Nvidia says artificial intelligence is now helping it design the chips that power AI systems, including one internal task the company says once took eight people about 10 months and now finishes in a single night on one GPU.
Bill Dally, Nvidia’s chief scientist and senior vice president of research, described the company’s approach during a conversation with Google’s Jeff Dean at the 2026 GPU Technology Conference, or GTC. He said Nvidia is trying to bring AI into every stage of GPU design.
Cell library porting became the headline example
The task at the center of Nvidia’s claim involves porting the company’s standard cell library to a new semiconductor process. Nvidia said that library typically includes about 2,500 to 3,000 cells, and that the work used to require eight people and roughly 10 months.
Nvidia’s NVCell program now completes that process overnight on a single GPU, Dally said. He also said the results were better than what human engineers produced.
Other internal tools are aimed at design and training
Dally also pointed to Prefix RL, a reinforcement learning-based tool Nvidia uses to evaluate chip design options through trial and error. He said the system can generate unusual ideas, but that its results are 20% to 30% better than human designs.
On the engineering side, Nvidia has also built internal large language models called Chip Nemo and Bug Nemo. The company trained them on its proprietary database and codebase so they can help junior engineers and explain complex concepts more clearly.
The broader effort suggests Nvidia is using AI not just in products it sells, but in the chip-design process itself. The company said the approach is helping it save time, reduce manual work, and make future process-node transitions easier as new GPUs are developed.
Nvidia has also been applying AI on the consumer side. The company recently introduced Alpamayo, which brings AI models to self-driving cars. What remains to be seen is how broadly Nvidia can scale these internal tools and whether the same methods can continue to outperform traditional engineering workflows as GPU design grows more complex.