Google announces that it has accelerated the development of AI-only processors using AI
Google is developing a dedicated processor 'Tensor processing unit (TPU) ' specializing in machine learning, and in May 2021 announced the fourth generation 'TPU v4'. A new Google research team has revealed that using AI to develop TPU has dramatically improved development speed.
A graph placement methodology for fast chip design | Nature
AI system outperforms humans in designing floorplans for microchips
Google's AI approach to microchips is welcome — but needs care
Google is using AI to design its next generation of AI chips more quickly than humans can --The Verge
Microchips such as TPUs include numerous components such as CPUs, GPUs, and memory controllers, as well as ultra-fine wiring that spans a total of several kilometers. During the development phase of a microchip, there is a process called a 'floor plan ' that determines the layout of these components and wiring, but according to the research team, a human floor plan can take months. There is. Therefore, the research team developed a method using AI to reduce the time required to execute the floor plan.
The research team likened the microchip board to the board of a board game such as chess or go, and each component and wiring to a chess piece or go stone. It's similar to learning a game. ' Furthermore, citing the fact that AI defeated the world champion of Go, he is developing the claim that 'AI can derive a floor plan that is superior to humans.'
World champion who lost to Go AI 'AlphaGo' retired as a Go player saying 'I can not beat AI' --GIGAZINE
The research team has developed an AI that can derive a power-saving layout with a small chip area by letting AI learn 10,000 floor plans tagged with power consumption, wiring length, microchip area, etc. Did. The figure below shows an image of a human floor plan (left) and an image of an AI floor plan (right). If you look at the figure, you can see that in the floor plan by humans, each component is arranged in an orderly manner, whereas in the floor plan by AI, each component is dispersed separately.
The pictures below are the result of human execution of the TPU floor plan (left) and the result of AI execution (right). The photos have been blurred for confidentiality, but a closer look reveals that the AI results have an irregular layout compared to the human results. According to the research team, it took several weeks for humans and 6 hours for AI to execute this floor plan, showing that AI can execute the floor plan at explosive speed.
The research team said, 'The AI-based microchip development method developed this time may help save thousands of hours of human working hours.' 'Powerful hardware designed by AI is AI's. We believe in creating a symbiotic relationship between AI and hardware, such as driving progress. '