Requirements for Edge AI on NVIDIA Jetson
This page describes the requirements for running Edge AI workloads on an NVIDIA Jetson device managed by Palette.
Deployment Mode
Palette manages the Jetson device as an Edge host registered in agent mode. You bring your own operating system, install the Palette agent on the device, and register it with Palette.
Appliance mode is not available on ARM64 devices. Provision ARM64 hosts, including Jetson devices, using agent mode. Refer to Edge Hardware Requirements for the current ARM64 support statement.
Network Requirements
In agent mode, the Palette agent on the device opens an outbound connection to Palette, and Palette never connects inbound. A Jetson host on a private network behind Network Address Translation (NAT) does not need inbound firewall rules or a Private Cloud Gateway (PCG).
The device requires outbound HTTPS (TCP 443) access to the following:
-
The Palette SaaS endpoint,
console.spectrocloud.com. -
The Palette image registries that host the agent image, packs, and images your cluster profile uses.
If your network restricts egress, export the proxy configuration in your terminal session before you install the agent.
Set the http_proxy and https_proxy variables, in both lowercase and uppercase forms. Refer to the proxy step in
Install Palette Agent for the exact commands.
Hardware Requirements
The following table compares the agent mode minimum requirements with the specifications of the NVIDIA Jetson AGX Thor Developer Kit. The minimum values match the agent mode prerequisites on Install Palette Agent.
| Component | Minimum (agent mode) | Jetson AGX Thor Developer Kit |
|---|---|---|
| CPU | 2 cores | 14-core Arm Neoverse-V3AE (64-bit) |
| Memory | 8 GB | 128 GB LPDDR5X |
| Storage | 100 GB, SSD required | 1 TB NVMe SSD |
| GPU | Integrated NVIDIA GPU | NVIDIA Blackwell, 2,560 CUDA cores, 96 Tensor cores |
Host Operating System
The Jetson device runs NVIDIA JetPack, which provides a Jetson Linux (L4T) operating system built on Ubuntu. For instructions on installing JetPack, refer to Prepare the Jetson Host.
For more information about NVIDIA JetPack and Jetson Linux, including troubleshooting, refer to the NVIDIA Jetson Software Documentation.
This guide is validated on the operating system versions in the following table.
| Component | Validated version |
|---|---|
| JetPack | 7.2.1 |
| Jetson Linux | L4T r39.2.1 |
| Ubuntu base | 24.04 LTS (Noble Numbat) |
Supported Kubernetes Distribution and CNI
Palette Optimized Canonical (edge-canonical) has no ARM64 build, so you cannot use it on a Jetson device. This guide
uses Palette Optimized K3s (edge-k3s) with the Flannel (cni-flannel) Container Network Interface (CNI), which is the
combination validated on the Jetson AGX Thor. You configure these layers in the cluster profile when you
register the host and deploy a cluster.
GPU Requirements
The NVIDIA GPU Operator pack requires a dedicated NVIDIA GPU and does not support embedded products such as NVIDIA Jetson. Do not use the GPU Operator pack on Jetson devices.
On a Jetson device, a workload reaches the integrated GPU through the NVIDIA container runtime that JetPack provides.
You do not add a GPU layer to the cluster profile, and you do not install a device plugin. Palette detects the GPU when
the host registers, but detection alone does not expose the GPU to your workloads. To give a pod access to the GPU, set
runtimeClassName: nvidia in the pod specification, and set the NVIDIA_VISIBLE_DEVICES and
NVIDIA_DRIVER_CAPABILITIES environment variables to all. Refer to
Enable GPU Access for Workloads for an example.
Palette Requirements
- Permissions in your Palette tenant to create cluster profiles and a Palette registration token.
Next Steps
After you confirm the requirements, continue to Prepare the Jetson Host.