AI speeds chip design
Nvidia says AI cut an 80 person‑month GPU design task to an overnight run on one project, shrinking what used to take 10 months and a team of eight to hours. (tomshardware.com) The company cautioned the result is task‑specific and not yet a full replacement for human engineers, but the claim points to faster iteration cycles in chip engineering. (videocardz.com)
A chip is built from tiny logic blocks, like a city built from standard bricks, and Nvidia said artificial intelligence now helps remake some of those bricks for a new factory process in hours instead of months. (videocardz.com) Bill Dally, Nvidia’s chief scientist, said at Nvidia’s March 2026 Graphics Technology Conference session with Google chief scientist Jeff Dean that one internal tool, called NB-Cell, cut a standard-cell-library port from about 80 person-months to an overnight run on one project. (nvidia.com) A standard cell library is a catalog of prebuilt circuit pieces that engineers reuse across a chip, and each time a design moves to a new semiconductor process those pieces have to be rebuilt and checked. Tom’s Hardware reported Dally described the old workflow as roughly eight engineers working for 10 months. (tomshardware.com) Dally said Nvidia is already using artificial intelligence in design exploration, standard-cell work, bug handling, and verification, which is the stage where engineers test whether a chip design behaves as intended before manufacturing. He also said fully autonomous chip design is still “far off,” according to reports from the session. (videocardz.com) That matters because modern graphics processing units contain billions of transistors, and every new manufacturing node forces teams to retune libraries, timing, power use, and physical layouts before a chip can be taped out for production. Nvidia’s public GTC materials described the March 16-19, 2026 event as focused on the next generation of artificial intelligence and accelerated computing. (nvidia.com) The company’s claim points to faster iteration inside one of the slowest parts of hardware development: the repetitive engineering work needed to adapt proven building blocks to a new process technology. Nvidia did not present this as a system that invents an entire graphics processing unit by itself. (tomshardware.com) Reports from the session said Nvidia trained an internal large language model on decades of graphics processing unit design data and uses other artificial intelligence methods, including reinforcement learning, across parts of its chip flow. That puts the company in the position of using its own computing hardware to shorten the design cycle for future hardware. (videocardz.com) The next test is whether results like NB-Cell spread from one task to more of the chip pipeline without breaking the checks that keep expensive silicon from failing after fabrication. For now, Nvidia is describing a faster tool for human engineers, not a replacement for them. (nvidia.com)
