ChipNeMo: NVIDIA's ChatGPT

The researchers, in their study, found that domain-specific LLMs, like ChipNeMo, designed to carry out specific tasks significantly improve their performance when compared to standard, one-size-fits-all models. They also found that they could make the LLMs five times smaller in size while still achieving similar or even better results in designing semiconductors.

The team found that in the context of chip designing, ChipNeMo performed better with as few as 13 billion parameters in comparison to much larger general-purpose LLMs like LLaMA2, which have 70 billion parameters. 

“I believe over time large language models will help all the processes, across the board,” said Mark Ren, an NVIDIA research director and lead author on the paper.

Bridging the supply-demand gap

NVIDIA is a big player in the AI race as the company’s valuation reached a $1 trillion market capitalization in May this year. In response to the global supply shortages, NVIDIA is planning on increasing its production of its GPUs, which power generative AI applications like ChatGPT. The company’s aim is to produce about 2 million units in 2024, up from the 500,000 targeted this year, reported Financial Timesin August.

The team also tested the LLM’s capability to generate code and found that it churned out concise software snippets, typically spanning 10 to 20 lines. This can help engineers to understand why a chip part doesn’t work and save them time.

“This effort marks an important first step in applying LLMs to the complex work of designing semiconductors,” said Bill Dally, NVIDIA’s chief scientist. “It shows how even highly specialized fields can use their internal data to train useful generative AI models.”

Packing billions of transistors into tiny spaces is often a hit-and-trial method. AI-powered assistants like ChatNeMo can enhance human productivity.

The researchers concluded that their future work will focus on further improving ChipNeMo models and methods to make them ready for production use.