could you elaborate a bit more on your answer…what I found about it is this: "correction to SPOOLES.2.2 for large input decks" - #3 by feacluster
and the bottleneck I identified isn’t PARDISO but ARPACK. I case you were talking about the i4 and i8 issue, both versions I have have PARDISO as solver, one compiled as i4 and another as i8 (both from @feacluster), in my linux system i8 version runs slower for the same problem, in windows 11 it’s almost the same