Skip to main content

astra-1680-deepx

View source on GitHub

This guide takes you from zero to a running DeepX M1 NPU inference demo on a Grinn Astra SL1680 SBC.

Prerequisites​

  • Hardware: Grinn Astra SL1680 SBC (grinn-astra-1680-sbc) with a DeepX M1 NPU on the PCIe bus, a display connected to HDMI, and a USB UVC camera for the Qt object-detection demo.
  • Media: an SD card (or the board's supported boot media) for first-time provisioning.
  • Host: the Avocado CLI installed, plus Docker (the SDK runs in a container).
  • Network: the board and host on the same network if you want to sideload iterative updates with avocado deploy.

Initialize​

Clone the references repo and enter the reference:

git clone https://github.com/avocado-linux/references.git
cd references/astra-1680-deepx

default_target is already set to grinn-astra-1680-sbc, so no --target flag is needed.

Install​

Resolve and fetch all extensions, packages, and SDK dependencies declared in avocado.yaml:

avocado install

Build​

# Base bring-up image (Weston desktop only):
avocado build -r dev

# Full DeepX stack (driver + runtime + models + Qt demo):
avocado build -r dev-deepx

For this reference the build step cross-compiles, inside the SDK against grinn-astra-1680-sbc:

  • deepx-driver — the DeepX PCIe kernel modules (dx_dma, dxrt_driver), built against kernel-devsrc.
  • deepx-rt — the DXRT C++ inference runtime and dxrt.service.
  • deepx-models — sample ONNX/DX models, fetched and staged for the NPU.
  • deepx-qt-example — the Qt5/QML YOLO detection app (built against Qt + OpenCV).

It then composes the extensions and builds the rootfs/initramfs images (with the dev permissions profile — empty root password, not for production).

Deploy​

First time (provision the board):

avocado provision -r dev-deepx

Follow the prompts to write the image to your boot media, then boot the board. Provisioning is a one-time step per device.

Iterative updates (already provisioned + reachable on the network):

avocado deploy -r dev-deepx -d <board-ip>

This sideloads only the changed extensions over SSH/HTTP — no reflash.

Verify​

After the board boots into the dev-deepx runtime:

# NPU kernel modules loaded:
lsmod | grep -E 'dx_dma|dxrt_driver'

# DXRT runtime service is active:
systemctl status dxrt.service

On the connected HDMI display, Weston starts in a kiosk session and the Qt YOLO demo (qt-deepx-example.service) draws bounding boxes over the live camera feed. Check the service if the window doesn't appear:

systemctl status qt-deepx-example.service
journalctl -u qt-deepx-example.service -f

Customize​

  • Models: replace or add models in the deepx-models extension (files/dx-models-*.sh) to run your own network on the NPU.
  • Application: the deepx-qt-example extension (and files/dx-qt-example-*.sh) is the day-to-day app to edit; swap in your own detection logic or UI. A headless inference suite is also available via the deepx-app extension.
  • Runtime selection: use -r dev for a minimal display-only image during bring-up, or -r dev-deepx for the full NPU stack.
  • Display: adjust the Weston kiosk behavior in files/weston-kiosk.conf / files/weston-kiosk.ini.
  • Permissions: the dev permissions profile sets an empty root password for convenience. Define a real profile under permissions: and point rootfs/initramfs at it before any production use.