NVL576

Search documents
华为徐直军:Atlas 950超节点算力超越英伟达
第一财经· 2025-09-18 09:16
"Atlas 950超节点,至少在未来多年都将保持是全球最强算力的超节点,并且在各项主要能力上都远 超业界主要产品。其中,相比英伟达同样将在明年下半年上市的NVL144,Atlas 950超节点卡的规模 是其56.8倍,总算力是其6.7倍,内存容量是其15倍,达到1152TB;互联带宽是其62倍,达到 16.3PB/s。即使是与英伟达计划2027年上市的 NVL576相比,Atlas 950超节点在各方面依然是领 先的。"徐直军表示,华为对于为人工智能的长期快速发展提供可持续且充裕算力,充满信心。(第一 财经记者 李娜) 9月18日,华为轮值董事长徐直军在演讲中表示,从大型AI算力基础设施建设的技术方向看,超节点已 经成为主导性产品形态。他透露,华为超节点Atlas 950 SuperPoD,算力规模8192卡,预计于今年 四季度上市,而Atlas 960 SuperPoD,算力规模15488卡,预计2027年四季度上市。 ...
华为徐直军:Atlas 950超节点算力超越英伟达
Di Yi Cai Jing· 2025-09-18 09:09
"Atlas 950超节点,至少在未来多年都将保持是全球最强算力的超节点,并且在各项主要能力上都远超 业界主要产品。其中,相比英伟达同样将在明年下半年上市的NVL144,Atlas 950超节点卡的规模是其 56.8倍,总算力是其6.7倍,内存容量是其15倍,达到1152TB;互联带宽是其62倍,达到16.3PB/s。即使 是与英伟达计划2027年上市的 NVL576相比,Atlas 950超节点在各方面依然是领先的。"徐直军表示, 华为对于为人工智能的长期快速发展提供可持续且充裕算力,充满信心。 (文章来源:第一财经) 9月18日,华为轮值董事长徐直军在演讲中表示,从大型AI算力基础设施建设的技术方向看,超节点已 经成为主导性产品形态。他透露,华为超节点Atlas 950 SuperPoD,算力规模8192卡,预计于今年四季 度上市,而Atlas 960 SuperPoD ,算力规模15488卡,预计2027年四季度上市。 ...
被抛弃的NVL72光互联方案
傅里叶的猫· 2025-07-17 15:41
Core Viewpoint - The article discusses the architecture and networking components of the GB200 server, focusing on the use of copper and optical connections, and highlights the flexibility and cost considerations in the design choices made by different customers [1][2]. Frontend Networking - The frontend networking in the GB200 architecture serves as the main channel for external data exchange, connecting to the internet and cluster management tools [1]. - Each GPU typically receives a bandwidth of 25-50Gb/s, with total frontend network bandwidth for the HGX H100 server ranging from 200-400Gb/s, while GB200 can reach 200-800Gb/s depending on configuration [2]. - Nvidia's reference design for frontend networking may be over-provisioned, leading to higher costs for customers who may not need such high bandwidth [2][4]. Backend Networking - The backend networking supports GPU-to-GPU communication across large-scale clusters, focusing on internal computational collaboration [5]. - Various switch options are available for the backend network, with initial shipments using ConnectX-7 cards and future upgrades planned for ConnectX-8 [6][10]. - Long-distance interconnections primarily utilize optical cables due to the limitations of copper cables over longer distances [6]. Accelerator Interconnect - The accelerator interconnect is designed for high-speed communication between GPUs, significantly impacting communication efficiency and system scalability [13]. - The GB200's NVLink interconnect has evolved from the HGX H100, requiring external connections due to the separation of NVSwitches and GPUs across different trays [14]. - Different configurations (NVL72, NVL36x2, NVL576) balance communication efficiency and scalability, with NVL72 being optimal for low-latency scenarios [15]. Out of Band Networking - The out-of-band networking is dedicated to device management and monitoring, focusing on system maintenance rather than data transmission [20]. - It connects various IT devices through baseboard management controllers (BMC), allowing for remote management and monitoring of system health [21]. Cost Analysis of MPO Connectors - The article estimates the value of MPO connectors in the GB200 server, indicating that the cost per GPU can vary significantly based on network architecture and optical module usage [22][23]. - In a two-layer network architecture, the MPO value per GPU is approximately $128, while in a three-layer architecture, it can rise to $192 [24]. - As data center transmission rates increase, the demand for high-speed optical modules and corresponding MPO connectors is expected to grow, impacting overall costs [25].