Switching Between iGPU and dGPU

The Clara AGX Developer Kit can use either the Xavier AGX module GPU (iGPU, integrated GPU) or the RTX6000 add-in card GPU (dGPU, discrete GPU). Only one GPU can be used at a time, with the iGPU being the default after flashing the Clara Holoscan SDK with JetPack. Switching between the iGPU and dGPU is performed using the nvgpuswitch.py script contained in /usr/local/bin.

To view the currently installed drivers and their version, use the query command:

            

            
$ nvgpuswitch.py query
iGPU (nvidia-l4t-cuda, 32.5.0-20201012161040)
        

To install the dGPU drivers, use the install command with the dGPU parameter (note that sudo must be used to install drivers):

            

            
$ sudo nvgpuswitch.py install dGPU
        

The install command will begin by printing out the list of commands that will be executed as part of the driver install, then will continue to execute those commands. This aids with debugging if any of the commands fail to execute for any reason. The following debug arguments may also be provided with the install command:

            

            
-d  Does a dry run, showing the commands that would be executed
    by the install but does not execute them.

-v  Enable verbose output (used with `-d’ to describe each of the
    commands that would be run).

-i  Run commands interactively (asks before running each command).

-l [LOG]
   Writes a log of the install to the file `LOG’.
        

The dGPU driver install may be verified once again using the query command:

            

            
$ nvgpuswitch.py query
dGPU (cuda-drivers, 455.32.00-1)
        

After the dGPU drivers have been installed, rebooting the system will complete the switch to the dGPU. At this point the Ubuntu desktop will be output via DisplayPort on the dGPU, and so the display cable must be switched from the onboard HDMI to DisplayPort on the dGPU.

            

            
export PATH=/usr/local/cuda-11.1/bin:$PATH
        

If at any time you want to switch back to iGPU, use the install command with the iGPU parameter:

            

            
$ sudo nvgpuswitch.py install iGPU
        

After the iGPU drivers have been installed, rebooting the system will complete the switch to the iGPU. At this point the Ubuntu desktop will be output via the onboard HDMI, and so the display cable must be switched from the DisplayPort on the dGPU to the onboard HDMI.

When switching between GPUs, CUDA is first uninstalled and then reinstalled by the script in order to provide the correct versions used by iGPU or dGPU (CUDA 10.2 and 11.1, respectively). Since many of the Clara Holoscan SDK installed packages depend on CUDA, this means that these Clara Holoscan SDK packages are also uninstalled when the active GPU is switched.

To reinstall the Clara Holoscan SDK packages after switching GPUs, the corresponding *.deb packages that were downloaded by SDK Manager during the initial installation can be copied to the Clara AGX Developer Kit and installed using apt. By default, SDK Manager downloads the *.deb packages to the following location:

            

            
~/Downloads/nvidia/sdkm_downloads
        

The current list of packages that must be reinstalled is as follows. Note that the version numbers may differ – if this is the case, use the latest version of the arm64 package that exists in the download directory.

• VisionWorks

  libvisionworks and libvisionworks-dev

• DeepStream 6.0

  deepstream-6.0_6.0.0-1_arm64.deb

• Clara Holoscan DeepStream Samples

  clara-holoscan-deepstream-sample_1.4.0-0_arm64.deb

• DeepStream GPU Video Test Source Plugin

  clara-holoscan-gstnvvideotestsrc_1.1.0-0_arm64.deb

Use apt to reinstall these packages:

            

            
$ sudo apt install -y libvisionworks libvisionworks-dev
$ sudo apt install -y ./deepstream-6.0_6.0.0-1_arm64.deb
$ sudo apt install -y ./clara-holoscan-deepstream-sample_1.4.0-0_arm64.deb
$ sudo apt install -y ./clara-holoscan-gstnvvideotestsrc_1.1.0-0_arm64.deb