Nvidia says AI cuts 10-month, eight-engineer GPU design task to overnight job — company is still ‘a long way’ from AI designing chips without human input

Nvidia is already using AI across multiple stages of chip design, from circuit-level optimizations to system-level exploration, and achieves orders-of-magnitude productivity gains and, in some cases, better-than-human results, according to Dally.

Article continues below

You may like

At the lowest level, AI has already transformed standard cell development, one of the most time-consuming steps when transitioning to a new fabrication process. Porting a standard cell library of roughly 2,500–3,000 cells previously required a team of eight engineers working for about 10 months, according to Dally. Nvidia has replaced this work with a reinforcement learning system called NB-Cell, which can now complete the same task overnight on a single GPU.

At a higher level, Nvidia has developed internal large language models — Chip Nemo and Bug Nemo — trained on proprietary architecture documentation covering all GPUs that Nvidia has ever developed. These LLMs can act like engineering assistants who can explain to junior designers how complex hardware blocks work. As a result, Nvidia no longer has to bother senior engineers about things that can be done by LLMs.

“We had a series of LLMs that we called Chip Nemo and Bug Nemo. We took a generic LLM, and then we fine-tuned it by feeding it all of the design documents proprietary to Nvidia,” Dally said. “So this is stuff you cannot get outside the company, it is all of the RTL, hardware design docs, all of the RTL for every GPU ever designed at Nvidia, all of the architecture specs for those. Now you have this LLM that is actually very smart about GPU design. […] When you have a junior designer, they can ask Chip Nemo, and Chip Nemo will explain [how GPUs work]. It improves productivity that way; it is a very patient mentor.”

Beyond cell libraries and engineering assistance, Nvidia is applying reinforcement learning to classical circuit design problems. For example, an RL-based system explores design options in a trial-and-error manner, and this approach helps to create chip designs that exceed human results in terms of area, power, and performance faster than humans can.

Get Tom’s Hardware’s best news and in-depth reviews, straight to your inbox.

“It comes up with totally bizarre designs that no human would ever come up with, but they are actually 20% or 30% better than the human designs,” said Dally.

In addition to using AI for place and route, Nvidia is also using AI to explore architectural designs.

In particular, Nvidia’s agent-based systems run large numbers of experiments, evaluate different design directions, and narrow down viable configurations. This greatly accelerates decision-making in the early stages of the chip development cycle when engineers must choose between various architectural trade-offs.

What to read next

Last but not least, Nvidia uses AI for design verification, one of the longest stages in the chip development cycle. Nonetheless, AI still cannot be responsible for the whole verification process, so Nvidia must emulate its designs and conduct actual experiments to ensure that everything works fine.

“We would like to collapse that space, what the really long pole is design verification,” said Dally. “We are particularly looking at how we can use AI to prove that designs work more quickly.”

In the long term, Nvidia’s chief scientist envisions chip development to shift to a multi-agent model in which specialized AI systems will handle different parts of the design, like human teams do today. For now, AI acts to cut development time by assisting engineers and improving design quality to levels beyond what humans can do, which in turn enables engineers to explore more design options than before.