Nvidia 推出 DSX POD 集装箱式 AI 工厂参考设计

Nvidia wants new AI capacity to show up at the data center fully built. At its GTC Berlin conference on Wednesday, the company introduced DSX POD, a reference design for shipping containers that partners fill with Vera Rubin racks, networking, and liquid cooling before they leave the factory.

Installing and wiring up racks in a conventional data center can take months, Nvidia says. But a DSX POD can be ready to deliver tokens in as few as 14 days, as long as the site already has the necessary power, cooling, and network connections (which are, to be fair, often bottlenecks, too).

The idea here is for data center operators to get to value faster, because, as Nvidia founder and CEO Jensen Huang said in a statement, “An AI factory creates value when it starts producing tokens — not when construction is complete.”

What’s in the box

Each container holds seven Vera Rubin NVL72 racks, or 504 Rubin GPUs. Ian Buck, Nvidia’s vice president of hyperscale and high-performance computing, said in a press briefing ahead of the announcement that a single unit packs about 10 million components and weighs around 28 metric tons.

The containers also come with roughly 20 kilometers of copper cabling and 17 kilometers of scale-out networking preinstalled and tested. Once a container arrives, crews set it in place, connect it to its neighbors, and run the final burn-in tests, Buck said.

Water use

Buck also noted that Nvidia designed the system to use no water for cooling. Vera Rubin can run on coolant as warm as 45 degrees Celsius, Nvidia says, so dry coolers, which work pretty much like a car’s radiator, can do the work instead of the evaporative cooling towers that are still the standard for many data centers today.

In a separate DSX announcement, Nvidia explains that operators in some climates may still need chillers on the hottest days. For a 2-gigawatt deployment, the company estimates that dry cooling can save about 54 million liters of water a day compared with cooling towers.

Data centers in a box

This server-in-a-container may feel a bit familiar, and indeed, tech companies do love to use containers, even if it’s sometimes more of a publicity stunt than a solution to a real problem.

Back in 2006, Sun Microsystems unveiled Project Blackbox, a data center in a 20-foot container, for example. A decade later, then-AWS CEO Andy Jassy ended his re:Invent 2016 keynote by having an 18-wheeler drive Snowmobile, a 45-foot container that could haul 100 petabytes of data, onto the stage.

Microsoft and AWS have since built containerized data centers for disaster zones and military missions.

All of those, however, are different from what Nvidia is doing here. DSX PODs are basically prefab parts that can be slotted into a data center. Buck described them as the building block for new AI factories.

Armada, which already sells containerized edge data centers, is one of 17 partners building them, alongside Bull, Dell, HPE, Schneider Electric, and Vertiv.

Nvidia says partners will start delivering DSX POD systems worldwide in early 2027.

The post Nvidia wants AI factories built in real factories appeared first on The New Stack.