IPS has been supported on GPUs for a long time...
What's never been explained is why it didn't scale well... Even NVIDIA
couldn't figure it out...
On Wed, Feb 12, 2014 at 6:19 PM, David A Case <case.biomaps.rutgers.edu>wrote:
> On Wed, Feb 12, 2014, Ross Walker wrote:
> >
> > Note, I agree the IPS implementation sucks, it should never have been
> done
> > as separate pairs_calc and I am looking to try and simplify this.
>
> There is a legitimate question of whether we need IPS in pmemd (in the
> released version -- we should certainly not throw away the code that we
> have). The original hope was that it might scale well to large processor
> counts on CPUs. I don't know if that has in fact ever been tested.
>
> Given the motivation for IPS, it really seems to be low priority to put
> this
> on GPUs. PME has become so much the "standard", one would have to sell
> the notion of using IPS based on parallel performance. Unless we have
> really
> good parallel performance with IPS, I don't see how it is going to win out
> over PME.
>
> Xiongwu may want to weigh in here -- I may be missing some key points here.
>
> ...dac
>
>
> _______________________________________________
> AMBER-Developers mailing list
> AMBER-Developers.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber-developers
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Wed Feb 12 2014 - 19:00:02 PST
