China News Service, Beijing, September 13. Title: Dialogue with Moore Thread Zhang Jianzhong: Building a domestic GPU, a private enterprise’s way of making "cores"
China News Service reporter Liu Liang
"If no one takes this step, this track may be absent from China for a long time." Zhang Jianzhong, founder, chairman and CEO of Moore Thread, recently accepted an interview with a reporter from China News Service to share his original intention of starting a business and his thoughts on the development of domestic GPUs (general computing units).
In 2020, Zhang Jianzhong, then NVIDIA's global vice president and general manager of China, resigned from his executive position and founded Moore Threads. At that time, the development of global artificial intelligence (AI) was on the eve of an explosion, and most of China's high-end GPUs relied on imports. "The creation of Moore Thread was mainly to fill this gap in China." Zhang Jianzhong said.
To fill this gap, Moore Threads chose to start with a "full-featured GPU". Zhang Jianzhong judged that future computing scenarios require an accelerated computing platform with comprehensive capabilities in artificial intelligence computing, graphics rendering, scientific computing and physical simulation, and ultra-high-definition video encoding and decoding.
"From the perspective of national science and technology development and industrial security, this is also the fundamental way to solve the 'stuck neck' problem." Zhang Jianzhong emphasized that if we do not build a full-featured GPU, we will still need to purchase graphics computing chips abroad in the future development process, and industrial security will still face risks.
Among the existing GPU routes, "full-featured GPU" is recognized as one of the most technically challenging routes in the industry. "In order to realize this vision, the entire team must endure loneliness and patiently complete technological breakthroughs in all aspects of full-featured GPUs. Whether it is 3D graphics computing, large model training and reasoning, or scientific computing and physical simulation, they all need to be fully developed." Zhang Jianzhong said.
In the past six years, the Moore Thread team has gradually promoted research and development and product implementation. Eighteen months after its official operation, Moore Thread successfully developed the first full-featured GPU chip. So far, it has successfully mass-produced five chips, launched a fifth-generation architecture, and built a full-stack product layout covering "cloud-edge-end"... Moore Thread continues to solve computing power problems and explores the construction of an independent full-stack computing power system.
Zhang Jianzhong introduced that the current Moore Thread Kua'e intelligent computing cluster has achieved expansion from kilocard to 10,000 card level, can support trillion parameter level model training, and many key indicators have reached the international mainstream level.
Moore Thread plans to start research on a 100,000-card scale cluster this year. Zhang Jianzhong introduced that training large models with trillions or even 10 trillion parameters in the future requires a computing power scale of 10,000 cards or 100,000 cards. Only by ensuring that these devices operate stably within a training cycle of 3 to 6 months can we avoid excessive training time from placing a burden on large model companies.
"We hope that through the next technical research, we can truly solve the pain point of the industry's computing power training platform being 'stuck'." Zhang Jianzhong said that in the future, the company will build larger-scale intelligent computing clusters to support the development of AI infrastructure.
The development of Moore's thread is a microcosm of the current exploration of independent innovation by many private technology companies in China. Zhang Jianzhong said that a private enterprise cannot accelerate the development of the entire industry alone. Only by working together on different tracks and different technical routes can the country's scientific research strength continue to develop.
As far as the GPU industry is concerned, Zhang Jianzhong said that currently, the domestic GPU industry is developing rapidly, with an increasing number of companies and different technical routes. This is a healthy development trend. But no matter which company it is, if it wants to make a breakthrough, it must jointly solve the common problems faced by the development of the industry.
He emphasized that in order for domestic chips to finally gain market recognition, they must go through a process from "not as good as others in price/performance" to "equal" to "surpassing", and ultimately achieve better performance, more competitive price/performance and better services. "This is also the only way for Chinese manufacturing to go global."
"Objectively speaking, the current domestic GPU chips are still far behind the world's advanced level, and more time needs to be spent to catch up." Zhang Jianzhong said.
Looking to the future, he believes that through the efforts of industrial partners, domestic GPU chips are expected to gradually reach a level comparable to foreign products in the future. With the further improvement of production capacity and supply chain management level, domestic GPUs will gradually have the ability to compete with global leading companies. (over)
