Nvidia has recently made it a lot easier to use Docker images in order to run Deep Learning and other GPU-based tasks. I'm mostly a hobbiest (though I do use GPUs professionally sometimes) and so I can't really justify buying a Dream Machine when you can get spot instances of (decent) GPUs on AWS for $0.10-0.20/hr.

However, as with all things, setting up AWS instances is a PITA. This nice guy Sean wrote a lovely tutorial on how to get everything going, but it is (a) out-of-date now that Ubuntu 16.04 Xenial is out, and its drivers are a bit behind. Moreover, it's a good explainer, but I just want a list of commands to execute on the box! Hence, this repository!

Launch an AWS box with some GPUs in it. I've been using g2.2xlarges. Make sure to request Spot Instances if you're cool with that.

Once you've gotten all that setup, just download nvidia.sh and run

sudo ./nvidia.sh

I haven't been able to turn this into an AMI. The machine boots and either the nvidia-docker plugin is screwed up, OR the nvidia kernel module doesn't load. If anyone has suggestions, let me know how to turn this into a functioning AMI.

This is all just a scripted and updated version of this awesome tutorial.

CC-BY-SA 4.0