http://journal.frontiersin.org/article/10.3389/fninf.2015.00012/abstract
Results for standard neuroimaging analysis pipelines differ somewhat
when run on two very similar Redhat-based Intel Linux clusters. The
conclusion was that these differences were due in part to differences in
single precision (32-bit) floating point.
The authors say: "A first step to correct these reproducibility issues
would be to use more precise representations of floating-point numbers
... "
Another conclusion might be that the results are inherently numerical
unstable if they really rely on minutiae of floating point
implementation.
The differences would surely have been greater comparing really
different operating systems like Windows and OSX.
The Google+ URL for this post was
https://plus.google.com/+MatthewBrett/posts/49XrQyjpH8Q