---
title: GPU Access
description: Enable NVIDIA CUDA or AMD ROCm GPU access for compute and vision workloads
---

## GPU Access

The `gpu` entitlement exposes the host GPU to your app for CUDA, ML inference, image processing, DeepStream, PyTorch, TensorRT, and MLX workloads. WendyOS does not expose GPU devices to apps unless the app declares the entitlement.

This covers **NVIDIA GPUs (CUDA)** on Jetson and x86 systems, and **AMD GPUs (ROCm)** on supported x86 Linux hosts. AMD hosts need the `amdgpu` driver, `/dev/kfd`, and a DRM render node. The container also needs ROCm libraries compatible with its GPU.

For a complete AMD example, follow [Run PyTorch on an AMD GPU](../../guides/tutorials/python/amd-rocm-pytorch/). Stagefile's `cuda: true` option remains NVIDIA-only; the ROCm example selects its wheel index explicitly.

On **Qualcomm Dragonwing** boards (such as the IQ-8275), the entitlement exposes the Adreno GPU's DRM render node, which the mesa freedreno/turnip, OpenCL, and Vulkan userspace in your image opens. The on-SoC Hexagon NPU is a separate accelerator: `wendy device info` reports it on its own `NPU:` line as `qualcomm (qnn)` when its FastRPC nodes are present — never as an Adreno compute backend — and apps reach it through the `npu` entitlement rather than `gpu`.

On NVIDIA Jetson, the latest WendyOS is built on **JetPack 7.2** (L4T R39.2.0, CUDA 13). Target JetPack 7.2 or later when selecting CUDA wheels or base images for your app.

Add the entitlement from your project directory:

```bash
wendy project entitlements add gpu
```

## Inspect GPU Capabilities

Check the target device version and hardware summary:

```bash
wendy device version
```

For a focused hardware capability view:

```bash
wendy device hardware list --category gpu
```

The output reports GPU-related device paths and properties when the target exposes them.

## Common Pairings

GPU apps often also need:

- `camera` for computer vision and live inference
- `network` for inference APIs or dashboards
- `audio` for voice AI pipelines
- `persist` for model caches and local output

For camera inference:

```bash
wendy project entitlements add camera
wendy project entitlements add gpu
```

For a networked inference server:

```bash
wendy project entitlements add gpu
wendy project entitlements add network   # prompts for the mode; there is no --mode flag
```

## App Configuration

A minimal GPU app uses:

```json
{
  "appId": "com.example.gpu",
  "platform": "linux",
  "version": "1.0.0",
  "entitlements": [
    { "type": "gpu" }
  ]
}
```

Then deploy:

```bash
wendy run
```

<CliShot flow="run" step="deploy" alt="The Wendy CLI deploying a GPU-enabled app to a WendyOS device" />

## Raspberry Pi Board Telemetry

On Raspberry Pi hosts, the `gpu` entitlement also exposes the VideoCore mailbox device (`/dev/vcio`). This lets your app read board telemetry — power, voltage and current, temperature, throttling state, and (on Raspberry Pi 5) the PMIC ADC channels — through the firmware property interface, the same channel the `vcgencmd` tool uses.

This is added only on Raspberry Pi hardware, and only when the `/dev/vcio` node is present. On NVIDIA Jetson devices the entitlement grants the CUDA/CDI access described above instead — the VideoCore mailbox is not exposed there — and on hardware that is neither, it has no effect.

```json
{
  "appId": "com.example.power-monitor",
  "platform": "linux",
  "version": "1.0.0",
  "entitlements": [
    { "type": "gpu" }
  ]
}
```

<Callout type="info">
  Access is read/write to the existing device node only — the container cannot create new device nodes (`mknod`). The grant applies solely when the agent detects Raspberry Pi hardware.
</Callout>