Hardware Setup

The Xisom Edge AI Box runs on an NVIDIA Jetson. Prepare the device, check how inference uses its GPU, and wire it into your plant network before you install the software.

Supported hardware

Item	Requirement
Device	NVIDIA Jetson TX2
Operating system	JetPack 4.5 (L4T r32.5)
Inference	ONNX Runtime on the integrated GPU (CUDA), with CPU fallback
Memory	4 GB or more. The installer warns below about 3.5 GB.
Free disk	About 6 GB where you copy the release bundle

GPU prerequisites

The JetPack image must have these in place before you install Xisom. The installer checks for them and stops if one is missing. It does not install them.

• The NVIDIA container runtime, registered with Docker. docker info must list an nvidia runtime.
• A genuine Jetson (L4T) system. The installer looks for /etc/nv_tegra_release.

Note

The Jetson has no nvidia-smi. GPU load and temperature on the dashboard come from the Jetson’s own system interfaces instead.

How inference uses the GPU

The inference service picks an execution provider when it starts. On the TX2 it uses CUDA, and it falls back to the CPU if CUDA cannot start. The dashboard shows which one is active.

CUDA (GPU)

The normal mode. Models run on the TX2’s integrated GPU through CUDA. The dashboard shows CUDA.

• Requires the NVIDIA container runtime on the host.
• Export models with ONNX opset 15 or lower. See Preparing Models with modelctl.
• On first start, the inference service initialises CUDA. Its health check allows a 90-second start period, so it shows starting for up to about a minute and a half. That is expected.

CPU (fallback)

The safety net. If CUDA cannot start, the service falls back to the CPU. Inference keeps running, with higher latency. The dashboard shows CPU (fallback) — that badge is the reliable signal. A warning is logged only when the container offers no CUDA provider at all.

• To make the box refuse to start instead of falling back, see Running on CPU when GPU expected.

TensorRT is not used on the TX2. The ONNX Runtime build for JetPack 4.5 needs a newer TensorRT than the platform ships, so the service skips it and runs on CUDA. There is no engine compile at first start.

Tip

If the dashboard shows CPU (fallback) on a box that should use its GPU, see Monitoring and the GPU not used — CUDA errors runbook.

Network requirements

• A LAN connection that operators’ browsers can reach. The dashboard listens on port 80 (and 3000 as an alternate), or on 443 when you turn on HTTPS.
• A route to your plant devices: OPC UA servers, MQTT brokers, and PLCs.
• No outbound internet access. You install and update from the offline bundle.

The inference service’s internal port (50051) is bound to the box’s loopback address only. It is not reachable from the network.

First boot

1. Connect power and the network cable.
2. Browse to the dashboard URL the installer printed: http://<box-ip> by default, or https://<box-ip> if you turned on HTTPS.
3. Sign in as admin with the password the installer printed.
4. Continue to Connect an input datasource.

If something goes wrong

• Box running on CPU (fallback), or inference stuck unhealthy after the start period — see the Troubleshooting runbook.

Next steps

Install from the offline bundle Air-gapped, USB-shippable install runbook.

Jetson Deployment Jetson-specific behaviour and checks.

Connect a datasource Stream sensor and process data in.