英伟达在OCP峰会开放Vera Rubin机架设计，联手思科、Meta、谷歌推「AI工厂」标准
圣何塞——英伟达（Nvidia）10月12日在圣何塞开幕的开放计算项目（OCP）全球峰会首日发布一系列硬件设计、网络技术与行业合作，目标是加快其所称的“AI工厂”——即为人工智能负载优化的专用数据中心——的建设速度。
最受关注的是开放动作：英伟达把第三代Vera Rubin MGX机架与托盘设计贡献给OCP，涵盖供电、散热与机械系统，使其模块化AI超级计算机的架构可被整个行业采用。英伟达AI工厂解决方案高级经理乔·德莱尔（Joe Delaere）介绍了该架构的三项创新：液冷计算托盘取消了软管、线缆与风扇，简化全生命周期的组装与维护；母线排加入液冷后，可在紧凑体积内支持5000安培电流，提高算力与功率密度；新推出的MQD与MQDB两种托盘内液冷快接头，旨在降低泄漏风险、让冷板更换更容易。
产品方面，英伟达宣布为智能体AI负载设计的Vera CPU MGX机架已经开始出货。英伟达称，按行业基准测试，Vera CPU在脚本执行、压缩与代码编译等负载上的性能是上一代Grace CPU的两倍；Meta则报告其在使用DCPerf基准测试的超大规模云负载中也测得两倍的性能提升。该机架提供面向智能体AI、推理与存储等不同负载的多种配置。
网络方面，英伟达与思科联合推出USFP超高密度小型可插拔模块，号称首款为400Gb基础网络设计的可插拔模块：单通道带宽从此前上限200Gb翻倍至400Gb，每200Gb能效提升1.3倍，带宽密度是现有200Gb方案的四倍，可支持1.6Tb/s高密度网络，最高32条电气通道、总容量12Tb/s。同时，Meta将在其AI工厂的Vera Rubin部署中采用英伟达Spectrum-6以太网交换机，硬件故障检测与流量重路由速度比软件方案快30倍；谷歌的Falcon传输技术也已获得英伟达Spectrum-X以太网平台支持。
标准方面，AMD、谷歌、Meta与英伟达联合发布《开放数据中心规范》v0.7，为下一代AI工厂的供电架构、冷却基础设施与物理布局提供蓝图；谷歌、Meta、微软与英伟达还共同推出OCP电网韧性规范，要求大型AI基础设施在电网扰动时自动响应、在不到两秒内恢复至接近满负荷运行。
德莱尔在峰会上说：“现在AI行业真正的瓶颈，是全世界能以多快的速度制造和部署AI工厂。”
来源：据NewsBytes PH 10月13日对OCP全球峰会首日英伟达发布的报道。
Nvidia Opens Vera Rubin Rack Designs at OCP Summit, Pushes 'AI Factory' Standards With Cisco, Meta, Google
SAN JOSE — Nvidia on October 12, the opening day of the Open Compute Project (OCP) Global Summit 2026 in San Jose, California, unveiled new hardware designs, networking technologies and industry partnerships aimed at accelerating construction of what the company calls "AI factories" — data centers optimized for artificial intelligence workloads.
At the center of the announcements, Nvidia contributed its third-generation Vera Rubin MGX rack and tray designs — covering power, cooling and mechanical systems — to OCP, making the architecture of its modular AI supercomputer openly available. Joe Delaere, Nvidia's senior manager of AI Factory Solutions, highlighted three innovations: a liquid-cooled compute tray that eliminates hoses, cables and fans, simplifying assembly and maintenance; liquid cooling added to the busbar, allowing it to support 5,000 amperes in a compact size for higher compute and power density; and two new in-tray liquid-cooling connectors, the MQD and MQDB, designed to reduce leakage risks and make cold plates easier to replace.
On products, Nvidia announced that its Vera CPU MGX rack, designed for agentic AI workloads, is now shipping. According to Nvidia, the Vera CPU delivers twice the performance of its predecessor, the Grace CPU, on workloads such as scripting, compression and code compilation, based on industry benchmarks; Meta reported a twofold performance improvement in hyperscale cloud workloads using its DCPerf benchmark. The modular rack comes in several configurations for agentic AI, inference and storage.
In networking, Nvidia and Cisco introduced the Ultra-dense Small Form Factor Pluggable (USFP) module, described as the first pluggable module designed for 400Gb base fabrics. It doubles per-lane bandwidth from the previous 200Gb limit to 400Gb, improves power efficiency by 1.3 times per 200Gb, and provides four times the bandwidth density of existing 200Gb solutions, supporting 1.6Tb/s network fabrics and up to 32 electrical lanes for 12Tb/s of capacity. Meta is adopting Nvidia's Spectrum-6 Ethernet switches for Vera Rubin deployments, allowing multiplane networks that detect hardware failures and reroute traffic 30 times faster than software-based alternatives, and Google's Falcon transport technology is now supported on Nvidia's Spectrum-X Ethernet platform.
On standards, AMD, Google, Meta and Nvidia published the Open Data Center Specification v0.7, a blueprint for next-generation AI factories covering power architecture, cooling infrastructure and physical layout. Google, Meta, Microsoft and Nvidia also introduced the OCP Grid Resilience Specification, under which large-scale AI infrastructure would automatically respond to utility grid disturbances and restore operations to near-full load in under two seconds.
"The real bottleneck for the AI industry right now is how fast the world can manufacture and deploy AI factories," Delaere said.
Source: NewsBytes PH, October 13, reporting on Nvidia's OCP Global Summit announcements.
图片来源：英伟达圣克拉拉总部，Coolcaesar 摄，CC BY-SA 4.0（Wikimedia Commons）
