Nvidia: AI sped up GPU design

Nvidia says it used artificial intelligence to turn one labor-intensive graphics chip design job from about 10 months into an overnight run. (nvidia.com) The company’s chief scientist, Bill Dally, said the task was porting Nvidia’s standard cell library to a new manufacturing process — the tiny, reusable logic blocks that get repeated across a chip. He said that work used to take eight engineers about 10 months, or roughly 80 person-months. (nvidia.com) Dally said Nvidia now uses a reinforcement learning system called NB-Cell to do that porting work overnight on a single graphics processing unit. He said the resulting cells “match or exceed” human work on size, power and delay, the three measures chip teams use to judge efficiency and speed. (videocardz.com) A standard cell library is the parts bin for a modern chip: logic gates, flip-flops and other basic circuits that engineers combine into larger blocks. Every time a chip company moves to a new process node, those parts have to be rebuilt to obey a new set of manufacturing rules. (nvidia.com) That job has gotten harder as chipmaking has moved below 5 nanometers, where routing tracks shrink, patterning rules multiply and automated tools struggle to keep layouts manufacturable. Nvidia researchers said those limits were a reason they built NVCell and later NVCell 2, earlier versions of the same automation effort. (research.nvidia.com, research.nvidia.com) Nvidia has been working on this line of design automation for years, not just since the recent boom in chatbots. In 2021, the company said NVCell could cut a task that took months for a 10-person team down to a process that ran in a couple of days. (nvidia.com) The company is also using other artificial intelligence tools deeper in the chip workflow. Dally has pointed to PrefixRL for circuit design exploration, while Nvidia researchers have described ChipNeMo, a custom large language model trained on internal design data to help engineers search documents, write code and track bugs. (nvidia.com, nvidia.com) Nvidia has also published work on multi-agent systems for hardware design, including tools for Verilog code generation, timing analysis and debugging. In a February 2025 technical post, Nvidia said one timing analysis agent delivered about 60 times faster turnaround than human engineers on that task. (developer.nvidia.com) But Nvidia is not saying artificial intelligence can design a full graphics processing unit on its own. Dally said the company is “a long way” from autonomous chip design, and Nvidia’s own research posts describe these systems as assistants that handle narrow, specialized steps inside a much larger human-led process. (videocardz.com, nvidia.com) That leaves Nvidia with a narrower claim: artificial intelligence is shaving time off some of the slowest, most repetitive parts of chip design. For a company shipping ever-larger graphics processors into an artificial intelligence boom, even one overnight shortcut can remove a bottleneck. (nvidia.com, nvidia.com)