Hi Filippo,
This seems strange. How long did the tests run for? They need to run for
at least a couple of minutes to get reasonable statistics for performance.
Can you check that ECC is turned off on the K20. You might also consider
rebooting the machine to make sure there isn't anything stuck running on
the K20. If you still see the same difference then please send me a copy
of your inputs and I'll try it here.
All the best
Ross
On 2/8/13 1:05 PM, "Filippo Pullara" <filippo.pullara.amber.gmail.com>
wrote:
>Dear all,
>
>We installed AMBER 12 with CUDA 5.0 and the newest patches and drivers in
>two GPU nodes. The first has GTX 680s, the second K20s.
>We do molecular dynamics and in order to run some test we used a system
>with 200.000 atoms.
>Here follow results:
>
>Node with GTX 680s, using only one GPU: 6 ns/day
>Node with K20s, using again only one GPU: 3 ns/day
>
>GTX 680 is two time faster than K20. We were not expecting that.
>Can please, someone give some suggestion?
>
>Thank you very much,
>Filippo Pullara
>
>
>_______________________________________________
>AMBER mailing list
>AMBER.ambermd.org
>http://lists.ambermd.org/mailman/listinfo/amber
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Feb 08 2013 - 13:30:05 PST
