At this point it doesn't really matter since GPU mining is obsolete.
Also, what's wrong with NVidia's GPGPU capabilities? The last time I had to write GPU code, I found that CUDA was much more mature - and much more pleasant to write than the equivalent OpenCL. Also, the benefits of being able to run OpenCL code on multiple compute platforms were being realized by CUDA with the development of a PTX to x86 compiler.
That would have made the difference even larger, but even at 1:1 instruction timing AMD would have had a large price advantage so long as we're talking about the last 2 generations of cards or so. If the price/performance optimum lies further back than that, it may well have fallen at a point in time when AMD / NVidia were closer to price/performance parity.
To clarify, by "performance" I mean "performance for my needs" which means "double precision float performance."
> Also, what's wrong with NVidia's GPGPU capabilities?
The price per double-precision FLOP was off by a factor of four at the consumer level last time I went shopping and the number of DFLOPS available at the consumer level was capped lower. It seemed like a move to force penny-pinching CUDA-dependent academics to upgrade to teslas.
> The last time I had to write GPU code, I found that CUDA was much more mature - and much more pleasant to write than the equivalent OpenCL.
It still is, but the gap has closed for many use cases (including mine). I'm not saying that CUDA isn't/wasn't a reasonable choice, especially a few years back, just that we are now paying the hidden price that comes from allowing a monopoly to develop.
https://en.bitcoin.it/wiki/Why_a_GPU_mines_faster_than_a_CPU...
The bitcoin community was the only one nimble enough to switch to AMD. For those of us in academia, we'll be paying the piper for the foreseeable future :/