Hi Sasha,
> Can you help with the issue of local bandwidth? Nvidia sells dual host
> PCI-E adapter card that effectively connects 4 GPUs in an S2050 to a
> single x16 slot. When pmemd.cuda is run locally (parallel or four
> serial
> processes), how much would this impact performance?
> In other words, should we even consider these cards for a host system
> intended to run pmemd.cuda?
Once again this comes down to an issue of the number of degrees of freedom.
Many years ago it was simple. You just looked at clock rate. Then came
multiprocessors and interconnects that added more degrees of freedom, now it
gets even more complicated. There is also the fact that we simply don't have
access to a big enough variation in hardware, or in a lot of cases the
patience to deal with sales people to get access to such hardware for
testing...
Given these caveats I will give you my best 'guess' at the performance
difference. I would think for running multiple serial jobs it would not make
much difference. These do not use the PCI-E connection very often. Only when
output needs to be written so if you have ntpr and ntwx set large (1000+)
then I suspect the difference will be marginal. For parallel runs though the
PCI-E connection is used on every step to communicate between cards so here
I think you will see a big difference. How big a difference? I have no idea,
but I would guess that it would be fairly large... It will really be a case
of suck it and see I think. If you get access to such a machine I'd love to
have the numbers to put up on the website to help everyone out.
All the best
Ross
/\
\/
|\oss Walker
---------------------------------------------------------
| Assistant Research Professor |
| San Diego Supercomputer Center |
| Adjunct Assistant Professor |
| Dept. of Chemistry and Biochemistry |
| University of California San Diego |
|
http://www.rosswalker.co.uk |
http://www.wmd-lab.org/ |
| Tel: +1 858 822 0854 | EMail:- ross.rosswalker.co.uk |
---------------------------------------------------------
Note: Electronic Mail is not secure, has no guarantee of delivery, may not
be read every day, and should not be used for urgent or sensitive issues.
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Aug 05 2010 - 07:30:05 PDT
