Very slow RDGRAD

[Michael_Patzschke]

Hello,

I am trying to run an x2c structure optimisation using TM version 7.8.1 The system has 103 atoms C,N,H and one lanthanide. I am using the x2c-TZVPPall basis set and the calculation is surprisingly slow. The gradient calculation on the same system using def2-TZVPP basis set (no x2c, sc-ECP on the lanthanide) takes 2 min on 32 cores. With x2c the calculation takes 32 min! The RIDFT parts on the other hand have comparable timings of about 1-2 min per SCF cycle. Is there any way to speed up the gradient calculation?

Kind regards,
Michael

PS Happy new year to all Tubomole developers and users!

[Michael_Patzschke]

Upon closer inspection it becomes evident that the most time consuming part of the regard calculation is indeed the x2c gradient. Here is were the calculation lingers for 25 min:

Code: [Select]

--------------------------------------
 Performing 1c X2C gradient calculation
 --------------------------------------

 maximum number of shells which are related by symmetry :  1

 
 Estimated memory requirements for X2C quantities in MB on 1 CPU in C1:
                  3208
 
 Evaluating integrals

 1e-integrals will be neglected if expon. factor < 0.127604E-12

 Constructing 1c X2C-Dirac matrix of irrep                      1 in c1     
 Evaluating integral derivatives
 Constructing 1c X2C-transformed 1e gradient contribution

In the latest version of orca the x2c gradient part is not very time consuming at all.

Help is appreciated!

[yannickf]

Hello,

for speed-ups simply use the local version, i.e. add $rlocal to your control. Then, the calculation time is substantially reduced. Please see https://doi.org/10.1063/1.5022153.

[uwe]

Hi Michael,

comparing two different quantum chemistry software packages can sometimes be a challenge. In the old days (I am not saying good old days, just old days ) with plain non-relativistic Hartree-Fock without any approximations things were much easier to compare.

What needs to be done is to check what the programs are actually doing, the defaults of ORCA and Turbomole are in many cases not identical. Or even close. DLU-X2C (as yannickf recommended to switch on by adding $rlocal) is the default in ORCA but not in Turbomole, at least not at the moment (versions 7.x). That's what the manuals say and I assume those are correct for both packages.
In addition, the choice of the operator (relativistic vs. non-relativistic) needs to be checked, search for the keyword picture-change in both manuals and make sure to have those set similarly.

As the ORCA manual says that the implementation follows the work that was done for Turbomole by Franzke, Weigend and others, the results should in principle be comparable - if, and only if, the same settings are used. However, the number of keywords at least in Turbomole for X2C is not small - they are given in the documentation, section 'All-electron relativistic approaches (X2C, BSS, DKH)'.

Best Regards