I found the best ccminer version is tried-and-true 2-year-old djm34's ccminer-sp-neoscrypt, on linux at least. I don't do Windows mining, but the releases for cuda6.5 and 7.5 are still there, however there's a possibility they might not be as good as KlausT's binaries on 8 or 9.1. Every time I've recompiled djm34 with newer cuda releases it's gotten better. Right now, djm34-sp on cuda 9.1 gives about 8% more hash/s than KlausT on 9.0 (will do 9.1 soon). Windows numbers may be similar if you can recompile.

I did a run-down of the four ccminer forks and nsgminer on neoscrypt for nvidia here on the Feathercoin UFO thread. I'm trying to update the list as I get the time.