AI cut GPU design time
Nvidia says it used AI to compress a GPU design task that previously took 10 months and eight engineers into an overnight job, while stressing humans remain central to the work. (tomshardware.com)
Designing a graphics processing unit starts with tiny reusable logic blocks called standard cells, the Lego bricks of a chip. Nvidia said it now uses artificial intelligence to rebuild one of those libraries for a new manufacturing process overnight instead of using eight engineers for 10 months. (tomshardware.com) Bill Dally, Nvidia’s chief scientist and senior vice president of research, described the work during a March 2026 conversation with Google chief scientist Jeff Dean at Nvidia’s Graphics Technology Conference. Dally said the old job amounted to about 80 person-months and the new flow runs on a single graphics processing unit overnight. (nvidia.com, tomshardware.com) The software behind that claim is called NVCell, a system Nvidia Research has been publishing on since 2020. In its 2021 paper, Nvidia said NVCell used reinforcement learning, a trial-and-error training method, to generate layouts with equal or smaller area for more than 90% of single-row cells in an industry standard-cell library. (research.nvidia.com) Nvidia updated that work with NVCell 2 in 2023, focusing on routability, the problem of fitting wires through a crowded chip without breaking manufacturing rules. The company said advanced nodes beyond 5 nanometers created more routing problems because routing tracks shrank while design rules multiplied. (research.nvidia.com, research.nvidia.com) Dally said Nvidia is using artificial intelligence in other parts of chip design too, including design-space exploration, bug handling and verification. Tom’s Hardware reported that Nvidia trained internal large language models on decades of graphics processing unit design data, but Dally said the company is still far from letting artificial intelligence design chips without human input. (tomshardware.com, nvidia.com) That caution matches Nvidia’s earlier public work on language models for chip design. In 2023, the company said large language models could help engineers search manuals, write scripts and summarize hardware bugs, while Mark Ren, an Nvidia Research director, said he believed the models would help “across the board” over time. (blogs.nvidia.com) Nvidia also published a June 2024 paper on using a large language model for standard-cell layout optimization, a sign that it is testing language models alongside reinforcement learning in the same corner of the design flow. That paper said automated tools still struggled to produce top performance, power and area results for more complex sequential cells. (research.nvidia.com) The immediate point is not that Nvidia has removed engineers from chip design. The point is that one of the slowest, most repetitive parts of moving a design to a new process node is becoming software-heavy, with humans reviewing, steering and signing off on the results. (tomshardware.com, research.nvidia.com) For Nvidia, which is shipping new architectures while demand for artificial intelligence chips stays high, shaving months off internal design work could matter as much as adding raw compute. Dally’s message at Graphics Technology Conference 2026 was narrower than “artificial intelligence designs chips now”: the machine is speeding a hard subtask, and the humans are still in charge. (nvidia.com, nvidia.com, tomshardware.com)
