Changing ALGO should make no difference in an ideal world. However, when you invoke spin polarization, you may find a different magnetic state from both algorithms. The best practice would be to ensure that you converge to the right solution.
That being said, the NORMAL algo is normally more robust than Fast. This in general might be a bad sign for your system. You can also try the ALL algo and see what that gives.
I see you have also added an incar, here is some general advice that might influence convergence.
- ADDGRID is a spooky keyword, I would say never use it but sometimes it helps. I suggest leaving it off, convergence issues tend not to be fixed by it in general in these magnetic systems
- You are using a small sigma value, I would suggest using 0.2 and reduce it to your desired value after convergence. It will not influence geometry much but will make convergence much easier.
- Leave the mixing settings at their defaults most of the time. You can try this approach as a first attempt, but if it doesn't fix the problem do not keep it.
- Consider running a spin paired calculation first as a single point calculation, save the WAVECAR/CHGCAR, then add magnetization. This often helps as well.
- EDIFF = 1e-8 is insanely accurate, use something more like 1e-4 or 1e-5 for geometry optimization. If you find that you cannot converge the geometry you can raise it or switch to a VTST geometry optimizer which uses forces which are less sensitive to this.