Inconsistency between the linux and widows OS in covarites regression

Hi there,
I did same analysis in different OS with same data from fmriprep (imported from fmriprep). The latest conn are used in the both OS. But I found after I abstracted the fisher-z FC in these two OS are slight different.
So I also checked the time series after regression by the latest conn, and I found the value of each voxel is also slightly different even though the general image looks similar.
Lastly, I selected one subject, and did a correlation analysis between fisher z FC from linux and windows. I found that the correlation between these two OS is around 0.71. While I also did a correlation between the FC regressed GSR and NonGSR, I found within the same OS, the correlation between GSR and NonGSR are around 0.988, but across OS are still around 0.71.
I am wondering that is this inconsistency acceptable? Or what causes this inconsistency. I should also indicate that I used a different version Matlab. (linux: Matlab R2016a; windows: Matlab R2017b)

Thanks
Hongru