Startup Ricursive Intelligence Bets on AI-Designed Chips to Ease the Compute Crunch
The founders of startup Ricursive Intelligence say more efficient chips designed by AI could help address the compute crisis. They presented their approach at TechCrunch Disrupt.
Anna Goldie and Azalia Mirhoseini, founders of the startup Ricursive Intelligence, explained on Wednesday at the TechCrunch Disrupt technology conference how their company is building AI that designs computer chips. The system learns from each iteration and improves later designs, with the aim of automating and speeding up chip design. The founders believe this could help ease the growing shortage of computing power.
Surging demand
Goldie said there is an understanding that demand for compute will grow at least 10,000-fold over the next five years, and that this estimate may even be conservative. In her view, the answer should not be building 10,000 times more data centers. Instead, chips should be made more efficient, using data center bandwidth as well as wafers and fabs more efficiently.
The roles of humans and AI
Goldie acknowledged that some people may be uneasy about AI designing its own chips. She argues that it would free humans to focus on applications and on exploring new possibilities, while AI handles execution. She said this shift is also happening in software development and other fields.
Unconventional layouts
The founders say AI can find unconventional layouts that outperform traditional ones. AlphaChip, an AI system they developed earlier, produced a circular, "donut-shaped" arrangement of chip components that at first surprised engineers but proved successful. Ricursive now optimizes entire chips instead of splitting them into dozens of smaller blocks, as is traditionally done. This yields intricate patterns, including wave-like designs, which the founders plan to discuss in more detail later.
New possibilities
Goldie stressed that if a new chip takes six months to two years to design, companies must be very selective about what they build. If a chip can be designed in about a week, custom chips become possible for a range of uses, such as space, augmented and virtual reality, and biological computing, provided it makes financial sense, with better energy efficiency.
