A 60,000-card scientific intelligence (AI for Science, or AI4S) computing cluster system, provided by Dawning Information Industry Co., Ltd., has been put into operation at the national supercomputing internet core node in Zhengzhou. As the largest AI4S computing cluster in China, it leverages full-stack intelligent convergence capabilities to meet diverse user needs—from cluster performance and software model optimization to research application efficiency and scientific agent development—ensuring the large-scale implementation of "AI + science and technology" across the country.
"The deployment of this cluster is not only a showcase of technological achievement but also a milestone in the deep integration of artificial intelligence technology and scientific research innovation in China," said Xu Hongxing, academician of the Chinese Academy of Sciences and president of the Henan Academy of Sciences. "We look forward to collaborating with experts and industry representatives from various fields to seize the opportunities presented by the 'AI+' era, starting from this new computational foundation."
The cluster is designed with six key advantages to address growing demands for large-scale parallel computing, training of trillion-parameter models, and high-throughput inference:
- Powerful Computing: Deployment of 60,000 cards delivers world-class intelligent convergence computing performance. - Full Precision Support: Based on independently developed core chips, the system supports 8/16/32/64-bit full-precision computing, enabling efficient processing of high-dimensional functions and complex scientific challenges. - High-Speed Interconnect: Utilizing China’s first InfiniBand-like lossless high-speed network, the scaleFabric series meets extreme bandwidth and low-latency requirements for AI4S clusters. - Storage-Compute Coordination: Through "super tunnel" and AI data acceleration designs, the system achieves three-layer transmission coordination from chip to application, avoiding I/O bottlenecks. - Flexible Scheduling: An intelligent scheduling mechanism allows dynamic allocation of computing, storage, and network resources, supporting over 10,000 concurrent job dispatches per second. - High Reliability: With intelligent operation and maintenance, a digital twin system, and immersion phase-change liquid cooling, the system achieves 99.99% availability for long-term stable operation.
According to Li Bin, senior vice president of Dawning, the cluster not only excels in hardware performance but also adopts a co-design approach to ensure compatibility with mainstream computing ecosystems and support iterative development of scientific applications, effectively breaking down barriers between traditional supercomputing and intelligent computing.
Through collaborations with universities, research institutes, and enterprises, the cluster has already delivered significant results across multiple fields. For example, a 30,000-card configuration accelerated protein folding simulations by 1,000 times compared to traditional algorithms. A 45,000-card setup achieved trillion-atom liquid water molecular dynamics simulation, breaking world records while improving efficiency by over three orders of magnitude. It has also enabled turbulence simulations at a scale of hundreds of trillions of grids, greatly enhancing research productivity.
With the cluster now operational, the national supercomputing internet platform (scnet.cn) has established China’s largest AI4S computing infrastructure, integrating over 3 million CPU cores and 200,000 GPU cards into a nationwide computing network. This provides affordable AI4S computing services to universities, research institutions, and companies across the country.
To further advance AI4S engineering, Dawning’s cluster integrates massive datasets, domain knowledge bases, large models, and scientific computing tools under the OneScience platform—China’s first one-stop development platform for scientific AI models. OneScience includes dozens of AI4S models and datasets, allowing users to develop scientific large models within hours in fields such as earth sciences, bioinformatics, fluid simulation, and materials chemistry, addressing challenges like complex development environments and data accessibility.
On the supercomputing internet platform, users can bypass complicated software configurations and IT processes. By simply stating requirements in natural language, a "super scientific computing agent" automatically decomposes tasks, invokes models, and allocates computing resources for end-to-end delivery, reducing research task completion times from days to hours.
From the launch of the scaleX 10,000-card super cluster in December 2025, to the deployment of a 30,000-card system in February 2026, and now the 60,000-card AI4S cluster, Dawning has demonstrated rapid progress in engineering large-scale computing systems. This "Dawning speed" reflects not only advances in deployment capabilities but also a full-stack technological leap from AI to AI4S, enabling the widespread adoption of "AI+" across scientific disciplines and helping China secure a leading position in AI industrial applications.
Comments