I've recently been running a bunch of mathematical modelling applications on Azure, on NC6 and NC6v3 instances. Although the ease of use and scalability are great, the pricing seems excessive, especially when compared with the non-GPU VM offerings.
For CPU-based computation tasks, I find that Azure VMs come close to competing with using other types of hardware, especially when considering all the typical cloud benefits. But for GPU instances, I'm having trouble justifying running models in the cloud when compared to the relatively low cost of "filling a room with gaming rigs".
The performance of the VMs certainly isn't bad, but it's not spectacular compared with in-house hardware (modern rigs with 1080Ti GPUs). And given that I'd want to be running these machines 24/365, ordering a large box of graphics cards and just replacing them as they blink out seems an awful lot cheaper - although clearly not very green...
What am I missing here?