Ricursive Intelligence bets on AI-designed chips to ease the compute crunch

On stage at TechCrunch Disrupt, Ricursive Intelligence founders Anna Goldie and Azalia Mirhoseini laid out their plan to tackle the looming compute crisis. The idea: hand chip design over to AI that learns from every iteration and makes each successive design better than the last.

Why it matters

Goldie said demand for compute is set to grow at least 10,000-fold over the next five years, and even that may be conservative. Nobody wants to build 10,000 times more data centers, so the bet is on efficiency — chips that make better use of bandwidth, wafers and fabs. «Even if we build data centers in space or something, you won't see the sun anymore,» she quipped.

Humans dream, AI executes

The thought of AI designing its own chips may unsettle some, but Goldie frames it differently: freed from the grind of design, people can focus on ideas and applications. She sees a broader shift in which humans bring the creative thinking while AI handles execution — already true in software development, and now coming to chip design.

Unconventional layouts

AI is challenging traditional design methods by finding unusual layouts that outperform the conventional ones, the founders say. Their earlier system, AlphaChip, produced a circular, «donut-shaped» arrangement of components that surprised engineers at first but ultimately worked well. At Ricursive they optimize an entire chip at once rather than splitting it into dozens of smaller blocks as is standard. That lets the AI create intricate patterns, including wave-like designs — details they promise to share later.

Cheap, custom chips, fast

The real payoff, Goldie argues, is speed. When designing a new chip takes six months to two years, you have to be extremely selective about what you build. But if a chip can be designed in about a week, it becomes financially sensible to make custom chips for all kinds of applications — space, AR and VR, biological computing — and to make them more energy efficient.

Source: TechCrunch