Eight engineers mobilized for ten months to port a standard cell library to a new node, NVIDIA now says it can do the same work in one night on a single GPU. Bill Dally detailed at the GTC an already concrete use of AI in chip design, far from automation integral of the flow.
NVIDIA AI in the design flow
During an exchange with Jeff Dean, chief scientist at Google, the chief scientist at NVIDIA listed several application points already active internally: design exploration, work on standard cell libraries, bug processing and verification. On the other hand, a fully automated end-to-end chip design still remains far away.
NB-Cell et prefix RL
The most specific example concerns NB-Cella tool based on reinforcement learning. Bringing NVIDIA’s standard cell library to a new manufacturing process has previously involved a team of eight people for about ten months, or 80 man-months, to process 2,500 to 3,000 cells.
According to Bill Dally, NB-Cell, now in version 2 or 3, runs this job overnight on a single GPU. The resulting cells match or exceed human designs on three metrics cited by NVIDIA: area, dissipation and delay. The productivity gain is massive.
Dally a aussi cité prefix RLanother internal tool applied to the placement of lookahead stages in a lookahead carry chain. He claims that the system produces layouts that no human would have come up with, with a gain of around 20 to 30% over human designs on key metrics.
Internal LLMs for GPUs and bugs
NVIDIA also operates internal LLMs called Chip Nemo et Bug Nemo. These models have been fine-tuned on the company’s proprietary data, including RTL and architecture documents covering GPUs designed over the years.
Chip Nemo and Bug Nemo every day
The use put forward is very concrete. A junior engineer can query the model to understand how a specific block works instead of constantly engaging a senior designer, while the system can also summarize bug reports and help assign them to the right module or engineer.
The most interesting point is not only the reduction in engineering time. NVIDIA also uses AI to explore placement and optimization solutions outside of usual human heuristics, which can directly impact the PPA, therefore the development speed and the competitiveness of a future GPU even before the manufacturing stage.
Source : VideoCardz
