The Virtuous Cycle of AI
Pete Singer
1 week ago
AI is everywhere and it’s evolving rapidly. What’s most exciting is to see how it’s being used for chip design and manufacturing in a “virtuous cycle” — a positive chain of events that reinforces itself through a feedback loop, leading to continued success.
Jim Chambers, Vice President, NVIDIA, described how AI is changing chip design, speaking at a recent technology breakfast event hosted by Applied Materials just before SEMICON West in Phoenix. “We’re currently experiencing an explosion in the diversity of LLM (large language models) use cases for chip design. Although RTL generation remains one of the primary use cases, there are new use cases coming in such as text message generation, all sorts of chatbots and co-pilots, as well as hardware security and more.”
 He also explain about NVIDIA is using physics-informed neural networks to embed the laws of physics in the neural network, with products such as PhysicsNeMo, Warp and cuLitho variants. “We have surrogate models with machine learning to replace complex equations like Navier-Stokes for fluid mechanics, and we have generative AI with diffusion models for fine-grained simulations,” he said. “If you like multidimensional partial differential equations, then Warp’s the tool for you.” Warp offers 3D spatial computing with multi-GPU acceleration for large-scale simulations. cuLitho variants enable large many-bodied models to simulate quantum wave functions.
“Using these tools, you can accelerate your workflows by 10 X, bringing it up to 10,000 X, and cuLitho is a great example. With cuLitho, we see a 40 X speed-up for lithography. Photomask generation that used to take weeks can now be done overnight. The work of 10,000 CPUs can be replaced with just a few hundred GPUs. And it’s not just a performance upgrade, it’s a strategic enabler for faster time to market, lower manufacturing costs, greater scalability, and improved supply chain resilience,” Chambers said.
“Imagine how good it can get when you accelerate all your workflows,” he added. “If you build digital twins for all the key elements in your world, then you have something incredibly valuable. For a fab, you can connect all the way down from a facility to the transistor level and back. Once you have this in hand, then you have a truly autonomous fab that enables lights-out manufacturing.”
In the end, AI will enable something that is truly speed of light, accurate and scalable, an engine that builds better technologies to support even more advanced AI, that will in turn build better technologies. It’s “a positive feedback loop that I hope will continue for a very long time,” Chambers said. |
