Inconsistent INISYM behavior between Intel- and NVIDIA-compiled VASP
Posted: Tue Feb 28, 2023 11:23 pm
I'm seeing inconsistent outputs between two VASP 6 executables compiled on different systems. I run VASP at NERSC, which has two different computing systems. On the first, VASP is built for CPUs with the Intel compiler and linked with Intel libraries; on the second, versions of VASP are built for both CPUs and GPUs with NVIDIA HPC SDK 22.7. In the attached files, I'm specifically comparing (Intel) vasp/6.3.2 and (NVIDIA) vasp/6.3.2-cpu, although I've tested several different versions and consistently encountered the same problem. Both sets of calculations use the vasp_ncl executable and start from identical input files.
If ISYM = 0, then both Intel and NVIDIA VASP behave identically and as expected. (These jobs do run quite slowly, and the attached Intel VASP calculation did not complete; but based on the in-progress outputs both versions seem to be converging to the same ferrimagnetic state.)
If ISYM = 2, then the Intel version also produces the expected output. However, the NVIDIA version detects different symmetries: INISYM returns the point group C_2, instead of the expected point group C_3h. We then get a different number of irreducible k-points and very different physical behavior (in particular, the magnetic and orbital moments are nearly wiped out).
I couldn't find any similar forum posts, so I'm not sure if this issue has been reported before. NERSC's VASP specialist is going to run some more tests, but at the moment we're not sure exactly what the root cause is, so we were hoping that the developers might have some insights. Thank you for your help!
If ISYM = 0, then both Intel and NVIDIA VASP behave identically and as expected. (These jobs do run quite slowly, and the attached Intel VASP calculation did not complete; but based on the in-progress outputs both versions seem to be converging to the same ferrimagnetic state.)
If ISYM = 2, then the Intel version also produces the expected output. However, the NVIDIA version detects different symmetries: INISYM returns the point group C_2, instead of the expected point group C_3h. We then get a different number of irreducible k-points and very different physical behavior (in particular, the magnetic and orbital moments are nearly wiped out).
I couldn't find any similar forum posts, so I'm not sure if this issue has been reported before. NERSC's VASP specialist is going to run some more tests, but at the moment we're not sure exactly what the root cause is, so we were hoping that the developers might have some insights. Thank you for your help!