Tech war: Chinese ChatGPT ambitions threatened by lack of advanced AI chips like Nvidia A100 GPU, experts say

Stacking up less powerful GPUs could help China match computing power needed for large language models, one engineer said, but long-run challenges remain

To commercialise a GPT model, a company might need more than 30,000 Nivida A100 chips, which are banned from export to China, according to TrendForce

Reading Time:3 minutes

Powering sophisticated large language models like the one behind OpenAI’s ChatGPT requires tens of thousands of graphics processing units that do the heavy lifting for artificial intelligence servers. Photo: Shutterstock

Published: 11:00pm, 2 Mar 2023Updated: 1:59am, 3 Mar 2023

To train the large language models (LLMs) necessary for artificial intelligence (AI) bots like ChatGPT, China may have to rely on quantity over quality in graphics processing units (GPUs) after being cut off by the US from the most advanced chips, entrepreneurs and engineers in the country say.

Referring to LLMs as the steam engine of the modern era, one AI technology entrepreneur on the board of the China Association for Artificial Intelligence said in a recent closed-door seminar that the US leads China in computing power primarily because of its edge in GPUs, which have a unique advantage in machine learning over central processing units (CPUs) because of how they use simultaneous computations.

“We don’t have a card as strong as the A100, but we can pull together less powerful [GPU] cards to win with quantity,” the entrepreneur said about Nvidia’s data centre chip that the US has blocked from export to China.

He, like others quoted in this story, asked not to be named because of the sensitivity of the topic. “I think that computing power is still a very important, decisive factor to win the competition down the road.”

Since the launch last November of ChatGPT, developed by Microsoft-backed start-up OpenAI, the chatbot’s astonishing rise in popularity has sparked a race among Big Tech firms to develop their own versions of the generative AI product.