NVIDIA trained an internal LLM on decades of GPU design data, says Bill Dally
NVIDIA is already using AI across several parts of its internal chip design flow, according to Chief Scientist Bill Dally during a GTC conversation with Google Chief Scientist Jeff Dean. Dally said the company is applying AI to design exploration, standard cell library work, bug handling, and verification, but added that fully end-to-end automated chip design is still far off.
One example Dally shared was NVIDIA’s NB-Cell tool. He said porting the company’s standard cell library to a new semiconductor process used to take a team of eight people about 10 months, or 80 person-months in total. According to Dally, the reinforcement learning-based tool now does that work overnight on one GPU, and the resulting cells can match or exceed human designs in size, power, and delay.
“So, we’re trying to use AI wherever we can in our design process, and there are a number of examples. For example, every time we have a new semiconductor process, we have to port our standard cell library to it. It’s about 2,500 to 3,000 cells, and that used to take a team of eight people about 10 months, so 80 person-months.
Then we developed a program based on reinforcement learning called NB-Cell. I think we’re up to NB-Cell 2 or 3 now. And it’s overnight on one GPU. The results are actually better than the human designs in measures such as cell size, power dissipation, and delay. It matches or exceeds the human designs.
It’s a huge productivity gain, and it eliminates an obstacle to moving to new processes, because now we can move cell libraries over very quickly.”
— Bill Dally, NVIDIA Chief Scientist
Dally also pointed to another internal tool called prefix RL, which targets the long-studied problem of placing look-ahead stages in a carry lookahead chain. He said the system produces layouts that “no human would ever come up with,” while improving key metrics by roughly 20 to 30% compared to human designs. That is a useful detail because it shows NVIDIA is not only using AI to save engineering time, but also to search for designs outside normal human intuition.
© NVIDIA At a broader level, Dally said NVIDIA has been running internal LLMs called Chip Nemo and Bug Nemo . These models were fine-tuned on proprietary NVIDIA material, including RTL and architecture documents for GPUs designed over the years. He said one of the practical gains is that junior engineers can query the model instead of repeatedly asking senior designers how a specific block works, while the system can also summarize bug reports and help assign them to the right module or engineer. Well, at least one company didn’t get rid of ‘junior’ employees and found a way to teach them more efficiently.
Source:
