Google is using machine learning to design the next generation of machine-learning chips. According to Google engineers, the algorithms are “comparable or superior” to human-made designs but can be created much faster. It is noted that work that takes people months can be done with the help of AI in less than 6 hours.
Google has been working on how machine learning can be used to create chips for several years. And, poking around, the company first applied its developments to a commercial product – 1 upcoming version of Google TPU (tensor processing unit) chips optimized for AI computing.
Google engineers note that this work has “serious consequences” for the chip industry. It should allow companies to more quickly explore the scope of architecture for future projects and make it easier to tailor chips for specific workloads.
The specific problem that Google’s algorithms solve is known as “floor planning.” It usually requires human engineers to use computer tools to find the optimal silicon die layout for the chip’s subsystems. These components include CPU, GPU, memory, which are linked together through many connections. Deciding where to place each component on a die affects the ultimate speed and efficiency of the chip. Thus, even minor changes in component placement can have huge consequences.
Google engineers note that designing “floor planning” takes “months of hard work” from people. But artificial intelligence algorithms can cope with this task much faster. Previously, AI algorithms have already proven their effectiveness in games and have outperformed humans in chess and go. Floorplanning is similar to game tasks. But instead of a game board, a silicon crystal is used, and instead of pieces such as knights and rooks, the logical components of a chip are used. So the challenge is to find the “winning conditions” for each board. In chess, it can be a checkmate. In chip design, it can be computationally efficient.
Google engineers trained the system using 10,000 datasets of various “floor plans.” Each option was marked with a specific “reward” function based on its success in various metrics, such as the length of the required connection and the energy consumption. The algorithm then used this data to distinguish between good and bad floor plans and, in turn, generate its own designs.
At the same time, it turned out that AI algorithms assemble chips completely different from humans. Instead of neatly arranged rows of components, the subsystems look like they are almost randomly scattered throughout the crystal.
A source: The verge