Experts from Google Research have taught artificial intelligence to design microcircuits. It turned out that he copes with the task much faster and better than a person. A study published in the journal Nature.
Designing a physical chip layout is time-consuming, even though some of the processes are automated. To make this process more efficient, specialists used machine learning technology. In particular, they developed several algorithms that perceived the task of creating a mock-up as a board game. Like chess, the board acts as a silicon wafer, and the pieces are the components (cores, graphics, memory, etc.). Algorithms see the main goal as maximizing efficiency – productivity and optimal energy consumption through the arrangement of components.
The researchers used 10,000 ready-made schemes for designing chips of different quality and then analyzed the results. These indicators guided google’s algorithms to distinguish more efficient chip designs from less efficient ones and create their own variants. It takes Google artificial intelligence less than six hours to create one chip mockup, while a human can take several months to complete the same task.
If we analyze the work of the algorithms, then it looks a little chaotic, as if the AI is carelessly scattering components all over the plate. Nevertheless, this does not negatively affect the results of the work. Research has shown that using neural networks to build chips can help make them more efficient. In practice, Google uses them to develop the next generation of Google’s tensor processors.