Earlier this year, NVIDIA caused quite a stir in the data-science community when they announced the DGX Spark: a desktop-sized AI supercomputer. It’s aimed at developers and data scientists who need to prototype, tune, and run AI models locally without constantly relying on expensive cloud hardware.

At the heart of the DGX Spark is the new GB10 chip, which combines NVIDIA’s 20-core Grace ARM CPU with the Blackwell GPU. Together, they deliver roughly 1 petaflop of AI compute and come paired with 128 GB of memory and 4 TB of fast local storage. The system ships with Ubuntu Linux 24.04, which immediately appealed to my inner UNIX/Linux fanatic.

As someone responsible for developing curriculum for our data-science program, I often need to create, train, optimize, and evaluate models… some of which are computationally demanding. Tasks like bootstrap aggregating or large-scale model evaluation can bring my laptop to its knees. So having workstation-class compute on my desk is a big deal. While the official NVIDIA DGX Spark is hard to obtain in Canada, several vendors ship the exact same machine under their brand (e.g., Dell, Lenovo, Asus, etc.). In my case, I went with the Dell Pro Max with GB10.

The setup process

You can set up the system in two ways:

1. Connect a monitor, keyboard, and mouse and go through the graphical setup wizard, or
2. Power it on and join the Wi-Fi hotspot printed on the sticker from another PC, then visit http://hostname.local from any browser on that PC where hostname is the hostname printed on the same sticker.

I chose option 2 because I wanted to see if the system could be set up and used completely headless (i.e., without a monitor, keyboard, or mouse).

During setup, you create your user account and choose whether to enable plain SSH or use the NVIDIA Sync app (which uses SSH under the hood). As someone who has been administering UNIX/Linux servers for years, I naturally chose regular SSH.

The graphical experience

Even though my plan was to run headless, I did plug in a monitor and input devices to get a feel for the desktop. The system is, quite unsurprisingly, extremely responsive. It runs a GNOME desktop on X.org (likely due to the ongoing challenges between Wayland and NVIDIA’s drivers).

The default neon-noir wallpaper on the Dell (below left) immediately gave me early-1990s retrowave vibes. It was reminiscent of 1980s and 1990s SGI and Sun UNIX workstation ads (below right). Those UNIX workstations popularized powerful RISC CPUs… and fittingly, the GB10 is itself a modern RISC design running Linux (which is basically UNIX).

The app selection includes the expected NVIDIA Settings tool and a shortcut to the NVIDIA DGX Dashboard, a web portal where you can install updates, change the hostname, view CPU/GPU performance, and – most importantly – launch JupyterLab, the go-to environment for data scientists. That’s where we run our Python code, view our outputs and graphs, and store our workflow notes.

The DGX Dashboard runs on port 11000, and JupyterLab runs on port 11002. This means you don’t actually need a desktop session at all!

Using the system headless

I rebooted the machine, unplugged the peripherals, and confirmed I could log in and obtain a command line shell via ssh user@hostname from my laptop (mDNS handles the hostname resolution on local networks).

To access the DGX Dashboard and JupyterLab from my laptop, I created two SSH tunnels:

ssh -f -N -L 11000:localhost:11000 user@hostname
ssh -f -N -L 11002:localhost:11002 user@hostname

Then I opened my browser and navigated to http://localhost:11000 to open the DGX Dashboard. Clicking Open in Browser on that screen launched JupyterLab in a new tab (http://localhost:11002):

Transferring data to a headless system is straightforward using the SSH copy command (scp). For example, to copy lala.txt in the current directory on my laptop to my home directory on the Dell Pro Max, I could run scp lala.txt user@hostname:/home/user command.

So yes, the entire system works flawlessly headless, which is exactly how I plan to use it.

Performance test!

For a quick benchmark, I grabbed a dataset and ran a Python script in JupyterLab that took roughly 30 minutes on my laptop last week (I have a feeling it was probably longer than that).

On the GB10 workstation? I couldn’t accurately measure the difference, because it completed instantly.

This bodes well for the months of data science work ahead ;-)

Final thoughts

The Grace Blackwell platform is, in a word, liberating. It gives data scientists the freedom to experiment locally at supercomputer speeds. In a way, it feels like a return to the classic UNIX workstation era, when SGI and Sun machines put serious compute power directly on a developer’s desk. Grace Blackwell feels like that idea reborn for the AI age: powerful compute on your desk, capable of running entire model-development workflows without ever touching the cloud. It’s a retro UNIX workstation with modern firepower.