I’ve been working with NVIDIA GPU VMs of late with the intention of running GPU enabled TensorFlow and PyTorch. I find working with NVIDIA software incredibly frustrating. In particular, installation instructions for CUDA / cuDNN simply do not work. The biggest culprit is missing or incorrectly versioned shared object libraries. Forums are loaded with developers having similar problems never with any viable solution. I’ve had better luck with NVIDIA distributed docker containers. There at least I can make it as far as running GPU enabled TensorFlow sample code.