Ricursive Intelligence pursues AI chip design to ease compute pressure

Ricursive Intelligence puts AI at the centre of chip design

Ricursive Intelligence founders Anna Goldie and Azalia Mirhoseini say their startup is developing AI that can design chips, learn from each iteration and improve later designs. Speaking at TechCrunch Disrupt, they positioned the approach as a response to rapidly rising demand for computing capacity.

Goldie said demand for compute could be at least 10,000 times higher within the next five years, describing that estimate as potentially conservative. The founders’ argument is not that the industry should build 10,000 times more data centres, but that it must make chips, bandwidth use, wafers and fabrication capacity more efficient.

Whole-chip optimisation challenges established workflows

Traditional chip design commonly divides a chip into dozens of smaller blocks. Ricursive instead says it is optimising the entire chip. The founders argue that this gives AI room to identify layouts and patterns that conventional processes may not produce.

They pointed to AlphaChip, an AI system they developed previously, as an example. AlphaChip generated a circular, donut-shaped placement of chip components. Engineers were initially surprised by the design, but the founders said it ultimately proved successful.

The pair also described more intricate possibilities, including wave-like patterns, and said they plan to share further details in the future. Their broader premise is that AI can take on execution in chip design while people bring ideas and concentrate on applications and new computing possibilities.

Faster custom silicon could widen the design agenda

Goldie contrasted a conventional chip-design cycle of six months, a year or even two years with the prospect of designing a chip in roughly a week. A shorter cycle, she said, would make teams less selective about which chips they pursue and could make custom designs financially practical for more applications.

The founders cited space, augmented reality, virtual reality and biological computing as areas that could benefit from more readily designed specialised chips. They also tied the case for custom hardware to greater energy efficiency.

For businesses planning compute-intensive products, the practical implication is to track whether AI-assisted whole-chip design can reduce the time and selectivity that currently limit application-specific hardware decisions.