由于人工智能产业的快速发展，数据中心的建设速度远远无法满足市场需求。因此，从 2027 年起，Nvidia 及其合作伙伴将推出经过全面预配置（包括布线、电气连接等）的 Vera Rubin 机架系统，这些系统可直接用于模块化的数据中心设施中。这些被称为 DSX 模块的计算设备本周在 GTC Berlin 大会上正式亮相；从功能上看，它们相当于“可移动的数据中心”。
Datacenters are taking too long to bring online to fill the demand of the AI boom. So starting in 2027, Nvidia and its partners will offer its Vera Rubin rack systems fully plumbed, wired, and ready for deployment in modular bit barns. The containers, which Nvidia calls DSX pods, were unveiled at GTC Berlin this week and are essentially the datacenter equivalent of a trailer home.
这些模块由 HPE、Dell、Schneider Electric 等合作伙伴共同开发，每个 DSX 模块都配备了人工智能应用所需的所有计算资源、互联设施以及电气和管道系统。用户只需将模块放置在准备好的数据中心场地中，连接电源、网络和冷却系统，即可立即开始运行并产生收益——至少这是 Nvidia 的宣传说法。
Built by partners like HPE, Dell, Schneider Electric, and others, each DSX pod will come fully furnished with all the compute, interconnects, electrical, and plumbing a growing AI factory might need. Just drop the pod on a prepared datacenter site, hook up power, networking, and cooling, and you'll be churning out tokens and raking in profit in no time.
每个 DSX 模块重约 28 公吨，内部包含 7 个 Vera Rubin NVL72 机架，每个机架可安装 504 个 GPU 和 252 个 CPU。据 GPUzilla 评估，这些模块可将数据中心的部署时间从原来的 6 个月缩短至仅 2 周，因为所有部件在出厂前都会经过严格的质量检测（包括连接方式和组件功能的验证）。这样一来，一旦发现任何问题，问题就能在工厂阶段就被及时发现，而无需用户在安装后花费大量时间进行排查。
Or that's Nvidia's sales pitch anyway. Inside the 28 metric ton containers you'll find seven Vera Rubin NVL72 racks packing 504 GPUs and 252 Vera CPUs. GPUzilla says the pods will cut datacenter deployment times from six months to as little as two weeks because it will validate every fitting, cable, and component before it leaves the factory. The idea being if there is something wrong, it'll get caught there rather than after customers have gone through the trouble of hooking everything up.
与传统的数据中心建筑类似，Nvidia 的 DSX 模块也具有高度的可定制性：用户可以根据自身需求选择不同的合作伙伴（如 HPE 或 Dell、Eaton 等）来定制这些计算设备。如果需要更多计算资源，只需再增加一个模块即可（相当于扩展了计算能力）。这种灵活性无疑会让其他云计算服务提供商感到羡慕。首批采用 Nvidia Vera Rubin 机架系统的 DSX 模块预计将于明年年初投入生产，这体现了 Nvidia 为应对供应链挑战而持续付出的努力——他们希望在这些挑战影响到公司收入之前就将其解决。
And just like a trailer home, Nvidia's DSX pods will be entirely customizable. Prefer HPE over Dell, or Eaton over Schneider? No problem, Nvidia has roped all its buddies into building these compute containers. And if you ever need more compute, just add another trailer and you've got yourself a double wide, which will surely be the envy of all your neocloud neighbors. The first DSX pods powered by Nvidia's Vera Rubin rack systems should roll off the production line early next year, and reflect the company's ongoing effort to quash supply chain challenges before they start eating into revenues.
在柏林举行的 GTC（GPU Technology Conference）上，Nvidia 公布了专为“无水数据中心”设计的新 DSX（Data Center Solutions）技术规格。这些数据中心采用闭环冷却系统及干式冷却装置，而非传统的需要大量水资源的冷却塔。这一举措延续了该公司在今年早些时候在 AI Infra Summit 上宣布的承诺：Nvidia 承诺通过优化技术来提升计算性能，同时让数据中心能够根据电网状况的变化灵活调整自身的工作负载。
Also at GTC Berlin, Nvidia laid out out new DSX specs for waterless datacenters, which use closed loop coolants and dry coolers rather than water-thirsty cooling towers. This effort builds on work announced at the AI Infra Summit earlier this year, when Nvidia touted its effeorts to squeeze more performance out of less juice, while also allowing bit barns to spin down their workloads in response to changing grid conditions. ®