[CP2K-user] [CP2K:15601] Proper use of ri_hfx.F module

[hut... at chem.uzh.ch]

Hi

RI-HFX doesn't bring you a scaling advantage over the original
code (unlike RI-MP2 or RI-RPA). For very large systems the original
code has a better screening behavior and should be faster.
For small and medium system it all depends on the basis sets.
Large basis sets and diffuse primitives are better handled by
the RI-HFX code. This means ADMM (small basis!) is better for the
original code. In addition, I don't know if the automatic basis
set generated for RI is taking the ADMM basis or the original basis
as starting point (you can check in the output).
Positive points for RI-HFX are memory usage (number of ERI calculated and
stored, no replicated density and Fock matrices in parallel runs). 
Systems that only need a few SCF iterations, e.g. during MD are also
better for RI-HFX. GPU usage is also on the plus side for RI-HFX, although
the tensor dbcsr code is not yet performing optimal. 

best regards

Juerg
--------------------------------------------------------------
Juerg Hutter                         Phone : ++41 44 635 4491
Institut für Chemie C                FAX   : ++41 44 635 6838
Universität Zürich                   E-mail: hut... at chem.uzh.ch
Winterthurerstrasse 190
CH-8057 Zürich, Switzerland
---------------------------------------------------------------

-----cp... at googlegroups.com wrote: -----
To: "cp2k" <cp... at googlegroups.com>
From: "Nicholas Winner" 
Sent by: cp... at googlegroups.com
Date: 06/21/2021 10:34PM
Subject: [CP2K:15601] Proper use of ri_hfx.F module

Hi devs,

I've been looking at the resolution of identy hfx module. I know that the forces are not implemented, so its not fully developed, but it promises to speed up HFX calculations so I'm interested in checking it out. 

My issue is that the performance is not very good, and I'm unable to find documentation on it because its experimental. What I have done is taken a normal ADMM input file and added the &RI section to the &HF section, and tried the two different RI flavors, MO and RHO, with no other modifications. The calculation runs fine, but the first step is much longer than a standard HF calculation, and what's more the subsequent SCF steps, which are normally very quick due to integral storage, are even longer than a normal HFX calc.

I'm wondering if this is due to the basis? I don't see any documentation for how to add an  RI basis for HFX, so I simply have my primary basis (DZVP) and then my BASIS_SET AUX_FIT cpFIT3 for the ADMM part. Is there a way to add a special RI AUX basis that would improve the speed? 

Any pointers are appreciated.

-Nick  
  -- 
 You received this message because you are subscribed to the Google Groups "cp2k" group.
 To unsubscribe from this group and stop receiving emails from it, send an email to cp... at googlegroups.com.
 To view this discussion on the web visit https://groups.google.com/d/msgid/cp2k/80d500db-2983-44e5-8b3f-9da4f87924c6n%40googlegroups.com.