Q: We have a large economics model running on both Windows and a Linux machine. The main development is on Windows. Recently we copied the model to Linux to run. The model gave us differences in the 5 significant digits. (XXXX.5374 was now XXXX.4157, for example). The model code is exactly the same (except for some obvious file path changes). Is this possible?
Such results are certainly possible. As long as both solutions satisfy the feasibility and optimality tolerances, they can both be correct. The reason for the difference could quite possibly be due to the fact that Windows and Linux have their own math libraries which can provide very small differences in an individual computation. One small difference can put a numerical algorithm on a different path causing the final answer when the convergence tolerances are met to be different by a larger number.
You might experiment by tightening the tolerances for the solver on each machine to see if you can get them to agree to another one or two digits.