Why do you need to stick to just the "vendor supplied CUDA"? Is CUDA for arm64 targets only distributed via Nvidia ISOs for the Jetson targets? The download site seems to offer an arm64-sbsa option.

I thought the big barrier was the custom kernel they distribute. It was a pain to use the old Jetsons before Nvidia finally enabled modern c group support for containerization with docker. Perhaps building a kennel for their ended platforms is simpler now?