NVIDIA compresses GPU design time
- NVIDIA Chief Scientist Bill Dally said at GTC 2026 that AI now handles one internal GPU design step that once took eight engineers 10 months. - The task is porting NVIDIA’s standard cell library to a new process node; Dally said an internal system now completes it overnight on one GPU. - NVIDIA has pursued AI chip-design tools for years, but verification and manufacturing checks still keep humans in the loop. (nvidia.com)
A chip is built from tiny reusable blocks called standard cells, and NVIDIA says AI now redraws those blocks for new manufacturing processes overnight instead of over months. (nvidia.com 1) (nvidia.com 2) At NVIDIA’s GTC 2026 event in San Jose, Chief Scientist Bill Dally said one internal job that used to take eight engineers 10 months now runs overnight on a single graphics processor. (nvidia.com) Dally described that job as porting the company’s standard cell library to a new semiconductor process, a step needed before larger chip blocks can be built on a new node. (nvidia.com 1) (nvidia.com 2) Standard cells are the Lego-like pieces of a chip: logic gates, flip-flops and other basic parts that engineers combine into bigger circuits. NVIDIA’s 2021 NVCell paper said reinforcement learning could produce equal or smaller area layouts for more than 90% of single-row cells in an industry-standard library. (nvidia.com) That older NVIDIA work had already cut a task that once took a 10-person team months into an automated run lasting a couple of days. The 2026 claim pushes the same category of work further, from days to an overnight run on one GPU. (nvidia.com 1) (nvidia.com 2) NVIDIA’s design-automation research group says it is applying Bayesian optimization, reinforcement learning and generative artificial intelligence across chip design, from hardware code to physical layout and sign-off. (nvidia.com) The company’s published work also shows why this does not amount to fully autonomous chip design. NVIDIA’s research describes advanced-node layouts as constrained by complex design rules, routability limits and manufacturing checks that still have to be satisfied before a design can ship. (nvidia.com) (nvidia.com) For NVIDIA, the immediate gain is speed in a bottleneck that repeats every time a design moves to a new process node. The opening claim was simple: one step that once consumed 80 person-months now runs overnight. (nvidia.com)
