Use both Integrated and Discrete GPUs on NVIDIA Developer Kits

Holoscan v2.2.0

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).

As of this release, when these developer kits are flashed to leverage the dGPU, there are two limiting factors preventing the use of the iGPU:

  1. Conflict between the dGPU kernel mode driver and the iGPU display kernel driver (both named nvidia.ko). This conflict is not addressable at this time, meaning that the iGPU cannot be used for display while the dGPU is enabled.

  2. Conflicts between the user mode driver libraries (ex: and the compute stack (ex: for dGPU and iGPU.

We provide utilities to work around the second conflict:

Refer to the IGX user guide to learn how to leverage the iGPU in containers while the IGX developer kit is flashed in dGPU mode.

To leverage both GPUs in Holoscan, you can create separate applications running concurrently per the IGX documentation above, where the iGPU application must run in the Holoscan iGPU container, and the dGPU application can run bare metal or in the Holoscan dGPU container.

You can also create a single distributed application that leverages both the iGPU and dGPU by executing separate fragments on the iGPU and on the dGPU.

The example below shows the ping distributed application between the iGPU and dGPU using Holoscan containers:


COMMON_DOCKER_FLAGS="--rm -i --init --net=host --runtime=nvidia -e NVIDIA_DRIVER_CAPABILITIES=all --cap-add CAP_SYS_PTRACE --ipc=host --ulimit memlock=-1 --ulimit stack=67108864 " HOLOSCAN_VERSION=2.2.0 HOLOSCAN_IMG="$HOLOSCAN_VERSION" HOLOSCAN_DGPU_IMG="$HOLOSCAN_IMG-dgpu" HOLOSCAN_IGPU_IMG="$HOLOSCAN_IMG-igpu" # Pull images docker pull $HOLOSCAN_DGPU_IMG docker pull $HOLOSCAN_IGPU_IMG # Run ping distributed (python) in dGPU container # - Making this one the `driver`, but could be igpu too # - Using & to not block the terminal to run igpu afterwards. Could run igpu in separate terminal instead. docker run \ $COMMON_DOCKER_FLAGS \ $HOLOSCAN_DGPU_IMG \ bash -c "python3 ./examples/ping_distributed/python/ --gpu --worker --driver" & # Run ping distributed (c++) in iGPU container docker run \ $COMMON_DOCKER_FLAGS \ -e \ $HOLOSCAN_IMG-igpu \ bash -c "./examples/ping_distributed/cpp/ping_distributed --gpu --worker"

The L4T Compute Assist is a container on NGC which isolates the iGPU stack in order to enable iGPU compute on the developer kits configured for dGPU. Other applications can run concurrently on the dGPU, natively or in another container.


These utilities enable using the iGPU for capabilities other than display only, since they do not address the first conflict listed above.

Previous Enabling Exclusive Display Mode
Next Deployment Software Stack
© Copyright 2022-2024, NVIDIA. Last updated on Jul 17, 2024.