familiar problems. When DeepMind’s AlphaGo plays Lee Sedol, the go-to human champion, this dynamic leads to a slew of challenges. The infamous “Movement 37” – A seemingly absurd placement of parts by AI that continues to lead to victory.
Google’s chip design algorithm isn’t all that amazing. But the floor plan still looks different from the man-made one. instead of the rows of components placed on the mold The subsystems appear to be almost randomly scattered across silicon. nature show the difference With a human design on the left and a machine learning design on the right. You can also see a general difference in the image below from a Google article (ordered human on the left and confused AI on the right), although the layout is blurry due to secrecy:
This article is of great importance. Because Google is currently using research commercially. But it’s far from the only aspect of AI-powered chip design. Google itself has explored using AI in other parts of the process, such as “architecture exploration,” and competitors like Nvidia are looking for other ways to speed up workflows. The intelligent circuitry of AI design chips for AI appears to have only just begun.