NVIDIA DGX Station: What Research Teams Can Run Locally

EPRI's researchers used a deskside DGX Station to run one timestep of a 50-member GenCast weather ensemble for electric-grid planning. In the test NVIDIA documents, that run took about 150 seconds; the case study estimates more than ten hours on a CPU-only server. That result belongs to EPRI's setup, but it shows why a research team might want this much compute within reach rather than waiting for a shared cluster slot.

A DGX Station purchase still has to answer ordinary questions: Will our model fit at the precision we use? Can our team share the machine without slowing each other down? Can the room power and cool it? Here is how to answer those questions before requesting a quote.

At Sungkyunkwan University School of Medicine, researchers load protein-structure databases, including an AFDB50 collection of more than 60 million structures, into memory for fast homology searches. Cornell computer-science professor Kilian Weinberger reports using the memory to train multibillion-parameter language models without waiting for scheduled compute. These are different jobs, yet both benefit from having large, accessible memory next to the researcher.

The work is not limited to language models. Carnegie Mellon University's Robotics Institute reports fine-tuning a drone-detection model on Station, then deploying it to a Jetson Orin NX for real-time inference. EPRI uses Station for weather prediction and downscaling research that informs electric-grid planning. The useful question for a buyer is which stalled experiment, slow data step or queue of jobs the system would remove for their own team. NVIDIA's examples are customer-reported use cases, not promised results for another lab.

DGX Station pairs one Blackwell Ultra GPU with a 72-core Grace CPU through NVLink-C2C. The GPU has 252 GB of HBM3e; the CPU has 496 GB of LPDDR5X. Together they form a coherent 748 GB address space, but data in CPU memory does not run at the same speed as data in GPU HBM. A model's weights, context cache, training states and active users all compete for space and bandwidth.

NVIDIA lists up to 20 PFLOPS of FP4 compute and support for models up to one trillion parameters. Those are platform limits under particular conditions, not a speed or model-fit guarantee. Ask the team to record the exact checkpoint, precision or quantization, context length, batch size and target latency. If the decision depends on a large model running without harmful offload or serving several people at once, request a test with those settings rather than extrapolating from parameter count.

NVIDIA rates Station at 1,600 W and requires a 20 A circuit. Have facilities check the circuit, outlet, heat load, clearance, delivery path and service access against the OEM unit in the quote. A deskside machine still needs an installation plan.

The shipping Linux configuration uses Ubuntu with NVIDIA AI Developer Tools. ASUS lists that environment for its ExpertCenter Pro ET900N G3. NVIDIA says DGX Station for Windows is expected from OEMs in December 2026; a current Linux ASUS unit should not be treated as a Windows purchase. Check Arm support for your containers, licensed software and model runtime before placing the order. For a shared machine, also decide who controls user access, updates, backups and storage.

Station includes ConnectX-8 networking, with up to 800 Gb/s of platform bandwidth, and NVIDIA now supports linking two Linux Stations through NVIDIA Sync. That peak does not specify the cables or switch in your room. Name the network it must join and ask the supplier to list the required transceivers and links. If visualization or simulation is part of the job, NVIDIA allows one additional RTX PRO GPU on supported configurations; confirm the card and OEM support in writing.

ASUS says its ExpertCenter Pro ET900N G3 is available to order worldwide and publishes several configurations. The base two-SSD version lists two 2 TB NVMe drives; other versions add drives, QSFP transceivers or an RTX PRO card. The Uniqcli catalog page below is a useful starting point, but its listing does not establish immediate stock or the lead time for your chosen configuration.

Ask for the manufacturer's full configuration, usable storage after any mirroring, required network parts, OS image, warranty, support entitlement and a dated delivery estimate. If a lab needs a different storage layout or graphics option, put that requirement in the request before a price is assembled.

MSI's XpertStation WS300 is another DGX Station-architecture option in the Uniqcli catalog. MSI's manufacturer page calls the listed WS300 part a barebones configuration, so request a complete bill of materials and clarify the installed storage, software and support before comparing it with the ASUS system.