How to Set Up DGX Spark for Headless Access Using ThinLinc (VDI / Linux Remote Desktop)

NVIDIA DGX Spark is a compact AI system designed for AI development, inference, and GPU-accelerated workloads. While you can manage the system using SSH, some workloads require a complete graphical Linux desktop.

ThinLinc provides a remote desktop environment that allows users to securely access graphical Linux sessions from another computer. It also supports multiple simultaneous users, persistent sessions, browser-based access, and native desktop clients.

In this tutorial, you will learn how to configure an NVIDIA DGX Spark for single-user and multi-user headless access using ThinLinc.

Prerequisites

Before starting, make sure you have:

  • An NVIDIA DGX Spark with DGX OS installed.

  • A working network connection.

  • A user account with sudo privileges.

  • SSH access to the DGX Spark.

  • The ARM64 ThinLinc Server package.

  • A local Windows, Linux, or macOS computer for remote access.

DGX Spark is an ARM64 platform, so you must use an ARM64-compatible ThinLinc Server build. Cendio has built and tested ARM64 ThinLinc Server software on NVIDIA DGX Spark.

Note: Obtain the current ARM64 ThinLinc Server beta/test package from Cendio before starting the installation. Do not install an amd64 ThinLinc Server package on DGX Spark.

Step 1 - Verify the DGX Spark System

Before installing ThinLinc, verify the operating system, CPU architecture, NVIDIA GPU, and CUDA environment.

1. First, check the CPU architecture.

uname -m

You should see:

aarch64

You can also verify the Debian package architecture.

dpkg --print-architecture

The output should be:

arm64

2. Check your DGX OS version.

cat /etc/os-release

DGX Spark uses an Ubuntu-based NVIDIA software environment designed for the platform.

3. Verify that the NVIDIA GPU is available.

nvidia-smi

The command should display information about the NVIDIA GPU and installed driver.

4. Check the CUDA toolkit.

nvcc --version

You can also verify the CUDA directory.

ls -l /usr/local/cuda

DGX Spark already includes the NVIDIA driver and CUDA software stack, so you generally do not need to reinstall them before setting up ThinLinc.

5. Finally, update the package index.

sudo apt update

Install some basic utilities.

sudo apt install -y openssh-server unzip wget curl dbus-x11 x11-xserver-utils

Make sure SSH is running.

sudo systemctl enable --now ssh

Check its status.

sudo systemctl status ssh

Once the DGX Spark environment is ready, you can install a desktop environment for ThinLinc.

Step 2 - Install XFCE Desktop Environment

ThinLinc requires a graphical desktop environment to provide remote Linux sessions. In this guide, we will use XFCE. It is lightweight and consumes fewer system resources than a full desktop environment, which makes it suitable when several users access the DGX Spark simultaneously.

1. Install XFCE and its additional utilities.

sudo apt install -y xfce4 xfce4-goodies

2. Verify that XFCE is installed.

command -v startxfce4

You should see:

/usr/bin/startxfce4

3. Next, create a dedicated ThinLinc administrator account.

sudo adduser dgxadm

Enter a secure password when prompted.

4. Add the user to the sudo group.

sudo usermod -aG sudo dgxadm

5. Verify the user.

id dgxadm

6. Switch to the new user.

su - dgxadm

You will use this account to install and test ThinLinc in the next step.

Step 3 - Install ThinLinc Server

With XFCE installed, you can now install the ThinLinc Server. ThinLinc uses a Master and Agent architecture. The Master accepts connections and manages sessions, while the Agent hosts the graphical sessions. A single machine can perform both roles.

For one DGX Spark system, select Master during setup. ThinLinc will configure the DGX Spark as a standalone server that runs both vsmserver and vsmagent.

1. Download and copy the ARM64 ThinLinc Server package to the DGX Spark.

For example:

scp thinlinc-arm64-server.zip \
dgxadm@DGX-SPARK-IP:~/

Replace DGX-SPARK-IP with the IP address of your DGX Spark.

2. Log in to the DGX Spark.

ssh dgxadm@DGX-SPARK-IP

3. Extract the ThinLinc Server archive.

unzip thinlinc-arm64-server.zip

The actual archive name may differ depending on the ARM64 test release supplied by Cendio.

4. Enter the extracted directory.

cd tl-*-server

5. Start the ThinLinc installer.

sudo ./install-server

ThinLinc installation has two stages. The install-server utility installs the appropriate packages, and tl-setup configures the server.

After the package installation completes, you will see a prompt similar to:

Run ThinLinc setup now [Yes/no]?

Type:

Yes

6. Accept the ThinLinc license when prompted.

Do you accept the terms of the license agreement [yes/No]?

Type:

yes

7. The setup wizard will ask which type of ThinLinc server you want to configure.

Server type [Master/agent]?

Select:

Master

Note: Selecting Master on the first ThinLinc machine creates a standalone installation. The same DGX Spark runs both vsmserver and vsmagent. Select Agent only when adding another server to an existing ThinLinc cluster.

8. If the setup wizard asks whether missing packages should be installed automatically, select Yes.

9. If prompted for Network File System support for local drive redirection, select Yes.

10. Allow the setup wizard to install required GTK+, Python, and other dependencies if prompted.

11. Configure the externally reachable address. Use the DGX Spark IP address or a hostname that remote users can resolve.

12. Enter the administrator email address when prompted.

13. Configure the ThinLinc Web Administration password. The Web Administration interface uses the username admin.

14. If ThinLinc detects AppArmor and asks whether to install its configuration, choose Yes.

15. After setup completes, verify the ThinLinc services.

sudo systemctl status vsmserver
sudo systemctl status vsmagent
sudo systemctl status tlwebaccess
sudo systemctl status tlwebadm

Each enabled service should show:

Active: active (running)

At this point, the DGX Spark is configured as a standalone ThinLinc server.

Step 4 - Verify ThinLinc Desktop Profiles

ThinLinc uses desktop profiles to decide which desktop environment should start when a user logs in. Because XFCE is installed, ThinLinc should make it available as a desktop option.

1. Verify XFCE again.

command -v startxfce4

2. Restart the ThinLinc Agent after installing or changing desktop environments.

sudo systemctl restart vsmagent

3. Confirm that the Agent starts normally.

sudo systemctl status vsmagent

When a user connects for the first time, ThinLinc should allow the user to start an XFCE session.

Step 5 - Configure the Firewall for ThinLinc

ThinLinc uses several ports for remote access and internal communication.

For a standalone DGX Spark where the Master and Agent run on the same machine, ports 9000 and 904 are internal ThinLinc service ports and normally do not need to be exposed to remote users.

1. Check the UFW firewall status.

sudo ufw status

2. Allow SSH.

sudo ufw allow 22/tcp

3. Allow ThinLinc Web Access.

sudo ufw allow 300/tcp

4. If you need remote access to the ThinLinc administration panel, allow port 1010 only from your management network.

sudo ufw allow from 192.168.1.0/24 to any port 1010 proto tcp

5. If UFW is currently disabled and you want to use it, enable it.

sudo ufw enable

6. List the firewall rules.

sudo ufw status numbered

Step 6 - Access DGX Spark Using ThinLinc Web Access

Now that ThinLinc is installed, you can test the remote desktop directly from a web browser. ThinLinc Web Access is an HTML5 client available through TCP port 300.

1. Find the DGX Spark IP address.

hostname -I

For example:

192.168.1.100

2. On your local system, open a web browser and go to: “https://DGX-SPARK-IP:300”

Your browser may display a security warning if the default installation uses a certificate that is not trusted by your device.

3. Enter your Linux credentials. For example:

Username: dgxadm
Password: ********

4. Click Log in. ThinLinc will start the XFCE desktop inside the browser.

5. Open a terminal inside the remote desktop and check the hostname and virtual display.

hostname
echo $DISPLAY

You should see the DGX Spark hostname and a virtual display allocated to your ThinLinc session.

Step 7 - Verify NVIDIA GPU Access

ThinLinc provides the remote graphical desktop, but CUDA applications launched from that desktop can still access the NVIDIA GPU.

1. Open a terminal inside your ThinLinc session and run:

nvidia-smi

The DGX Spark GPU should be detected.

2. Check CUDA.

nvcc --version

3. If PyTorch is installed, verify CUDA availability.

python3 -c "import torch; print(torch.cuda.is_available())"

A working CUDA environment should return:

True

4. Check the GPU detected by PyTorch.

python3 -c "import torch; print(torch.cuda.get_device_name(0))"

This confirms that GPU compute workloads started from the ThinLinc desktop can use the DGX Spark GPU.

You can use the same environment to run applications such as:

  • PyTorch

  • TensorFlow

  • TensorRT

  • CUDA tools

  • JupyterLab

  • AI inference workloads

  • Model development tools

  • NVIDIA containers

Note: GPU compute and desktop graphics acceleration are separate. A CUDA application can use the NVIDIA GPU even if the XFCE desktop itself uses software rendering.

Step 8 - Access DGX Spark Using ThinLinc Client

While Web Access works directly from a browser, the native ThinLinc Client generally provides a richer remote desktop experience. ThinLinc provides native clients for Linux, Windows, and macOS.

1. Download the appropriate ThinLinc Client for your local operating system from the ThinLinc download page.

For an Ubuntu client, install the downloaded package. For example:

sudo apt install ./thinlinc-client_*.deb

2. Launch ThinLinc Client from the Applications menu.

3. Enter the server and user information.

Server: DGX-SPARK-IP
Username: dgxadm
Password: ********

4. Click Connect. ThinLinc should open your XFCE desktop.

The native client also provides features such as clipboard integration, local drive redirection, audio redirection, connection tuning, and session reconnection. If you disconnect without logging out, the remote desktop session remains running and can be resumed later.

Step 9 - Configure Multi-User Access

ThinLinc is designed to host multiple independent Linux sessions. Each ThinLinc user should have a separate Linux account on the DGX Spark.

1. Create the first additional user.

sudo adduser alice

2. Create another user.

sudo adduser bob

3. Create a third user if you want to test three simultaneous sessions.

sudo adduser charlie

4. Verify the accounts.

id alice
id bob
id charlie

5. Connect as alice using one ThinLinc Client, connect as bob from another computer or browser, and connect as charlie from another client if required.

Each user should receive a separate XFCE desktop. The Master manages session allocation, while the Agent hosts the graphical sessions. On this standalone configuration, both components run on the same DGX Spark.

Step 10 - Test Multi-User GPU Access

You can now verify that different ThinLinc users can access GPU resources.

1. Connect as alice and run:

nvidia-smi

2. Connect simultaneously as bob and run the same command.

nvidia-smi

3. From an administrator terminal, continuously monitor GPU activity.

watch -n 1 nvidia-smi

4. Start GPU applications from different user sessions and observe the GPU process list where supported.

ThinLinc manages the remote graphical sessions, but it does not automatically divide GPU capacity among users. GPU resource management should therefore be planned separately if multiple users run demanding workloads at the same time.

Step 11 - Test Headless Access

Next, confirm that ThinLinc continues to work when no monitor is attached to the DGX Spark.

1. Shut down the DGX Spark.

sudo poweroff

2. Disconnect the monitor.

3. Power on the DGX Spark and wait for the operating system to start.

4. From your local computer, verify SSH access.

ssh dgxadm@DGX-SPARK-IP

5. Check the ThinLinc services.

systemctl is-active vsmserver
systemctl is-active vsmagent
systemctl is-active tlwebaccess

They should return:

active

6. Connect using ThinLinc Client or Web Access.

ThinLinc should create the XFCE session normally without a physical monitor because it hosts independent virtual graphical sessions rather than requiring users to work directly with the local console display.

Step 12 - Access ThinLinc Web Administration

ThinLinc includes a web-based administration panel for monitoring and managing the remote desktop environment.

1. On your local computer, open: “https://DGX-SPARK-IP:1010”

2. Enter the credentials configured during tl-setup.

Username: admin
Password: YOUR-ADMIN-PASSWORD

3. Sign in to the ThinLinc Web Administration interface.

You can use the interface to view information about:

  • ThinLinc servers

  • User sessions

  • System load

  • Licenses

  • Configuration

  • Active users

Note: For security, allow access to the Web Administration interface only from trusted administration networks.

Troubleshooting ThinLinc

ThinLinc Client Cannot Connect

First, verify SSH connectivity. If SSH does not work, fix network connectivity before troubleshooting ThinLinc.

ssh username@DGX-SPARK-IP
sudo systemctl status ssh
sudo systemctl status vsmserver
sudo systemctl status vsmagent
sudo ss -lntp | grep :22
sudo ufw status

ThinLinc Desktop Does Not Start

Verify that XFCE is installed and review the ThinLinc service journals.

command -v startxfce4
journalctl -u vsmagent -n 100 --no-pager
journalctl -u vsmserver -n 100 --no-pager

ThinLinc Web Access Does Not Open

Check the Web Access service, verify that port 300 is listening, and review the firewall rules.

sudo systemctl status tlwebaccess
sudo ss -lntp | grep :300
sudo ufw status

NVIDIA GPU Is Not Available

Test the GPU outside ThinLinc using SSH. If nvidia-smi fails over SSH, troubleshoot the DGX Spark GPU software stack before changing ThinLinc.

nvidia-smi
lsmod | grep nvidia

CUDA Works but OpenGL Uses Software Rendering

This can be normal. The ThinLinc desktop can use software OpenGL while CUDA applications still use the NVIDIA GPU.

OpenGL renderer: llvmpipe
CUDA available: True

Install and configure VirtualGL only if your graphical application requires accelerated OpenGL rendering.

Conclusion

In this tutorial, you configured an NVIDIA DGX Spark for headless remote desktop access using ThinLinc. You installed an XFCE desktop, configured the DGX Spark as a standalone ThinLinc Master and Agent, enabled browser and native client access, and verified access to the NVIDIA GPU from remote sessions.

You also configured multiple Linux accounts so several users can connect to the same DGX Spark with independent graphical desktops. Each user can disconnect from a ThinLinc session and reconnect later without terminating the applications running inside it.

For AI, CUDA, PyTorch, TensorRT, and other compute workloads, users can access the DGX Spark GPU directly from their ThinLinc sessions without additional graphics configuration. VirtualGL is only required when an application specifically needs GPU-accelerated OpenGL rendering.

You now have a headless ThinLinc environment that can be used as a single-user remote AI workstation or a shared multi-user DGX Spark system.

Sometimes I log directly onto the Spark, and play a game on Steam(there used to be quite a few hoops to jump through, buts it’s much easier now since there’s an official Steam arm build). Other than that though, I try not to add anything at all to the overhead. I want every byte of ram and cycle of compute to serve my inference.

What’s the advantage of using it like this, instead of just having it serve an inference api?