Relying on the best-selling cutting-edge AI chip H100 worldwide, Nvidia is undoubtedly one of the most prominent "trendsetters" in this AI wave this year. The development of generative AI requires a significant amount of computing power, and Nvidia's cutting-edge AI chips are the preferred choice for the development of Large Language Modeling (LLM), almost monopolizing global computing power.
So, if we say that the cutting-edge AI chip, the "chicken", has given birth to the globally popular "egg" of generative AI. Is it possible for this increasingly mature "egg" to hatch new "chicks" in the future? Nvidia is currently making the latest attempt.
A research paper released by NVIDIA on Monday describes how to use generative artificial intelligence to assist in one of the most complex engineering tasks: designing chips.
Modern chips are circuits composed of billions of transistors. One of the most challenging tasks in the technology industry is to figure out how to arrange these transistors on silicon wafers.
NVIDIA's GPU chip is undoubtedly one of the most complex chips in the industry, and has become the core technical support behind generative AI tools such as ChatGPT. As shown in the following figure, under a microscope, state-of-the-art chips such as the Nvidia H100 look like a carefully planned metropolis, built from billions of transistors connected to "streets" ten thousand times thinner than hair.
In order to build such a "digital giant city", multiple engineering teams and thousands of R&D personnel often need to coordinate for up to two years. Some groups are responsible for determining the overall architecture of the chip, some are responsible for making and placing various ultra small circuits, and some are responsible for testing their work. Each task requires specialized methods, software programs, and computer languages.
In response, Mark Ren, Research Director of NVIDIA and the first author of the latest research report, stated that& Quot; I believe that over time, large language models will help participate in all processes.
How can AI help develop chips?
Nvidia demonstrated its research attempt by using so-called large-scale language models and utilizing data accumulated over the past 30 years in chip design history for model training. After investigating possible use cases with NVIDIA engineers, the research team chose three use cases as the starting point: chat robots, code generators, and analysis tools.
The research paper provides a detailed introduction to how NVIDIA engineers can create a customized LLM called ChipNeMo for internal use. This LLM is trained on internal company data to generate and optimize software, and provides assistance to human designers.
Bill Dally, Chief Scientist of NVIDIA, stated, "It has been proven that many of our senior designers spend a considerable amount of time answering questions from junior designers. Using chat robots can help answer questions from junior designers, thereby saving them a lot of time
Nvidia also found in its research that by adding a large amount of specific data accumulated by the company, an intermediate chat robot can become more accurate than an advanced chat robot. Nvidia stated that this helps to control the cost of the system.
Another feature showcased by the company is the use of artificial intelligence to generate code. Dally said that engineers typically spend a lot of time searching for parts of the chip that don't work and using testing tools to identify the cause. For testing purposes, an artificial intelligence system can quickly write a piece of code to operate the tool.
Ren, who has worked in the field of electronic design automation for over 20 years, pointed out that in the long run, engineers hope to apply generative artificial intelligence to every stage of chip design, thereby significantly improving overall productivity.
Dally emphasized, "Our goal is not to achieve process automation or replace manual labor, but to utilize our existing personnel, empower them with stronger capabilities, and make them more productive