Nvidia says AI sped chip design

Nvidia says it is already using artificial intelligence inside chip design, with one internal tool shrinking a months-long library-porting job to an overnight run. (videocardz.com) The claim came from Nvidia chief scientist Bill Dally during a March 2026 conversation with Google DeepMind chief scientist Jeff Dean at the company’s GTC conference in San Jose. Nvidia’s on-demand session page says the talk focused on hardware, systems and algorithms for the 2026-to-2030 period. (nvidia.com) A chip is built from thousands of small reusable logic blocks called standard cells, and those blocks must be adapted each time a company moves to a new manufacturing process. Dally said Nvidia’s tool, called NB-Cell, now ports about 2,500 to 3,000 cells overnight on one graphics processor instead of a process that used to take eight engineers about 10 months. (videocardz.com) (letsdatascience.com) Nvidia said the NB-Cell system is based on reinforcement learning, a method that improves by trying actions and scoring the results. Dally said the resulting cell layouts can match or beat human work on size, power and delay, the standard measures for chip design tradeoffs. (letsdatascience.com) (videocardz.com) Nvidia described the work as one piece of a broader push to use machine learning in electronic design automation, the software stack engineers use to draw, check and optimize chips before factories make them. The company’s research group says it is working on graphics-processor-accelerated design tools, artificial intelligence for design automation and large language models trained for chip-design tasks. (research.nvidia.com) This is not Nvidia’s first public step in that direction. In October 2023, Nvidia researchers introduced ChipNeMo, a domain-tuned large language model trained on internal chip-design data to answer engineering questions, summarize bugs and help write scripts for design tools. (blogs.nvidia.com) (arxiv.org) Nvidia said in that earlier work that modern chip programs can take multiple engineering teams as long as two years, which helps explain why shaving months from one repeated task would matter inside a company shipping new graphics processors on tight cycles. The same 2023 project framed the goal as raising designer productivity rather than handing the full process to a model. (blogs.nvidia.com) (community.element14.com) Dally also said full end-to-end automated chip design is still far off, even as Nvidia applies artificial intelligence to design exploration, bug handling and verification. That leaves the current pitch narrower: automate specific bottlenecks that repeat every time a new process node arrives. (videocardz.com) Other chip and software companies are chasing the same idea. Communications of the Association for Computing Machinery reported in 2024 that Nvidia, Intel, Advanced Micro Devices, International Business Machines, Google and Apple were all using artificial intelligence in parts of design, simulation, verification and testing. (cacm.acm.org) For Nvidia, the immediate claim is not that a model designed a whole graphics processor by itself. It is that one of the slowest repeat jobs in chip development can now be compressed into a single night on a single graphics processor. (videocardz.com)