NVIDIA has not only built powerful AI GPUs and driven a new wave of AI around the world, but it has also deployed AI on a large scale internally, including the GPU chip design process. NVIDIA chief scientist Bill Dally mentioned in a conversation with Google chief scientist Jeff Dean,NVIDIA has extensively applied AI in the chip design stage, including design exploration, standard cell library development, bug processing, verification and other stages.

However, he emphasized thatIt’s still too early to use AI for fully end-to-end automated chip design, but when it will be realized, he is not willing to predict easily.

Dally shared a specific case of AI chip design.

In the past, NVIDIA required a team of 8 people to work for 10 months to migrate the standard cell library to a new process technology, which is a workload of 80 person-months.

Now, NVIDIA has a tool based on reinforcement learning, NB-Cell, which has been iterated for two or three generations. Now, only one GPU graphics card can be run overnight to complete the above work.

More importantly,The units generated by AI tools reach or even exceed the level of manual design in terms of area, power consumption, and delay., allowing for rapid deployment of new processes.

Dally also mentioned another internal toolPrefix RL, aiming at a long-term research problem, that is, the layout of the lookahead stage in the carry lookahead chain.

He claimed,The layout generated by the AI ​​tool is "something a human engineer could never think of," and the key performance indicators are 20-30% higher than manual design.

This shows that NVIDIA uses AI not only to improve efficiency and save time and labor, but also to explore design solutions that exceed conventional human intuition.

Prefix RL

On a more macro level, Dally also revealed,NVIDIA already runs two large language models internally: Chip Memo and Bug Nemo.

These large models are fine-tuned based on NVIDIA's proprietary data, including register transfer level (RTL) code and architectural documentation from years of GPU design.

Dally said that one of the practical benefits they bring is that when junior engineers encounter problems, they can directly ask questions to the large model and get answers. They no longer need to repeatedly consult senior designers, who can also focus on higher-value work.

At the same time, they can also help summarize bug reports and assist in assigning them to corresponding modules or engineers.

It is worth mentioning that NVIDIA does not seem to be laying off junior employees because of the efficiency improvements brought by AI tools. Instead, it is cultivating them to make rapid progress through more efficient methods.

Compared with many companies that frequently use AI to replace and eliminate employees, perhaps this is the most balanced application of AI.