[QE-users] GPU version gives different result from CPU version

[Ivan Carnimeo]

Dear Shimin,
it does not look like a problem with compilation, nor GPU related.

Assuming that all calculations are correctly converged, it seems that eigenvalues show some fluctuations by varying the size of the band groups (1 vs 16 cores per band group).

You could try to run a few calculations with -nb 1 and changing the total number of cores using plane-wave parallelism only,  to check if the eigenvalues still fluctuate also without using the band parallelism.

Another possible issue could be due to the presence of two stationary states close in energy, so that, by slightly changing the parallelization options, one or the other is picked as the solution.

To check this, you could take the calculation converged with one parallelization scheme and restart it using the density obtained from the other parallelization scheme, to test if the minimization leads to the same state or the other.

Best,
IC

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20240802/27e284fc/attachment.html>