Ricursive Intelligence founders Anna Goldie and Azalia Mirhoseini have detailed their startup's approach to using artificial intelligence (AI) for chip design. Speaking at TechCrunch Disrupt on Wednesday, the founders explained that their AI system is being developed to automate and accelerate the chip design process, learning and improving with each iteration.
Goldie stated that there is an understanding of a potential 10,000-fold increase in compute demand within the next five years. She suggested that making chips more efficient could help manage this demand without needing a proportional increase in data centres.
The founders claim that AI is challenging traditional chip design by discovering unconventional layouts. They cited AlphaChip, a previous AI system they developed, which produced a circular, "donut-shaped" arrangement of chip components that proved successful. They also explained that their approach involves optimising entire chips rather than partitioning them into smaller blocks, allowing for intricate patterns like wave-like designs.
Goldie also argued that AI-designed chips could enable the creation of new types of chips that were previously unfeasible. She suggested that reducing chip design time from months or years to approximately a week could make it financially viable to design custom chips for diverse applications, including space, augmented reality (AR), virtual reality (VR), and biological computing, while also improving energy efficiency.