中科曙光正式公布scaleX万卡超集群：每节点640卡、总算力超500亿亿次

Core Viewpoint - The launch of the scaleX supercluster by Zhongke Shuguang marks the first real machine appearance of a domestic AI supercluster with compatibility for open standards, showcasing significant advancements in computing power and efficiency [1] Group 1: Product Features - The scaleX supercluster consists of multiple scaleX640 ultra nodes, each housing 640 cards, totaling 10,240 acceleration cards with a computing power exceeding 5 EFlops (500 billion billion operations per second) [1] - The self-developed scaleFabric network chip provides a bandwidth of 400 Gb/s and latency below 1 microsecond, with the capability to expand to 100,000 cards, resulting in a 30% reduction in network costs [1] Group 2: Cooling and Efficiency - The system employs full immersion phase change liquid cooling, enhancing the computing density in a single cabinet by 20 times, achieving a Power Usage Effectiveness (PUE) of 1.04 [1] - A three-level collaborative optimization improves the training and inference efficiency of large models by 30-40%, with GPU utilization rates increasing by up to 55% [1] Group 3: Reliability and Availability - The cluster boasts a high availability rate of 99.99%, with less than 4 minutes of downtime over a 30-day period [1]