Dear Ekaterina,

That's right, different machines and different operating systems will produce slightly different results, as shown in this study:

Glatard, Tristan and Lewis, Lindsay B and da Silva, Rafael Ferreira and Adalat, Reza and Beck, Natacha and Lepage, Claude and..., "Reproducibility of neuroimaging analyses across operating systems", Frontiers in neuroinformatics (2015).

This is not specific to one toolbox or even neuroscience, this impacts all domains of computing. Indeed, the underlying low-level libraries doing the calculations, and the processors, have different instructions sets and specifications, and thus introduce slight biases that can have accumulating effects at the macroscopic scale (ie, the end result of an analysis). It's like trying to translate a sentence in multiple languages: the translation might be very close and convey the global meaning, but it will never reflect totally perfectly the original meaning nor style.

A way to reduce this issue is to make a virtual machine and use it to run all your calculations, but even this will not ensure perfect reproducibility.