The price for a custom build with these specs is ~$8k: http://pcpartpicker.com/p/NP4MNG Spending that money on an EC2 GPU instance would be a better use of money unless you really need a local workstation.
I'd recommend a local custom build over EC2. EC2 GPU instances are virtualized via a Hypervisor, which dramatically reduces performance on multi-GPU networks. and this doesn't take into account the large amounts of disk space needed for the training set.
does anyone know which hypervisor do they use ? Can I build a local EC2 GPU instance with these GPUs ? I'm quite amazed that they are able to get the drivers, etc working with these GPUs on top of a hypervisor
Really? If you're serious enough to need something like this, you likely need it for more than 4-5 months which is the break even point for a single instance (for your quoted price anyway) and I'm pretty sure this'll have much higher performance.
