Nvidia speeds GPU design with AI

Designing a graphics processor starts with tiny building blocks called standard cells, the basic logic pieces that get repeated millions of times on a chip. Nvidia said its artificial intelligence system now ports that cell library to a new manufacturing process overnight instead of taking eight engineers about 10 months. (nvidia.com) (tomshardware.com) Bill Dally, Nvidia’s chief scientist and senior vice president of research, described the work during a March 2026 conversation at the company’s Graphics Technology Conference with Jeff Dean, chief scientist at Google DeepMind and Google Research. Dally said Nvidia trained an internal large language model on decades of graphics processor design data and is using artificial intelligence across several parts of its chip-design flow. (nvidia.com) (videocardz.com) A standard cell library is a catalog of prebuilt parts such as logic gates, memory elements and other circuit blocks that engineers reuse when they assemble a processor. When a chipmaker moves a design to a new semiconductor process, that library has to be rebuilt or adapted so the parts still meet targets for size, power use and speed. (tomshardware.com) (letsdatascience.com) Dally said Nvidia’s reinforcement learning tool for that job, referred to in reports as NVCell or NB-Cell, runs on a single graphics processor and produces results that match or beat human work on area, power and delay. He also said Nvidia is applying artificial intelligence to design exploration, bug handling and verification, not just cell-library work. (tomshardware.com) (letsdatascience.com) (videocardz.com) That work lands as Nvidia is trying to ship ever more complex processors into a market shaped by artificial intelligence demand, while also laying out new product road maps at Graphics Technology Conference 2026 in San Jose from March 16 to March 19. Shortening internal design steps can help a chip company move faster between manufacturing nodes and product generations. (nvidia.com) (blogs.nvidia.com) (cnbc.com) Nvidia did not present this as fully automated chip design. Dally said the company is still “a long way” from artificial intelligence designing chips without human input, and the reported use case is a narrow but labor-intensive step inside a larger engineering process. (tomshardware.com) (videocardz.com) Chip companies have used software automation for decades through electronic design automation tools, which check rules, place circuits and verify that a design works before it is manufactured. Nvidia’s claim points to a newer layer, where machine learning systems learn from past chip work and search for better layouts or fixes faster than teams can do by hand. (tomshardware.com) (nvidia.com) The immediate takeaway is narrower than “artificial intelligence designs chips now.” Nvidia is saying one repetitive, expert-heavy stage of graphics processor design can be compressed from roughly 80 person-months to one night, while human engineers still decide what gets built and whether the result is good enough to tape out. (tomshardware.com) (nvidia.com)