An update to my AWS tests. This time Using GPU - G2.2xlarge instance.
Stop the clock! Setting up the nvidia driver to use the GPU was not pleasant and I nearly gave up. I'm not going into it here but it was very messy
My original cpu only test is here
This test was done using the same scene file in blender 2.76a - a 2000x2000 resolution basic "Aiko with hair and clothes" render..
AWS G2.2xlarge Nvidia Grid (4Gb) 1,536 cuda, Xeon E5-2670 CPU, 15Gb Ram - 69 seconds
MyPc GTX760 (4gb) Core2Quad, 8GB Ram - 2 minutes
So the G2.2xlarge instance was quite a bit slower at rendering the scene than the CPU C4.8xlarge instance (25 SECONDS TO RENDER) used previously and faster, as expected, than my home pc.
As I now have windows server set up with the Nvidia Driver, my upcoming G2.8xlarge test should be better (I can now also make an AMI copy of the instance and use Spot Price on new Instances)