Nvidia: AI speeds chip design
Nvidia says AI compressed a GPU design task that once took eight engineers and ten months into an overnight run, though the company cautioned humans still provide framing and verification. The claim illustrates how AI can accelerate engineering workflows while leaving review and systems understanding to people. (tomshardware.com)
A modern graphics processing unit is built from thousands of tiny logic blocks, and Nvidia said one step in arranging those blocks now runs overnight instead of taking months. (tomshardware.com) Nvidia Chief Scientist Bill Dally said the task is porting a standard cell library, a catalog of prebuilt logic parts that has to be adapted each time the company moves to a new semiconductor manufacturing process. He said that work used to take eight engineers about 10 months, or roughly 80 person-months. (videocardz.com) Dally said Nvidia now uses a reinforcement learning system called NB-Cell for that job, and that the latest versions can process about 2,500 to 3,000 cells overnight on one graphics processing unit. He said the resulting cells match or exceed human designs on size, power use, and delay, which is the time a signal takes to travel through a circuit. (videocardz.com) Chip design has become harder as manufacturing nodes shrink and the number of checks rises, which is why electronic design automation software from companies such as Cadence and Synopsys has become central to the industry. Nvidia said in March that Cadence, Siemens, Synopsys and other industrial software companies are building Nvidia-powered artificial intelligence agents to plan, optimize and verify chip and system workflows. (investor.nvidia.com) Nvidia is also using large language models inside engineering teams, not just search-and-optimize systems such as NB-Cell. Its 2023 ChipNeMo research paper said domain-adapted models were tested for engineering assistant chatbots, electronic design automation script generation, and bug summarization and analysis. (research.nvidia.com) That earlier Nvidia work framed the models as assistants trained on internal design data rather than autonomous chip architects. The company’s researchers said ChipNeMo was built by adapting general large language models with chip-design documents, code, and retrieval tools so engineers could query specialized knowledge more directly. (research.nvidia.com) Dally also said Nvidia is “a long way” from letting artificial intelligence design whole chips without people, according to Tom’s Hardware’s account of his remarks. He said humans still set objectives, understand tradeoffs across the whole system, and verify the outputs before they are used. (tomshardware.com) That leaves Nvidia describing artificial intelligence as a speed tool for narrow, expensive parts of semiconductor work, not a replacement for the engineers who decide what the chip is supposed to do. The immediate change is in cycle time: a job that once occupied a team for most of a year can now be run between one workday and the next. (tomshardware.com)
