NGC | Catalog
CatalogContainersL4T Compute Assist

L4T Compute Assist

Logo for L4T Compute Assist
This container allows running compute-only applications using CUDA and TensorRT on the integrated GPU of Holoscan devkits (Clara AGX DevKit, IGX Orin DevKit) while other applications can run on dGPU, natively or in another container.
Latest Tag
April 1, 2024
Compressed Size
2.24 GB
Multinode Support
Multi-Arch Support
l4t_35.3-trt_8.5.2 (Latest) Security Scan Results

Linux / arm64

Sorry, your browser does not support inline SVG.

Disclaimer: this container is deprecated as of the Holoscan SDK 1.0 release. Refer to your developer kit user guide for up-to-date instructions.


NVIDIA developer kits like the NVIDIA IGX Orin or the NVIDIA Clara AGX have both a discrete GPU (dGPU - optional on IGX Orin) and an integrated GPU (iGPU - Tegra SoC). At this time, when these developer kits are flashed to leverage the discrete GPU (using HoloPack), the integrated GPU cannot be used due to conflicts between the CUDA libraries for dGPU and iGPU stack.

This container allows running compute-only applications using CUDA and TensorRT on the iGPU of these developer kits by isolating the iGPU stack from the rest of the host device. Other applications can run concurrently on the dGPU, natively or in another container.

For Holoscan documentation, visit the Holoscan developer page.


  1. Set up your developer kit in dGPU mode:

    Developer Kit User Guide HoloPack
    NVIDIA IGX Orin Guide 2.0
    NVIDIA Clara AGX Guide 1.2
  2. Load the nvgpu kernel driver for iGPU:

    sudo insmod $(find /usr/lib/modules -name nvgpu.ko -type f,l | head -n1)
  3. Install the Holoscan SDK

  4. Either:

Running the container

  1. Log in to the NGC docker registry

  2. Copy the version you want to use from the Get Container drop-down at the top of this webpage, and set this as your NGC_CONTAINER_IMAGE_PATH in your terminal:

    • for HoloPack 2.0, use l4t_35.3-trt_8.5.2
    • for HoloPack 1.2, use l4t_34.1.2-trt_8.4.0
    # For example
  3. Start the L4T Compute Assist container:

    export HOLOSCAN_SDK_INSTALL_PATH="/opt/nvidia/holoscan" # choose other path if installed somewhere else on your host
    export APP_PATH="/path/to/your_app"
    export DATA_PATH="/path/to/your_data" # if needed by your_app
    docker run -it --rm --net=host \
      --privileged \
      --runtime=runc \
      --ipc=host \
      --cap-add=CAP_SYS_PTRACE \
      --ulimit memlock=-1 \
      -v ${APP_PATH}:${APP_PATH}:ro \
      -v ${DATA_PATH}:${DATA_PATH}:ro \
    • --privileged is to run with privileged permissions to access the iGPU driver.
    • --runtime=runc is to ensure you are not using the nvidia docker runtime which would load dGPU drivers instead of iGPU. If this does not work, remove the nvidia runtime from your defaults in /etc/docker/daemon.json (you'll need to add --runtime=nvidia when running your other containers for dGPU).
    • HOLOSCAN_SDK_INSTALL_PATH is the path to where you have installed the Holoscan SDK on your host.
      • Debian package: /opt/nvidia/holoscan
      • Python wheel: in your environment's dist-packages/holoscan
      • If you want to mount the SDK from a Holoscan dGPU container instead of your devkit host:
        • add --name holoscan_dgpu -v /opt/nvidia/holoscan when running your dGPU container
        • replace -v ${HOLOSCAN_SDK_INSTALL_PATH} ... by --volumes-from holoscan_dgpu when running the L4T Compute Assist container
    • -e PYTHONPATH is set for python3 to find the Holoscan python module if it is not installed from a wheel.
    • APP_PATH is the path of your built Holoscan application on your host.
    • DATA_PATH is the path of any data on your host which you might need to run your application.
    • --ipc=host, --cap-add=CAP_SYS_PTRACE, --ulimit memlock=-1 are needed for the distributed applications using UCX to work. (See
  4. Once in the container, you can run ${APP_PATH}, or any other commands to leverage CUDA and/or TensorRT on the iGPU.


  • You can run deviceQuery inside the container to confirm you are using the iGPU device (Xavier/Orin) and CUDA drivers (11.4).
  • You can run tegrastats on the host to visualize the iGPU load when running an app in the container (GR3D_FREQ percentage).


deviceQuery: no CUDA-capable device is detected

Ensure that the nvgpu kernel driver is loaded (see prerequisites).

tegrastats: does not show GR3D_FREQ

Ensure that the nvgpu kernel driver is loaded (see prerequisites).

deviceQuery: returns "Quadro RTX" capable device (dGPU), 11.6 CUDA drivers

Ensure that you're not using the nvidia container runtime (refer to --runtime=runc instructions above).

tegrastats: GR3D_FREQ is at 0%

Ensure that:

  • you're not using the nvidia container runtime (refer to --runtime=runc instructions above)
  • your app is running in the iGPU container
  • your app uses CUDA for GPU computation


By pulling and using the container, you accept the terms and conditions of this End User License Agreement.