> Don't make this mistake. One of my machines is AMD based (ASUS
> CROSSHAIR IV + Phenom II X6 1090T.4.0Ghz, Suse 11.4) and the results
> with two GTX580 3GB and maximum overclock are (NPT):
^^^^^^^^^^^^^^^^^
:'( -------------------------|
> JAC: 51.18 ns/day ( just 46.7 not overclocked), i.e. the scale is not
> good the CPU is..
We are working on addressing this for the AMBER 12 release but it will need CUDA 4.1 and peer to peer support. I need to check with NVIDIA to find out if it will ultimately be supported by the GeForce cards.
> The same tests on my second machine (Intel 2600K.4.6 and Asrok Z68
> Extreme7 Gen3) :
> JAC: 56.36 ns/day
I suspect this can be related to the front side bus speed and hence memory bandwidth differences between the two machines. What is the CPU memory bandwidth like for these two machines? It shouldn't have any real effect on single GPU performance but for parallel GPU performance the CPU memory bandwidth is critical. This is why 2 nodes by 1 GPU per node is faster than 1 node by 2 GPUs per node. Since the CPU memory bandwidth per GPU is better in the first case even when taking into account the IB interconnect.
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 |
| NVIDIA Fellow |
|
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 Wed Nov 30 2011 - 15:00:03 PST
