Nvidia spoke about how its artificial intelligence helps with the development of new GPUs.
PrefixRL is the AI. PrefixRL uses deep reinforcement learning to outperform traditional Electronic Design Automation tools (EDA) built by Cadence or Synopsys, as well as Siemens/Mentor.
PrefixRL can create arithmetic circuits entirely from scratch. They are smaller and more efficient than traditional tools. Nvidia claims that PrefixRL has created almost 13,000 such schemes in the Hopper architecture.
64b adder circuits created with PrefixRLAI (left) are 25% smaller that circuits made by modern EDA tools (right), but just as fast and as functionally equivalent.
Nvidia points out that AI like this requires large computing resources. Learning how to create a 64-bit addition circuit required 256 CPU cores per GPU, and 32,000 GPU hours.