This is not a first attempt. At Google, the pair created AlphaChip, an AI tool that could produce workable chip layouts in hours, a job that normally takes human designers a year or more. AlphaChip was used to help design Google’s Tensor Processing Units, the custom AI chips Google runs in its data centers. Accounts differ on how many generations: the company’s funding announcement says four, while an earlier profile said three.
One AlphaChip result became famous inside the field. It arranged chip components in a circular, donut-shaped layout that surprised engineers at first but went on to work well. The founders say this is the point: AI can find layouts humans would never try.
Ricursive is now going further. Instead of the usual method of splitting a chip into dozens of smaller blocks and optimizing each one, it works on the whole chip at once. The founders say that lets the AI produce intricate patterns, including wave-like designs they plan to share more about later.
The work has not been free of controversy. One Google colleague spent years trying to discredit the AlphaChip research and was fired in 2022, according to an earlier report.