[gmx-users] GPU HW questions

Szilárd Páll pall.szilard at gmail.com
Mon Aug 22 16:31:27 CEST 2016

On Wed, Aug 17, 2016 at 3:49 PM,  <melichercik at leaf.nh.cas.cz> wrote:
> Hello,
> I have 2 processor Xeon E5-2660 node and I want to add the GPU. I'm thinking
> about Radeon Fury (without X) which I can get bellow 350 eur, or 2 pcs. of RX
> 470. My systems have usually between 100 000 and 200 000 atoms - I suppose it
> can influence the CPU-GPU ratio. Or do you think something different would be
> better? (some Geforce? as they are more expensive and I have several Radeons
> working with GMX to my satisfaction).

Fury Nano is what I would recommend, those cards are pretty good,
close to the GTX 980 Ti in performance (and to the 1070). The NVIDIA
cards will likely be a bit faster with small simulations systems
(<40-50k), but at 100k the Fury should reach its peak too.

> More general question is what is most influencing factos for GPU calculations?
> Is it number of cores/shaders? Or access to the memory or something different?
> I'm thinking about the situation that you have e.g. some GPU and you want to
> upgrade it a little bit (e.g. use old one with slower CPU and buy a little
> more powerful). Well, if it is possible to state something in general. :)
> Thanks in advance.

With the current code what matters most is number of ALUs and
frequency as the offloaded non-bonded kernels are compute-bound.
However, that does not mean you can compare #ALUs*Freq across
different vendors (not even across different generations when the
architectures change).


> Milan Melichercik
> --
> Gromacs Users mailing list
> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/GMX-Users_List before posting!
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-users or send a mail to gmx-users-request at gromacs.org.

More information about the gromacs.org_gmx-users mailing list