You're absolutely right, HOWEVER, I would call that the exception rather than the rule, as IIRC it was largely due to AMD supporting some operators in a more native/optimized form than Nvidia did at the time, than any sort of true general purpose superiority. Given how quickly mining moved to ASICS I wouldn't use that as an advocation that they're (ASICS) superior to Nvidia for GPGPU functionality. (Special-purpose specially-designed, no doubt, but that's not really "General")

Edit: read a sister post later article on the breakdown of the perf differences, and it seems to slightly affirm what I'm saying (re: bitcoin aligning with a functionality AMD performed very strongly in), although I'll admit to have only skimmed it.

I will say though, even despite this, it doesn't address the known heat/power issues even the "heyday" amd cards seemed to suffer from and that were still present when I was comparing them while looking at a 1070 recently.

EDIT: since I can't edit my original post any more, I'd note here that I didn't intend to turn this into a NVIDIA/AMD debate, even if we ignore that for a moment and consider it a duopoly situation instead of monoply, I'd ask readers to consider my core point in that light.

Try to write some GPU software on AMD cards, then compare that to the NVidia/CUDA experience. I don't know about the Bitcoin thing (maybe in the end, for very long lived code, the lower cost would be more important than the developer pain and time time required) but for pretty much all the HPC code I see, getting things running sooner and with more polished dev tools beats price/performance ratios every time. If you're spending a few thousand on Titans anyway, getting a few more cycles per watt doesn't matter all that much anymore; except for Bitcoin mining where your main cost is energy (by design).

So I think there's a very clear reason for that specific case.

I'm... confused why you'd say this. It's simply true: NVIDIA's consistently ahead on these tasks on performance/energy cost metric.

You can do most any computation on any GPU with the right SDK support (and of course, the CPU hasn't gone anywhere). It's just fantastically less efficient.