That's fundamentally incorrect.
There are currently more available CUDA or double OCL + CUDA implementations than there are OCL engines.
VRay RT in example offers both.
And while nVIDIA isn't putting much stock in OCL and optimizing accordingly (unsurprisingly), it can and will run both unless there are hardware targeted optimizations.
AMD will lock you with OCL only for now.
That precludes you a lot of products such as Octane (CUDA only), RedShift (CUDA only) and some pretty decent CUDA extensions for other softwares.
At this point in time CUDA is quite simply more adopted, more mature, better documented, and much better served. Giving it up for some mythical apps that benefit from OCL is not reasonable. Not unless you are putting together a folding rig, or a bitcoin miner and so on. In DCC OCL is largely not of the relevance we’d all hope for yet.
Double precision floating point performance, measured in GFLOPS, is typically the amount of floating point operations (numbers after the decimal point) a GPU can handle in 1 second.
Incorrect at best.
FLOPs are floating point operations per second, NOT double precision ops. Big difference. DP FlOps aren't an atomic operation, you don't meausre by them.
Double precision involves a lot more to be taken into consideraiton, not last that many videocards are artificially crippled in their DP for market phasing (IE: GTX 6 and 7 cards, but not the 5s, quadros or Titans).
GPUs are based on parralelism and so, they naturally have more GFLOPS than CPUs (CPUs are optimized for interger calculations, but come with a subsidiary Floating point unit (FPU) or Vector unit (VU) or both, as seen in the Playstation 2’s primary processor, the great Emotion Engine)
I'm not entirely sure where you are going with this.
It's a 386 to Pentium set of notions. Modern CPUs are not that simple.
There is a lot more differentiating the two than that, and CPU architecture these days is ridiculously complex. A number of other factors will also come into play (IE: whether the compiler was set to, or even able, to take advantage of some features).
Based on all these, i will advice Dual AMD Radeon HD 7970s in CrossfireX. AMD is superior to Nvidia in OpenCL, the API by which GPGPU is based (the means by which you can render with your gpu) It also seems superior to the GTX 780 in DX11 acceleration, which translates into your nitrous viewport frame rate. I hope you understand all i’ve said,
You keep mentioning OCL as if it’s the premiere, or even the only, GPU rendering platform. That is a million miles off the truth.
nVIDIA has such an overwhelming dominance in the DCC market that nobody in their right mind would make an OCL only commercial product. It’s easier to find something CUDA only than it is to find something OCL only between products of any relevance.
Crossfire will do absolutely nothing for your viewport, and is generally regarded as a waste of money outside of gaming.
On top of that the 79xx has considerable syncronicity issues when it really gets taxed (IE: offline rendering on a GPU).
The k4k is daylight robbery. Unless you need quadro specific features (12bit colour buffers, proper stereo support in Nuke etc.), in which case the k5k or the k6k are where it's at, a Titan or a 7xx, or a 580 if on a budget, tend to be better bang for buck.
I have a Titan and don't regret it, but I do a fair chunk of CUDA work. Unless you need the uncrippled DP (which most softwares don't require) or the abundant ram the 770 and 780 are better value.
On a budget, especially if you need DP on a budget (unlikely) the 5xx remain strong cards.
