Nvidia: AI slashes GPU design time
Nvidia says AI reduced a GPU design task that previously took ten months and eight engineers to something that can be done overnight, while stressing AI still needs human oversight for chip design. The example frames AI as a way to compress expert workflows rather than replace expert judgement. (tomshardware.com)
Designing a graphics processing unit starts with tiny logic blocks called standard cells, the Lego bricks of a chip. Nvidia said its AI tools now generate those cell libraries overnight instead of using a team for months. (nvidia.com) At Nvidia’s GTC conference in San Jose in March 2026, Chief Scientist Bill Dally said porting a standard cell library to a new manufacturing process used to take eight engineers about 10 months. He said the job covers roughly 2,500 to 3,000 cells. (nvidia.com; videocardz.com) Dally said Nvidia’s reinforcement-learning system, called NB-Cell, now does that work overnight on one graphics processing unit. He said the resulting cells match or beat human versions on size, power use and delay, which is the time a signal needs to move through a circuit. (videocardz.com; nvidia.com) A standard cell library is the base kit chip designers reuse every time a company moves to a new factory process from a manufacturer such as Taiwan Semiconductor Manufacturing Co. That step has to be done before teams can assemble larger parts of a processor from those blocks. (nvidia.com; nvidia.com) Nvidia did not present this as push-button chip design. Dally said the company is using AI in design exploration, standard cell work, bug handling and verification, while fully automated end-to-end chip design remains far off. (videocardz.com; nvidia.com) The company also said internal large language models trained on Nvidia’s design history help junior engineers search decades of architectural knowledge. That puts AI in the role of a faster assistant inside an expert workflow, not a replacement for the engineers signing off on the chip. (nvidia.com) Nvidia has been publishing research on this problem for years. Its 2020 NVCell paper described reinforcement learning for standard-cell layout, and a 2023 NVCell 2 paper reported gains in routability and clean layouts on advanced-node cells. (nvidia.com; nvidia.com) Other parts of the chip flow are getting the same treatment. On April 9, 2026, Siemens said its Veloce system, used with Nvidia technology, let designers run and capture trillions of pre-silicon verification cycles before chips were manufactured. (siemens.com) Nvidia’s claim is narrower than “AI designs chips now.” The company’s own description is that AI is compressing one labor-heavy step in chip design from 80 person-months to a night of computing, while people still decide whether the result is good enough to build. (videocardz.com; nvidia.com)
