Is it possible to have MXNet 0.12 running on a g2 instance? The g2 instances have a GRID K520 gpu. If so, which drivers and CUDA Toolkit have worked for you?
I would expect it to work but I see that g2 instances are deprecated in favor of g3.
The quickest and easiest way to verify is to spin up an instance using the Amazon Deep Learning AMI. You could then see what drivers and libs are installed there if you were wanting to configure your own AMI.
+1 I agree with simonco@. I would suggest using MXNet 1.0.0 and G3 or P2/P3 instances.