国产万卡超集群真机亮相 中科曙光发布scaleX系统

Core Viewpoint - The launch of the scaleX WanKa supercluster by Zhongke Shuguang at the HAIC2025 conference marks a significant advancement in large-scale intelligent computing systems, showcasing China's capabilities in AI infrastructure [1][3]. Group 1: Product Features - The scaleX WanKa supercluster is designed for trillion-parameter models and complex tasks, featuring innovations in supernode architecture, high-speed interconnects, storage performance optimization, and system management [3][4]. - It includes a world-first single cabinet-level 640-card supernode, achieving over 5 EFlops of total computing power and a PUE value as low as 1.04, enhancing computing density by 20 times [4]. - The system utilizes a self-developed RDMA high-speed network, achieving 400 Gb/s bandwidth and under 1 microsecond communication latency, improving performance by 2.33 times compared to traditional IB networks while reducing overall network costs by 30% [4]. Group 2: Optimization and Management - The supercluster features deep optimization of storage, computing, and transmission, enhancing data transfer efficiency and increasing AI accelerator card resource utilization by 55% [4]. - It incorporates digital twin technology for intelligent scheduling and management, ensuring 99.99% availability and supporting management of tens of thousands of nodes and users [5]. - The scaleX supercluster supports over 400 mainstream large models and is applicable in various fields such as financial risk control, geological energy exploration, and scientific intelligence [5]. Group 3: Collaborative Ecosystem - The "AI Computing Open Architecture" initiative, launched by Zhongke Shuguang in collaboration with over 20 AI industry chain enterprises, aims to lower the R&D threshold for intelligent computing clusters and promote ecosystem development [5].