Workflow
昇腾AI+盘古
icon
Search documents
华为云CEO:384超节点每卡性能可达英伟达H20三倍
Guan Cha Zhe Wang· 2025-08-30 03:38
Core Viewpoint - The importance of chips is acknowledged, but the ability to provide the required computational results for customers is emphasized as more critical [1] Group 1: Huawei Cloud's Developments - Huawei Cloud is undergoing significant organizational restructuring, with a focus on enhancing computational power through its Ascend AI cloud services and Tokens services [1] - The CloudMatrix384 super node, launched in April, integrates 384 Ascend NPUs and 192 Kunpeng CPUs, achieving a computational scale of 300 PFlops [2] - The Tokens service has been integrated with the CloudMatrix384 super node, achieving a maximum of 2400 TPS and 50 ms latency, surpassing industry standards [2][3] Group 2: Performance and Growth Metrics - Huawei Cloud's overall computational capacity has increased by nearly 250% compared to the previous year, with the number of Ascend AI cloud service customers rising from 321 to 1714 [5] - The CloudMatrix384 super node can support training of large models, with the capability to connect 432 super nodes to form a 160,000-card AI cluster [2] - The deployment of over 40 CloudMatrix384 super nodes in Guizhou is part of a strategy to create a national computational network [5] Group 3: Market Position and Strategic Focus - Huawei Cloud ranks second in China's cloud service market with an 18% share, while Alibaba Cloud holds a 33% share [6] - The market demand is shifting from "cloud adoption" to "AI integration," prompting Huawei Cloud to streamline its operations and focus on maximizing the advantages of its Ascend AI and Pangu model combinations [6][7] - The organizational restructuring aims to concentrate resources on core areas that can leverage AI capabilities effectively [6][7]