Nvidia used AI to slash a GPU‑design task
Nvidia says it used an internal AI model to reduce a GPU‑design task that previously took eight engineers ten months to an overnight job. The company trained an LLM on decades of GPU‑design data and applied it to a narrowly defined, verifiable engineering workflow rather than claiming fully autonomous chip design. (Tom's Hardware)
Designing a graphics processing unit starts with millions of tiny choices, and Nvidia said one of those jobs now runs overnight with an internal artificial intelligence tool instead of taking eight engineers 10 months. (tomshardware.com) Nvidia Chief Scientist Bill Dally said at the company’s March 2026 GTC conference that the task was porting roughly 2,500 to 3,000 standard cells to a new manufacturing process. Standard cells are the basic Lego-like building blocks chip teams reuse to assemble larger circuits. (videocardz.com) Dally said Nvidia’s internal tool, called NB-Cell, uses reinforcement learning, a training method based on trial and error, and can do that standard-cell work on one graphics processing unit overnight. He said the resulting cells were better on size, power, and delay than human-designed versions on that task. (videocardz.com) Chip design has long depended on electronic design automation software, which checks whether circuits fit, run fast enough, and can actually be manufactured. Nvidia’s own design automation research group says it works across the flow from register-transfer level design and verification to physical design and sign-off. (research.nvidia.com) That same research group says it is applying Bayesian optimization, reinforcement learning, and generative artificial intelligence to chip-design problems, and training custom large language models for chip design tasks. Nvidia has been building that effort in public for several years, not just since this month’s claim. (research.nvidia.com) In October 2023, Nvidia researchers published ChipNeMo, a domain-adapted large language model for chip design. The paper said the model was tuned for three uses: an engineering assistant chatbot, electronic design automation script generation, and bug summarization and analysis. (research.nvidia.com, arxiv.org) IEEE Spectrum reported in October 2023 that ChipNeMo drew on Nvidia’s internal archive of design documents, code, and bug reports built up over about 30 years of chip work. Dally said then that the point was to capture institutional memory so junior engineers could query the system instead of repeatedly interrupting senior designers. (spectrum.ieee.org) Dally told GTC attendees in March 2026 that Nvidia is also using artificial intelligence for design exploration, bug handling, verification, and circuit layout. A separate internal layout tool, PrefixRL, improved some key metrics by 20% to 30%, according to accounts of the session. (videocardz.com, nvidia.com) Nvidia did not present this as fully autonomous chip design. The 2023 ChipNeMo paper said there was still “room for improvement” between its results and ideal outcomes, and Dally’s 2026 remarks described narrow, verifiable jobs inside a larger human-run workflow. (research.nvidia.com, tomshardware.com) The immediate effect is less about a machine “inventing” a new graphics processing unit than about compressing one repetitive bottleneck in a process that already spans architecture, verification, layout, and manufacturing checks. Nvidia’s claim is that one of those bottlenecks is now short enough to fit into a single night. (research.nvidia.com, tomshardware.com)
