On Tue, Sep 6, 2011 at 1:25 PM, Alexander Lebedev awlman@rambler.ru wrote:
Thank you very much for your response!
May I also ask how about running CUDA and non-CUDA analyses for the same dataset? I am asking because I don't have CUDA support on one of my machines as well (It has ATI videocard). I have found a thread on that topic (http://www.mail-archive.com/freesurfer@nmr.mgh.harvard.edu/msg16832.html) and even asked the author - and the question still hangs... What do you think?
If you require bit-for-bit agreement, then you basically can't change _anything_ about the computer configuration. It is the nature of floating point arithmetic that a+(b+c) != (a+b)+c There's nothing we can do about this - it's how computers work.
Now, the results still ought to agree in some statistical sense, but I never got that far with verifying it across a variety of brains and platforms.
Richard