Yeah, Nvidia's consumer cards are greta value in theory for CUDA but they aren't designed for full load operation for hours on end.
conversely, their Tesla cards are downclocked quite a bit to make them be able to work in those condition but they cost 5x more so teh value proposition drops pretty quickly.
Yeah, no doubt they have a lot of potential. I was an early adopter of GPU rendering when Vray released their OpenCL kernel a couple of years ago. I bought a pair of GTX590's to take full advantage. I quickly found that there was really little that I could do using it so I sold the cards and forgot about it. I then tried it again when they released their CUDA kernel last year with an expanded feature set. It was more usable but when I turned on all of the stuff that makes my renderings look good it's performance dropped to maybe 1.5x what my CPUs could do and there was still stuff missing... I then quickly hit the 6GB memory limit and that was that for GPU rendering.
I also tried iRay after a collegue reocmmended it. That was nice and he does get good results for car renderings but not much else. The GPU renderers have to be written from scratch so its taking it a while to build them up to the level that CPU renderers have evolved over the past 25 years. The other issue is that CPU renderes are constantly improving as well so they're a moving target for GPUs. My personal rule is that you should buy into what you use and what makes you money now and for the next 1-2 years. Tops.