Re: [AMBER] Fwd: GPU pmemd.cuda runs very slow (3ns/day)???

From: Jason Swails <jason.swails.gmail.com>
Date: Thu, 4 Jun 2015 08:15:12 -0400

On Thu, Jun 4, 2015 at 8:04 AM, Gerald Monard <
Gerald.Monard.univ-lorraine.fr> wrote:

> Hi,
>
> I've checked my numbers, with a GTX780, on a system of 261k atoms
> (similar size as yours), I have 9ns/day. Given that a GTX760 has twice
> less number of cores than a GTX780, your numbers are not fast, but they
> are not slow either.
>

​Not only does the GTX 760 have half the CUDA cores as the GTX 780, but the
cards that I've seen have lower memory bandwidth as well.​ So 3 ns/day is
not surprising.

One thing you can try to speed up your simulation is to use a 8 A cutoff
instead of 10. As long as you're using PME (which you are), the cutoff
will have no effect on the electrostatic interactions, and the van der
Waals interactions at those distances are negligible (and corrected by the
long-range correction quite well for typical systems).

HTH,
Jason

-- 
Jason M. Swails
BioMaPS,
Rutgers University
Postdoctoral Researcher
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Jun 04 2015 - 05:30:04 PDT
Custom Search