Energy difference on different Platforms

[slaubach]

I run VASP on a Athlon Linux cluster compiled with the makefile:

makefile.linux_ifc_ath

and on the Jülich IBM cluster JUMP running on POWER6

For a calculation on LiCoO2 on those two machines I get two different TOTEN for the same input namely:

-335.795219 eV on JUMP
-335.916387 eV on the Athlon cluster

I have no idea, what the problem might be. I tried different optimizations (O3 and O2) without any effect on the JUMP Cluster. But I had to use O2 on the athlon cluster to get the same energy result for parallel and serial calculations.

Thanks in advance

[admin]

1) please check if the number of bands were the same for both runs.

2) To test whether the error is due to optimization, re-run with lowest optimization level on both machines.

3) Otherwise, there may e.g. be an error in one of the libraries (LAPACK, BLAS) on one of the machines.

[Brane]

I have the same problem on IBM Power6, i.e., the calculated total energy is wrong when I use -O3 or -O2. As expected, if the optimiztion level is changed to the lowest one (-O0), the energy is correct. I think some subroutines need the lowest optimiztion level, but it is diffucult to identify them. Does anybody have a clue?

Thanks.