[QE-users] QE+LIBXC on AMD EPYC CPUs poor performances

[Pietro Davide Delugas]

Dear Giuseppe, this is very strange.  There is nothing tightly CPU-architecture-specific in those code parts; it's plain Fortran. Architecture-specific optimisation is expected from the Fortran compiler; in my experience, gfortran is equally not the best for both architectures (and whether good or bad, it would be equally good or bad for all the code, not just for one single kernel).  Intel and AOCC compilers usually do much better architecture-specific optimisation.


The first thing to check is that you are actually using only the physical cores, so (number of MPI ranks) × (number of OpenMP threads per process) should be at most 128, assuming nothing else heavy is running on the machine. Use the env variable OMP_NUM_THREADS to set the number of OpenMP threads per process.

That said, on general grounds, 7.6 will be available on the Quantum Espresso website by this weekend, and pw behaviour with SCAN and R2SCAN will be significantly improved.  Better give it a try next week.
Pietro
________________________________
From: users <users-bounces at lists.quantum-espresso.org> on behalf of Giuseppe Zollo via users <users at lists.quantum-espresso.org>
Sent: Thursday, July 23, 2026 16:31
To: users at lists.quantum-espresso.org <users at lists.quantum-espresso.org>
Subject: [QE-users] QE+LIBXC on AMD EPYC CPUs poor performances

Dear all,
I'm experiencing very very poor behavior of QE 7.5 + LIBXC 7.0.0 on AMD EPYC 7713 CPU regardless of the compilation. I've compiled using gcc+ native AMD libraries aocl or gcc+MKL or intel+MKL. In any case, after successful compilation, I receive very poor behavior using r2SCAN meta-GGA. The system is not large  (less than 130 atoms) and the same calculation behaves reasonably well on two Intel(R) Xeon(R) Gold 6226R (16 cores each) (much older and much lower benchmark), nearly 300 secs per scf iteration step. Using two AMD EPYC 7713 cpus (64 cores each, much faster and with NUMA access) the performance is even worse than two Xeon(R) Gold 6226R.
I've spent days trying to fix it and it seems it could be related to the AMD EPYC architecture.
I wonder if anyone has experienced the same and if there are indications on how to handle this drawback.
Thanks
Giuseppe

--
Giuseppe Zollo
Dipartimento di Scienze di base e Applicate per l'Ingegneria
University "La Sapienza" of Rome
via A. Scarpa 14-16, Rome-Italy
phone: +39 06 49766947


Fai crescere le giovani ricercatrici e i giovani ricercatori

con il 5 per mille alla Sapienza

Scrivi il codice fiscale dell'Università 80209930587
Cinque per mille<https://www.uniroma1.it/it/node/23149>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.quantum-espresso.org/pipermail/users/attachments/20260724/dc74522c/attachment.html>