US computing firm NVIDIA is adapting AIs to help its human engineers build better computer chips.
Tech companies have been scrambling to secure supplies of NVIDIA鈥檚 chips 鈥� graphics processing units that sell for tens of thousands of dollars each 鈥� in the midst of an AI gold rush and subsequent chip shortage. NVIDIA is the leading supplier of chips for training and developing AIs, and its latest efforts put large language models (LLMs) to work improving such chips.
鈥淭his could become an instance of the 鈥楢I improving chips, chips improving AI鈥� virtuous cycle,鈥� says at the University of California, San Diego.
Advertisement
NVIDIA鈥檚 engineers customised Llama 2 models 鈥� developed and released by Meta 鈥� by training the AIs on specialised data from NVIDIA鈥檚 own chip design and verification processes. The specially trained LLMs were renamed ChipNeMo models.
Their initial AI applications included a chatbot assistant that can answer questions about NVIDIA鈥檚 chip architecture and design, a generator that writes snippets of code for chip design software and a tool capable of automatically summarising and updating descriptions of known software bugs.
The chatbot assistant鈥檚 responses were rated 7.4 out of 10 by human experts, whereas the bug summarisation tool earned 4 to 5 points on a 7-point scale and the code generator achieved just over 50 per cent correctness. The researchers did not include human test scores, but said these AI scores 鈥渟till show a considerable gap with human expert performance鈥�.
It is too early to tell how such scores may lead to either improving chip design quality or saving labour time, says Kahng. But even if they are not yet beating humans, he suggests these AI tools could still prove helpful.
鈥淭hese examples are still in a sweet spot of mechanical and small-scale tasks that involve code and text, where humans can verify correctness of outputs, or where best efforts 鈥� for example, catching some bugs, if not all bugs 鈥� are helpful,鈥� says Kahng.
Training AI chatbots on data specific to designing chips 鈥� as NVIDIA has done 鈥� may allow them to generate 鈥渂etter-quality answers to technical questions than the general ChatGPT answers I am getting鈥�, says at the University of California, Berkeley. He uses ChatGPT to produce code scripts that save him time. However, at this point he describes NVIDIA鈥檚 early results as 鈥渆xpected milestones鈥� that are 鈥渘either insignificant nor groundbreaking鈥�.
鈥淚t will be interesting to see what the next, say, five success stories will look like,鈥� Kahng says.
arXiv
