DIGITS on Jetson TX2 vs Virtual Machine

I’m currently without an NVIDIA host machine. Thinking about setting up a Ubuntu VM as a host for now, but also considering just using the Jetson TX2 for both training and inference. I know DIGITS is not officially supported on TX2, but it seems some people have had success making it work?

Primary question - does anyone know on what setup DIGITS would perform better - virtual Linux machine with a virtual GPU (all running off an external HDD), or Jetson TX2?

Secondary question - if you have managed to install DIGITS on the TX2, would you mind disclosing your formula?

Unless you have a physical GPU with pass-thru support in the VM (i.e. ESXi hypervisor or similar), it’s unclear if the virtual emulated GPU would support CUDA and accelerated training. Note that you can always run Caffe without CUDA enabled, but it would be slow. I’d try AWS or onboard TX2. DIGITS may not be formally supported on Jetson but it seems others who’ve tried haven’t hit blockers. Mostly you want to remember to use NVcaffe-0.15 branch (instead of the experimental/FP16 branch of NVcaffe which is typically used on Jetson that supports FP16 inference but not training).