AI-Driven Chip Design
Ricursive Intelligence co-founders Anna Goldie and Azalia Mirhoseini presented their work at a Disrupt conference on Wednesday, arguing that AI-driven chip design could help address a projected surge in compute demand. Goldie noted that the industry may face at least a 10,000x rise in demand for compute within the next five years, a figure she suggested could be conservative. Improving chip efficiency is one way to avoid the necessity of building orders of magnitude more data centers.
Innovative Design Approaches
The startup is developing systems that design chips, learn from each iteration, and refine subsequent layouts with the aim of automating and accelerating the chip design process. The founders claim their AI challenges conventional methods by optimizing whole chips rather than splitting designs into dozens of smaller blocks. For instance, they cited AlphaChip, a system they previously developed, which generated a circular, “donut-shaped” arrangement of components. Engineers were initially surprised by that layout, but it ultimately proved successful in tests. They added that the AI has produced intricate, wave-like patterns and other unconventional topologies, which the company plans to disclose in more detail later.
Speed and Efficiency in Chip Production
Goldie argued that faster design cycles could change which chips are practical to build. “If it takes six months or a year or two years to design a new chip, then you really have to be extremely selective about the chips that you design,” she said. “But if you can design a chip in like a week… you can design custom chips for all sorts of different types of applications in space and AR, VR, and biological computing, and it makes sense financially to do that, and we can be more energy efficient.”
Future Developments
The founders contend that delegating execution to AI would allow engineers and designers to focus on applications and new use cases, while the AI handles iterative optimization. They believe this approach could enhance manufacturing and wafer use efficiency and enable chip variants that were previously impractical due to design time and cost. Ricursive Intelligence plans to continue refining its tools and to share further technical details about its AI-generated layouts and design process in the future.
Original source: TechCrunch AI