Recently, the reporter walked into the Wuxi Domestic Intelligent Computing Center. The modular liquid-cooled GPU cabinets were arranged in an orderly manner. Other than the buzzing sound of the air conditioner, the MTT S5000 with 2,000 Mohr threads was running almost at full load. It only took more than two months for this domestic GPU cluster to be fully loaded from the launch in June of this year. As the operator of the project, Zhang Tao, deputy general manager of Wuxi Huishan Investment Promotion Group, told reporters: “Up to now, more than 92% of the cluster computing power is already in use, and paying users account for more than 60%. Overall, computing power is in short supply, and many customers are waiting for our cluster computing power to be freed up before further testing.” The reporter learned that on this computing power platform, more than 20 customers are already testing or official business, covering model training and inference, etc., of which the training business accounts for more than 62%. Fang Zhaoxia, vice president of Moore Thread and general manager of Ecological Cooperation, explained that this training mainly focuses on later training. Fang Zhaoxia revealed that at the beginning of the first phase of the 2,000 card deployment, planning considerations were already being carried out according to the future 10,000 card scale. “We have solved the problems in the 10,000 card cluster project. The real challenge is 100,000 cards. This is a systematic project involving collaboration in all aspects of calculation, storage, heat dissipation, and data exchange.” She explained that building a 100,000 card cluster in 2027 is a very important strategy for Moor Threads, and a more intensive “supernode” architecture will be adopted according to customer needs. In September of this year, JD Cloud and Moore Thread announced that they have reached an in-depth cooperation to build a 100,000 card full-function GPU cluster to create an autonomous and controllable large-scale computing power base to meet the low latency and high reliability training and implementation needs of physical AI.
Zhitongcaijing · 2d ago
Share
Listen to the news
Recently, the reporter walked into the Wuxi Domestic Intelligent Computing Center. The modular liquid-cooled GPU cabinets were arranged in an orderly manner. Other than the buzzing sound of the air conditioner, the MTT S5000 with 2,000 Mohr threads was running almost at full load. It only took more than two months for this domestic GPU cluster to be fully loaded from the launch in June of this year. As the operator of the project, Zhang Tao, deputy general manager of Wuxi Huishan Investment Promotion Group, told reporters: “Up to now, more than 92% of the cluster computing power is already in use, and paying users account for more than 60%. Overall, computing power is in short supply, and many customers are waiting for our cluster computing power to be freed up before further testing.” The reporter learned that on this computing power platform, more than 20 customers are already testing or official business, covering model training and inference, etc., of which the training business accounts for more than 62%. Fang Zhaoxia, vice president of Moore Thread and general manager of Ecological Cooperation, explained that this training mainly focuses on later training. Fang Zhaoxia revealed that at the beginning of the first phase of the 2,000 card deployment, planning considerations were already being carried out according to the future 10,000 card scale. “We have solved the problems in the 10,000 card cluster project. The real challenge is 100,000 cards. This is a systematic project involving collaboration in all aspects of calculation, storage, heat dissipation, and data exchange.” She explained that building a 100,000 card cluster in 2027 is a very important strategy for Moor Threads, and a more intensive “supernode” architecture will be adopted according to customer needs. In September of this year, JD Cloud and Moore Thread announced that they have reached an in-depth cooperation to build a 100,000 card full-function GPU cluster to create an autonomous and controllable large-scale computing power base to meet the low latency and high reliability training and implementation needs of physical AI.
Disclaimer:Webull uses external vendor Google Translation Service for news translations where we endeavour to ensure these are correct, however, we recommend that you please double-check this information accordingly. Webull is not responsible for translation errors or issues.
What's Trending