I am currently having jetpack 4.3 on the AGX board. There is one application using cuda 10.0, and that prevents me to upgrading to Jetpack 4.6 that has cuda 10.2.
I noticed if I use “apt dist-upgrade” method to upgrade from 4.3 to 4.6, the system leaves cuda 10.0 untouched. I can then install cuda 10.2, so that my system has both 10.0 (for old applications) and 10.2 (for new applications).
If I use clean flash.sh or OTA image upgrade, the system will only have cuda 10.2. If I pull the old cuda 10.0 deb packages from jetpack 4.3 and install them on jetpack 4.6, so that I have both cuda versions, do you see a problem?
Another option I am thinking is to run 4.3 docker container image inside 4.6, for the old application that depends on cuda 10.0. I checked the Dockerfile from nvidia / container-images / l4t-base · GitLab but they don’t seem to be maintained. If you try to build images using those Dockerfiles, they don’t match the images pulled from NVIDIA L4T CUDA | NVIDIA NGC
Where can I find the latest Dockerfiles?
BTW, the jetpack 4.6 docker image is correctly marked as for "Architecture": "arm64" (docker inspect nvcr.io/nvidia/l4t-base:r32.6.1 ). BUT jetpack 4.3 image is incorrectly marked as "Architecture": "amd64" (docker inspect nvcr.io/nvidia/l4t-base:r32.3.1)
I haven’t got a chance to test that. If there are dependencies with GPU driver on the board, does that mean we won’t be able to use a container with older l4t version that comes with old cuda version?