On Fri, Mar 27, 2015 at 7:31 AM, Glenn Carrington [bs10g3c] <
bs10g3c.leeds.ac.uk> wrote:
> Hi,
>
> Thanks very much for the feedback and to Jason for the input files. Some
> useful pointers there.
> I think there has been a misuderstanding though, in that the x2 speedup on
> the GPU is Berendsen versus Langeivein.
No misunderstanding. I'm very surprised that Langevin causes a 2x
slowdown compared to NVE, let alone Berendsen (which should be basically
the same cost as NVE, but a tiny bit cheaper).
What Langevin does is generate 3N normally distributed random numbers each
step. Contrast that to the nonbonded interactions, of which there are N^2
of them. Langevin should definitely be more expensive than Berendsen or
Andersen or NVE. But for a system of 47K atoms, the nonbonded term should
be *so* much more expensive than the integration step (regardless of the
integrator you use) that you should barely notice the difference.
All the best,
Jason
--
Jason M. Swails
BioMaPS,
Rutgers University
Postdoctoral Researcher
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Mar 27 2015 - 05:00:05 PDT
