[gmx-developers] GROMACS performance information

[Johanna K. S. Tiemann]

Dear all

We are planning a new HPC cluster (budget ~$300,000) for GROMACS (typical
simulation size: 100,000 atoms) and I have a few questions regarding its
performance characteristics which are really crucial for that.

1. Can 4 GPUs be used efficiently simultaneously in one compute node? I
have seen some benchmarks and they usually show bad scaling to more than 2
GPUs. However, my plan is to use more CPU power for each node than in all
the benchmarks I found. Is that going to help or is the bottleneck
somewhere else than CPU power?

2. Is NVLINK going to have a positive performance impact?

3. We consider using AMD EPYC CPUs instead of Intel Xeon for
price/performance reasons. However, I saw some recent benchmarks which
showed substantial performance gains when using AVX-512 (which only Intel
supports at the moment). However, these benchmarks were done without GPU
acceleration. I would guess that most of the calculations that benefit from
AVX-512 are actually offloaded to the GPU anyway. So my question boils down
to: is there any disadvantage when using AMD EPYC CPUs that I should be
aware of?

Any help would be highly appreciated. Best regards and thanks in advance,

Alexander Vogel & Johanna Tiemann

-- 
Johanna Tiemann, MSc
AG ProteInformatics / Hildebrand lab
Institute of Medical Physics and Biophysics
University Leipzig
Härtelstr. 16-18, 04107 Leipzig

Phone: +49 341 - 97 157 60
mobile: +49 176 - 832 798 51
skype: minga-madl
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20180125/cc2a939e/attachment.html>