Ceilidh ... Re^2: The same input file produces different single point results on different computers

Firefly and PC GAMESS-related discussion club

Re^2: The same input file produces different single point results on different computers

Dear Dr Granovsky,
Thank you very much for your detailed answer.
Can I ask another questions?
1. Can the conlusion be done that all DFT calculations of lanthanide compounds, that can be found in literature, are not correct in principle ?
2. On the other side, I am still worried with message
SCHWARZ INEQUALITY OVERHEAD: 2775 INTEGRALS, T= 0.31
What does T value mean ? This value was found to be significantly different (for example, 0.08 and 0.29) for the same task on different machines.

On Fri Jan 22 '10 10:42pm, Alex Granovsky wrote
-----------------------------------------------
>Well, similar problems actually have been already discussed several times on this forum...

>The atomic state (term) you are trying to model is seven-fold
>degenerate and hence, strictly speaking, the only applicable methods
>are those that perform averaging over all seven degenerate states,
>e.g. SA-CASSCF. However, as you are using single-state single-reference
>method (UDFT), it is natural to expect the existence of multiple
>solutions of SCF equations. Both your outputs are correct and wrong
>at the same time - they are correct in the sense that there are no
>any errors introduced by the program and the final answer is indeed
>the converged SCF solution..., while they are wrong in the sense that
>these not-exactly-equal energies and different SCF solutions are just
>artifacts of the use of the improper theoretical model.

>As to the reasons of this instability... the CPUs are slightly
>different so there does exist the difference in round-off errors
>etc... The HCORE guess results in properly degenerated seven
>f-orbitals, however, as they are degenerate, the exact one which
>will become populated in the final solution is indeed defined by
>these round-off errors... Just look at the converged orbitals
>and you'll find them different in these two cases.

>Regards,
>Alex Granovsky
>
>
>On Fri Jan 22 '10 8:48pm, Dmitriy wrote
>---------------------------------------
>>The situation is absolutely different: the difference between energies is significant - in third decimal place. Moreover, the conditions are identical - tasks are performed in sequential mode (but in parallel mode the same discrepancy is observed), but the discrepancy in energies arises already at the first step of SCF.

>>On Fri Jan 22 '10 8:01pm, Roman Zubatyuk wrote
>>----------------------------------------------
>>>Seems to be already discussed.
>>>http://classic.chem.msu.su/cgi-bin/ceilidh.exe/gran/gamess/forum/?C39bb915069e1-7317-802+00.htm http://classic.chem.msu.su/cgi-bin/ceilidh.exe/gran/gamess/forum/?C39bb915069e1-7317-802+00.htm

Sat Jan 23 '10 1:34am

This message read 872 times