Google is developing a TPU (Tensor processing unit), a processor dedicated to machine learning, and announced the fourth-generation TPU v4 in May 2021. The Google research team is making it clear that they have dramatically improved the development speed by using AI in the development of a new TPU.
A microchip such as a TPU contains numerous components such as a CPU, GPU, and memory controller, as well as fine wiring spanning several kilometers. In the microchip development stage, there is a process called floor plan that determines the layout of these components and wiring, but according to the research team, it may take several months for a human-made floor plan. Therefore, the research team developed a method to use AI to shorten the time it takes to execute a floor plan.
The research team points out that learning the floor plan to AI by comparing the components and wiring to board games such as chess and Go is similar to learning board games such as chess and Go by comparing the microchip board to board games such as chess and Go. Also, citing the fact that AI beat the world champion of Go, they claim that AI can come up with better floor plans than humans.
The research team developed an AI that learned 10,000 floor plans, such as power wiring length and microchip area, to derive a layout with a smaller chip area and power saving. The human-drawn floor plan is packed with components, whereas the AI-made floor plan is in a dispersed state.
According to the research team, it took humans several weeks to execute the floor plan, while the AI took 6 hours. The research team said that the AI-driven microchip development technology developed this time can help save thousands of hours of human labor, and that a symbiotic relationship between AI and hardware will be born, such as powerful hardware designed by AI helping AI progress. it is stated that Related information can be found here.