AI服务器AI server

用于AI运算的服务器

每条都摘自下方编号的来源原文;只写事实,不写预测或建议。

技术简介

  • 戴尔PowerEdge XE9680是为AI应用设计、最多支持八颗全互连GPU的服务器。2
    The PowerEdge XE9680 is a state-of-the-art server that is built for effectively running the most demanding AI applications, supporting up to eight fully interconnected GPUs.
  • NVL72把72颗Rubin GPU和36颗Vera CPU放在单个液冷机柜中,作为一台机器运行。3
    72 NVIDIA Rubin GPUs and 36 NVIDIA Vera CPUs in a single liquid-cooled rack operating as one machine; Eighteen 1U compute trays, each with four Rubin GPUs and two Vera CPUs, connect through nine sixth-generation NVIDIA NVLink switch trays delivering 216 TB/s of scale-up bandwidth; Each rack provides 20.7 TB of HBM4 memory and up to 54 TB of LPDDR5X
  • GB300 NVL72采用Blackwell Ultra架构,在单个机柜内整合36颗CPU和72颗GPU。4
    其中在“ AI Factory ”展区,端出 NVIDIA GB300 NVL72 系统,采用 NVIDIA Blackwell Ultra 架构,整柜拥有 18 个运算节点,每个节点搭载 2 颗 CPU 与 4 颗 GPU,将 36 颗 CPU 与 72 颗 GPU 整合于单一机架内。

已有应用

  • 工业富联2025年中期说明会确认GB200服务器已有出货。5
    During the Q&A session, Chairman Brand Cheng addressed investor questions and confirmed that GB200 server efficiency and yield had improved significantly since the first quarter, with shipments accelerating and expected to maintain strong momentum in the third quarter.
  • 超微2026年9月23日宣布已出货搭配直接液冷系统的英伟达Vera Rubin NVL72机柜。3
    SAN JOSE, Calif., Sept. 23, 2026 /PRNewswire/ -- Super Micro Computer, Inc. (NASDAQ: SMCI), an AI, Enterprise, Storage, and 5G/Edge IT Total Solution Provider, is now shipping NVIDIA Vera Rubin NVL72 racks integrated with Supermicro's Data Center Building Block Solutions® (DCBBS) and its direct liquid cooling stack (DLC-2).
  • 超微的Vera Rubin NVL72可扩展单元方案在16个计算机柜中配置1152颗NVIDIA Rubin GPU。3
    The DCBBS Blueprint based on one Vera Rubin NVL72 Scalable Unit provides 1,152 NVIDIA Rubin GPUs with 331 TB of HBM4 memory across 16 compute racks, with balanced cooling capacity, power delivery, high-performance storage, context memory storage, and networking.
  • Flexcompute使用搭载英伟达加速计算的戴尔PowerEdge XE系列服务器运行AI物理仿真。6
    At Flexcompute, Dell PowerEdge XE-Series servers with NVIDIA accelerated computing power AI-driven physics simulations that help engineers complete simulations in minutes instead of days.
  • 纬创德州D1工厂已在美国制造并量产英伟达GB300 Grace Blackwell Ultra Superchip。7
    FORT WORTH, July 21, 2026 — Wistron Corporation ("Wistron") celebrated the grand opening of its D1 AI smart facility in Fort Worth, Texas, the site where the first NVIDIA GB300 Grace Blackwell Ultra Superchip was built and mass-produced in the United States.
  • 超微Super AI Station采用NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip,可桌边或5U机架部署。8
    First unveiled at GTC Spring in March 2026 and making its Asia debut at COMPUTEX, Supermicro’s Super AI Station is a tower-sized local AI supercomputer powered by the NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip that can be deployed deskside or mounted in a 5U rack slot for quick and easy upscaling.
  • 富士康在2025年鸿海科技日展示了GB300 NVL72 AI服务器机柜。4
    其中,“ AI Factory ”展示 NVIDIA GB300 NVL72 AI 服务器机柜系统、AI 资料中心高效散热全方案及内部高速连接器等 AI 基础设施关键解决方案,同时呈现超级算力中心与生态系统平台,全面展现在 AI 领域研发、制造与应用的垂直整合实力。
  • 富士康的智慧制造、智慧EV与智慧城市平台使用NVIDIA GB300 NVL72提供AI算力。4
    作为 AI 工厂的核心运算引擎,NVIDIA GB300 NVL72 为富士康三大智慧平台——智慧制造、智慧 EV 与智慧城市——提供强大 AI 算力,驱动富士康在 AI 时代持续创新。

为何此时

  • 工业富联2026年第三季度AI-ASIC服务器出货量同比增长1.5倍,AI-GPU机柜出货量同比增长70%。1
    其中,AI-ASIC服务器出货量同比增长1.5倍、环比增长80%,AI-GPU机柜出货量同比增长70%、环比增长15%。
  • 超微2027财年第一季度营收初估上调,主要原因是AI平台需求增加。9
    Revenues for the first quarter of fiscal year 2027 are estimated to be in the range of $18.5 billion to $19.5 billion versus guidance of $14.5 billion to $15.5 billion, primarily due to increased demand for AI platforms.

主要参与者

  • 工业富联在2026年第三季度已出货AI-ASIC服务器及AI-GPU机柜。1
    其中,AI-ASIC服务器出货量同比增长1.5倍、环比增长80%,AI-GPU机柜出货量同比增长70%、环比增长15%。
  • 戴尔AI方案团队使用PowerEdge XE9680与PowerEdge R760xa两类AI优化服务器测试单节点模型训练。2
    The Dell AI Solutions team used two classes of AI-optimized servers to test a deployment of single-node training models, the PowerEdge XE9680 and the PowerEdge R760xa.
  • 超微将NVL72机柜与完整液冷系统一起测试验证。3
    Supermicro tests and validates every rack with the full liquid cooling stack, speeding up time-to-online when deployed.
  • GB300 NVL72使用富士康自制的机壳、水冷头、高速电缆、水冷盘、分歧管和风扇。4
    GB300 NVL72 结合多项富士康自制的关键零组件,包含 NVIDIA MGX rack 机壳、水冷头、高速电缆、水冷盘、分歧管与风扇,充分展现富士康在垂直整合设计与精密制造方面的强大实力。

相关卡片

来源

  1. fii-foxconn.com · 高基数下持续高增长 工业富联前三季度营收净利均超去年全年_FII
  2. dell.com · InfoHub | Dell USA · 2026-10-06
  3. ir.supermicro.com · Supermicro Now Shipping NVIDIA Vera Rubin NVL72 Racks · 2026-09-23
  4. 鸿海 · 鸿海科技日逾200件展品 AI Factory展现生态系垂直整合能力 - 富泰华工业(深圳)有限公司 · 2025-11-24
  5. fii-foxconn.com · Fii 2025 Interim Results Briefing: AI-Driven Growth Accelerates Momentum_FII
  6. dell.com · Built for Innovation: The Full Stack Behind Dell AI Leadership | Dell USA · 2026-09-15
  7. wistron.com · Wistron · 2026-07-21
  8. 超微电脑 · Supermicro Announces General Availability of the NVIDIA DGX GB300-Powered Super AI Station at COMPUTEX 2026 · 2026-06-01
  9. ir.supermicro.com · Supermicro Provides First Quarter of Fiscal Year 2027 Preliminary Business Update · 2026-10-20