Hmm, autotune on windows is super inaccurate at the moment. The gettimeofday() routine I am using on Windows seems to have a resolution that is not accurate enough. And I think I skip the "take 3 measurements and average" policy when in scrypt-jane mode... But it is not a good idea for low n-Factors...
This needs to be refined before a new release.
Christian
It could be really slow sometimes. I was thinking about trying to add a few lines to cudaMiner --benchmark mode so that it exits after autotune finds the best startup config and make a massive batch file with all the different kernels/lookup gaps/N factors and just let it play while I'm at work. Not sure if the GPU would like to jump up and down all day though.