What Is Nvidia’s DGX Station? A New, Specialized Desktop Line for AI Work

What Is Nvidia’s DGX Station? A New, Specialized Desktop Line for AI Work What Is Nvidia’s DGX Station? A New, Specialized Desktop Line for AI Work

SAN JOSE, CALIF.—After introducing an AI-focused mini PC, dubbed “Project DIGITS,” at CES 2025, Nvidia is preparing to unleash a bigger cousin: a full-fledged tower desktop for AI workloads.

At the company’s GTC conference on Tuesday in San Jose, Jensen Huang, Nvidia’s CEO, announced the DGX Station, saying, “This is what a PC should look like.” (You can view an edited-highlights version of Huang’s epic keynote in the video above.)

Indeed, Huang showed a mockup of the innards of a sample DGX Station: a board resembling a traditional PC motherboard, but with certain components enlarged and in slightly different spots than you’d expect. He also confirmed the DGX Station board contains some PCI Express slots, enabling it to connect to an Nvidia graphics card. 

Nvidia DGX Station and DGX Spark

(Credit: Nvidia )


Not for Gamers: The DGX Station Is for AI Experts Only

Some of its interior may look relatable, but the DGX Station isn’t meant for consumers. Instead, Nvidia is marketing the product toward enterprise customers, such as researchers and software developers, looking to run intensive AI workloads. Indeed, the company dubs it a “desktop supercomputer.”

That’s because the product features a data-center-grade GB300 “Grace Blackwell Ultra” chipa type that Nvidia has traditionally referred to a “Superchip”that’s specifically designed for wrangling with cutting-edge AI models. This combined CPU/GPU contains 72 CPU cores, along with 784GB of unified memory shared between the processing and graphics portions. Pro AI devs would apply this kind of massive memory loadout to large training and inferencing tasks.

Nvidia DGX Station

(Credit: Michael Kan)

An extremely high-speed interconnect, dubbed “NVLink-C2C,” will join the Grace CPU and Blackwell GPU portions of the chip to enable all this data flow. Likewise, for external connectivity, a high-speed network card, the ConnectX-8 SuperNIC, will enable the kinds of transfers required for chaining multiple DGX Stations together, or to perform high-speed local-network data movement of the kind these workflows require.

Nvidia didn’t share how much a DGX Station will cost, but we wouldn’t be surprised if a single unit went for $10,000 or more, considering how expensive Nvidia’s other enterprise-grade GPUs can be. Expect the DGX Station to arrive later this year from Nvidia’s manufacturing partners, all familiar desktop-workstation faces, among them Asus, Boxx, Dell, HP, and Supermicro.


DIGITS to Sparks: Nvidia’s AI Mini PC Gets a New Name

In the meantime, the company is starting to take reservations for its AI-developer-centric mini PC, formerly known as “Project DIGITS.” At GTC, Nvidia revealed it had renamed the product “DGX Spark.”

Nvidia DGX Spark

(Credit: Michael Kan)

Like the DGX Station, the Spark mini PC isn’t designed for your typical PC buyer, but for software developers, researchers, and students interested in developing generative AI programs. The product features Nvidia’s GB10 Superchip, which can support AI models up to 200 billion parameters in size—similar to OpenAI’s GPT-3.

Recommended by Our Editors

Nvidia notes that the Blackwell GPU in the DGX Spark can support up to 1,000 TOPS (trillions of operations per second) to support training and inferencing needs. Also, the CPU and GPU will be equipped with Nvidia-specific interconnect technology that should allow for as much as five times the bandwidth between the chip elements and its memory, versus conventional PCI Express 5.0.

Nvidia DGX Spark

(Credit: Michael Kan)

Nvidia detailed a good bit more about DGX Spark’s inner specs on its site once the GTC keynote wrapped up. The Spark will feature 128GB of unified LPDDR5x memory and have Nvidia’s AI software tools preinstalled. The “Grace” portion of the GB10 will have 20 Arm cores. The system will also support a form of the ConnectX networking mentioned earlier, here ConnectX-7, for tethering two Sparks together to work on extremely large AI models. The system will come equipped with a 1TB or 4TB SSD, and have provision for Wi-Fi 7, USB (four USB Type-C, to 40Gbps), and Bluetooth connectivity, much like any other high-end desktop.

Nvidia DGX Spark Interior

(Credit: Nvidia)

DGX Spark will run by connecting to a regular electrical socket. But don’t expect the mini PC to play Windows games. The DGX Spark comes with Nvidia’s DGX OS, the company’s custom version of Ubuntu Linux.

The pre-order page for the DGX Spark

Nvidia’s pre-order page for the DGX Spark. (Credit: Nvidia)

Nvidia says that models of the DGX Spark from its partners will start at $2,999. (The pre-order page teased a $2,999 Asus Ascent GX10 configuration with a 1TB SSD.) An Nvidia-branded Founders Edition version of the DGX Spark with 4TB of storage will go for $3,999, and Nvidia is also offering a bundle of two 4TB DGX Sparks with a high-speed connection cable, for $8,049. PC vendors including Asus, Dell, HP, and Lenovo are all preparing DGX Spark and DGX Station products under their own brands.

Get Our Best Stories!

Sign up for What’s New Now to get our top stories delivered to your inbox every morning.

This newsletter may contain advertising, deals, or affiliate links.
By clicking the button, you confirm you are 16+ and agree to our
Terms of Use and
Privacy Policy.
You may unsubscribe from the newsletters at any time.

Newsletter Pointer

About Michael Kan

Senior Reporter

Michael Kan

I’ve been working as a journalist for over 15 years—I got my start as a schools and cities reporter in Kansas City and joined PCMag in 2017.


Read Michael’s full bio

Read the latest from Michael Kan

Add a comment

Leave a Reply

Your email address will not be published. Required fields are marked *

This site uses Akismet to reduce spam. Learn how your comment data is processed.

Keep Up to Date with the Most Important News

By pressing the Subscribe button, you confirm that you have read and are agreeing to our Privacy Policy and Terms of Use