Recursive Intelligence founders Anna Goldie and Azalia Mirhoseini, who pioneered Google's AlphaChip project for TPUs, discuss how AI can revolutionize chip design. They highlight the mismatch between the rapid advancement of AI models and the slower pace of chip development, arguing that AI-driven chip design can bridge this gap. The discussion covers how their reinforcement learning agent was able to self-improve, discovering curved placements that humans had not considered, and the potential of synthetic data to overcome training data limitations. They envision a future where custom silicon becomes more accessible, enabling new applications and a Cambrian explosion of chip designs, ultimately transforming the industry from fabless to "designless."
Sign in to continue reading, translating and more.
Continue