NVIDIA, known as the world’s most profitable chip-making company, has recently introduced a custom large language model named ChipNeMo, a breakthrough technology pivotal for artificial intelligence tools such as ChatGPT. This in-house development by the company aims to assist in generating and optimizing software, supporting human designers in the intricate task of constructing semiconductors. The technology, crafted by NVIDIA’s researchers, holds promising prospects, particularly in advancing the company’s work in graphics processing, artificial intelligence, and related technologies.
The advent of ChipNeMo serves as a significant aid to NVIDIA engineers in semiconductor design, a notably complex endeavor. Constructing these intricate devices, packed with millions and billions of transistors, demands the collaboration of multiple engineering teams laboring for extended periods, often spanning over two years. The superior quality of NVIDIA’s semiconductors is a testament to the arduous nature of their creation, contributing to their high demand and premium pricing in the market.
Researchers discovered that leveraging domain-specific Large Language Models (LLMs) like ChipNeMo, tailored to execute particular tasks, significantly amplifies their performance compared to conventional, one-size-fits-all models. Furthermore, they managed to shrink the size of LLMs by five times while retaining similar or even improved results in semiconductor design.
In the realm of chip design, ChipNeMo showcased superior performance with as few as 13 billion parameters, compared to the much larger, general-purpose LLMs like LLaMA2, boasting 70 billion parameters.
Mark Ren, an NVIDIA research director and lead author on the project, expressed optimism: “I believe over time large language models will help all the processes, across the board.”
NVIDIA, a major contender in the AI landscape with a market valuation hitting $1 trillion earlier this year, is addressing global supply shortages by scaling up GPU production. These GPUs power generative AI applications like ChatGPT. The company aims to increase production to approximately 2 million units by 2024, a substantial rise from the 500,000 units targeted for this year, as per reports by Financial Times in August.
The team also explored the LLM’s capability to generate concise code snippets, aiding engineers in troubleshooting and saving time.
Bill Dally, NVIDIA’s chief scientist, highlighted the significance of this effort: “This effort marks an important first step in applying LLMs to the complex work of designing semiconductors. It shows how even highly specialized fields can use their internal data to train useful generative AI models.”
The process of fitting billions of transistors into confined spaces often relies on trial and error. AI-driven assistants like ChipNeMo have the potential to significantly augment human productivity in this domain.
The researchers affirmed that their forthcoming work will revolve around further refining ChipNeMo models and methodologies, making them more viable for practical use in production settings.