This is good! Thank you!
Thats '250 atoms per nonbond cell. I have bitonic merge sorts in register
space out to 512 atoms.
InefficientAF in register space, but screamtastically efficient otherwise...
B40C sort remains ~1.5x better than CUB in the 10K to 100K atom range,
doesn't break even until 500K. This sucks. But giving up on moving minds
and deciding to build instead.
On Fri, Jun 25, 2021 at 1:33 PM David Cerutti <dscerutti.gmail.com> wrote:
> 12.0A, 2A skin I would imagine. But the systems that use this cutoff will
> probably be membrane systems, so they will not be smaller than perhaps 80A
> on a side.
>
> Dave
>
>
> On Fri, Jun 25, 2021 at 3:25 PM Scott Le Grand <varelse2005.gmail.com>
> wrote:
>
> > Hey all, I am thinking about a radical rewrite of the sort code used in
> > CUDA AMBER.
> >
> > I need one datapoint. What's the biggest nonbond cutoff plus skin in
> > practical use currently?
> > _______________________________________________
> > 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
>
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Fri Jun 25 2021 - 14:00:03 PDT
