Zhang Jianzhong discusses Moore Threads’ full-function GPUs, 100,000-GPU computing plans and the challenges facing China’s GPU industry.
“Had no one taken the first step, China might have remained absent from this field for a long time.”
Zhang Jianzhong, founder, chairman and CEO of Chinese GPU maker Moore Threads, remarked in a recent interview with China News Service, as he discussed his decision to leave Nvidia and found Moore Threads.
In 2020, Zhang left his position as Nvidia’s global vice-president and China general manager to establish Moore Threads. At the time, China relied heavily on imported high-end GPUs, while only a small number of companies were working in the field.
“Setting up Moore Threads was mainly about filling that gap,” Zhang said.
For Moore Threads, filling the gap meant developing what Zhang calls a “full-function GPU”. Rather than focusing on a specific type of computing task, he believes future computing scenarios will require a platform that can support AI computing, graphics rendering, scientific computing, physical simulation and ultra-high-definition video processing.
Taking On a Difficult GPU Route
A full-function GPU covers a wide range of computing workloads. As a result, Zhang sees it as one of the more technically demanding routes in GPU development.
“To achieve this vision, the whole team has to be patient and work through technical challenges in every part of the process,” he said. “Whether it is 3D graphics, large-model training and inference, or scientific computing and physical simulation, we need to develop across the board.”
Over the past six years, Moore Threads has moved from research and development to product deployment. The company developed its first full-function GPU 18 months after it began operations. It has since mass-produced five GPU chips, introduced five generations of architecture and built a product portfolio covering cloud, edge and end devices, according to Zhang.
The company’s MTT KUAE intelligent computing cluster has also expanded from the 1,000-GPU level to 10,000 GPUs. Zhang said it can support the training of models with trillions of parameters, with several key indicators reaching levels comparable to international peers.
The next challenge is larger.
Moore Threads plans to work towards a 100,000-GPU computing cluster this year, Zhang said. The company is not simply looking to increase the number of GPUs. At that scale, the engineering challenges grow across hardware, systems, software-hardware coordination, reliability and operations.
Training models with trillions, and eventually tens of trillions, of parameters will require computing infrastructure at the scale of tens of thousands or even 100,000 GPUs, Zhang said.
More importantly, such a system must remain stable over long training cycles. For large AI models, a training run can last three to six months. Any major interruption can add significantly to the cost and time of development.
Competition Beyond One Company
Zhang sees Moore Threads’ development as part of a broader effort by Chinese technology companies to develop domestic computing technologies.
In his view, no single private company can accelerate an entire industry on its own. Instead, companies need to work on different technologies and follow different technical routes.
“The future world will inevitably be the result of many technologies working together,” Zhang said. “Only then can an ecosystem thrive.”
He also described the growing number of Chinese GPU companies and their different technical approaches as a positive development. However, he said the industry still needs to address common challenges before domestic GPUs can compete more widely in the market.
For Zhang, market acceptance will ultimately depend on more than simply having a domestic alternative. Chinese chipmakers need to improve performance, price competitiveness and services, he said.
At the same time, Zhang acknowledged that Chinese GPUs still have ground to make up.
“Objectively speaking, domestic GPUs are still behind the world’s advanced level and will need more time and effort to catch up,” he said.
Looking ahead, Zhang expects continued cooperation across the industry to help Chinese GPU products gradually reach levels comparable to overseas products. He also expects improvements in manufacturing capacity and supply-chain management to strengthen the position of domestic GPUs in global competition.
As manufacturing capacity and supply-chain management improve, Zhang expects domestic GPUs to gradually develop the ability to compete with leading global companies.
By Liu Liang
If you liked this article, why not read: China’s Next Stage | Private Enterprises Drive Innovation
