中科曙光支撑国家级算力枢纽，3套万卡超集群同时落地

Group 1 - The core point of the article is the launch of the scaleX 30,000-card supercluster system by Zhongke Shuguang, marking the largest domestic AI computing power pool in China, capable of supporting trillion-parameter model training and high-throughput inference [1][3] - The scaleX supercluster represents a significant advancement in the deployment and practical application of domestic AI computing infrastructure, having been showcased at the HAIC conference and operational within two months [3] - Zhongke Shuguang has overcome key technical challenges such as high-speed interconnection, tight coupling of storage and computing, and efficient resource scheduling to enable the scaleX supercluster [3] Group 2 - The AI industry is experiencing explosive growth in global intelligent computing demand, necessitating infrastructure that is ecologically compatible and flexible to support diverse applications [3] - The scaleX supercluster is designed to be compatible with mainstream software ecosystems like CUDA, allowing for mixed deployment of various domestic acceleration cards, thus lowering the adaptation barrier for developers [3] - The supercluster has completed optimization for over 400 mainstream large models and can connect to thousands of applications, fostering a collaborative innovation ecosystem in the AI industry [4] Group 3 - The scaleX supercluster provides deep empowerment across multiple scenarios, including support for trillion-parameter model training and fault recovery, and has enhanced inference efficiency for major internet users [5] - In the AI for Science domain, the supercluster has supported a domestic materials research model that achieved top rankings internationally, significantly improving protein research efficiency by 3 to 6 orders of magnitude [5] - Zhongke Shuguang aims to deepen technological research and application practices, promoting domestic intelligent computing to serve various sectors of economic and social development more efficiently and inclusively [5]