China has officially launched its first fully homegrown 100,000-card artificial intelligence (AI) supercluster, marking a major leap in national computing infrastructure, according to the National Development and Reform Commission (NDRC).
The system, known as the Dawning 8000 (Dengfeng), links to the Zhengzhou core node of the national supercomputing internet and combines scientific and intelligent computing power to meet emerging demands such as advanced materials, drug development, and AI model training.
The national supercomputing internet is a nationwide computing network which now aggregates over 3.5 million CPU cores and 250,000 GPU cards, serving more than 1.4 million registered users with over 7,300 applications, 1,500 adapted large language models and over 11.3 million visits per month on average.
Combining scientific computing and intelligent computing, the cluster represents a forward-looking deployment for potential new computing needs in the future.
"The new computing cluster delivers peak performance equal to humanity computing nonstop for 200 years. It now supports more than 300 tasks across 26 fields, including new materials, drug innovation, and artificial intelligence," said Bu Jingde, chief engineer of high-end computing at the National High-Performance Computer Engineering Technology Research Center.
China has designated eight national computing power hub nodes and plans to build 10 data center clusters under its East–West Computing Resources Transmission Project, or East Data, West Computing.
Alongside Zhengzhou, several other nodes are also undergoing a new round of capacity expansion.
The national computing power hub node in the Guangdong-Hong Kong-Macao Greater Bay Area has put a new domestically produced 10,000-card cluster into use to meet the intelligent computing needs for next-generation large language model development.
"Domestic large models are moving from hundreds of billions of parameters to trillions of parameters. Our new intelligent computing cluster can cut the training time for trillion-parameter large models from one year to around six months, while also providing full-chain checking and testing environments for large models," said Liu Zhijun, head of the planning and construction of China Telecom's Guangdong-Hong Kong-Macao Greater Bay Area Integrated Data Center.
China launches massive homegrown AI supercomputer cluster
