Nvidia: AI speeds chip design
Nvidia says AI reduced a GPU design task that previously took ten months and eight engineers to an overnight job, while stressing that fully AI‑led chip design remains a long way off. The company framed this as automation accelerating technical iteration but still requiring human oversight. (tomshardware.com)
Nvidia says it now uses artificial intelligence to finish one graphics processor design task overnight instead of taking eight engineers 10 months. (tomshardware.com) The task is called standard cell library porting: moving thousands of tiny, reusable circuit building blocks to a new chip manufacturing process. Bill Dally, Nvidia’s chief scientist, said the job covers about 2,500 to 3,000 cells and now runs on a single graphics processor. (videocardz.com) Dally said the artificial intelligence system was trained on decades of Nvidia graphics processor design data, and that the resulting cells match or beat human work on size, power use and delay, which is the time a signal takes to travel through a circuit. (videocardz.com) Chip design is a long chain of steps, from writing the logic to checking for bugs to arranging transistors on silicon, and each step has to fit the rules of a factory process from companies such as Taiwan Semiconductor Manufacturing Co. Nvidia’s design automation research group says it works across logic synthesis, physical design, verification and manufacturing sign-off. (research.nvidia.com) That matters because Nvidia is trying to build new artificial intelligence chips on a faster cadence as demand from data centers keeps rising. At its March 2026 Graphics Technology Conference in San Jose, the company used its keynote and technical sessions to argue that gains now depend on speeding up the full stack, including the way chips themselves are designed. (nvidia.com) Nvidia did not present this as fully autonomous chip design. Dally said the company uses artificial intelligence in design exploration, library work, bug handling and verification, but said artificial intelligence is still far from designing chips without human input. (tomshardware.com) The company has been talking about this direction for at least two years. In August 2024, industry publication eeNews Europe reported that Nvidia had trained a large language model on Verilog, a hardware design language, and was using autonomous agents to support graphics processor, central processor and networking chip work. (eenewseurope.com) The near-term use case is not a machine inventing an entire processor from scratch. It is software handling narrow, repetitive engineering work faster, so human designers can review results and move to the next round sooner. (tomshardware.com)
