From: luc Renambot (renambot_at_gmail.com)
Date: Sun Jul 26 2009 - 18:52:30 CDT

On Jul 26, 2009, at 6:22 PM, Axel Kohlmeyer wrote:

> On Sun, 2009-07-26 at 16:20 -0500, luc Renambot wrote:
>> Hello,
>
> luc,
>
>
>>
>> I just got a new Linux system with 4 Nvidia GPUs (and eventually 6 of
>> them, 9800 GT 1GB RAM). I'd like to benchmark it.
>
> i am curious. why so many? why not 2 or 3 GTX 275 cards instead
> (and perhaps one 9800GT for the display)? transferring data across
> the PCIe bus is one of the big bottlenecks in GPU computing and
> - at least on paper - the GTX 275 should have about twice the
> performance than the GT 9800 while needing in total much less power.
>

Correct, but the machine was built to drive a lot of displays, not
especially for GPGPU.
However, I'm curious to see how it performs.

>> I got VMD installed and it recognize all the GPUs. Nice !
>> But I couldn't find on the web site a file or a benchmark to stress
>> test my new system.
>
> i don't have any example for VMD that you could use,
> but have you tried other GPU accelerated codes?
>

I have the CUDA SDK with the examples installed. I was interested to
check a real app.

>
>> Would anybody like to share such a file or script ?
>> I'd be happy to share results and configuration if anybody is
>> interested.
>
> i'd be interested to see how that machine works with
> HOOMD. http://www.ameslab.gov/hoomd/benchmarks.html
> n.b. the GPU parallelization in HOOMD is very simpleminded
> and thus not very efficient (i've seen 20-40% speedup from
> using a second GPU and next to nothing from a third),
> but you could just run 4 individual benchmarks backgrounded
> at the same time and tied to different GPUs via the
> --gpu=# flag and that should give you a good estimate for
> the best case scenario (i've been able to run with the
> 4 GPUs in a tesla S1070 on a single node machine with
> two 8xPCIe-2.0 busses reasonably efficient at a 15% performance
> hit, which is quite surprising given that it was on a
> dell poweredge 1950 with a horrible memory and PCI bandwidth).
>

Thanks, I'll check that.

> a second option would be to try the GPU accelerated
> compilation of NAMD.
>

I didn't see that NAMD was gpu-enabled also. Is that public ?

Luc

> cheers,
> axel.
>
>>
>>
>> Thanks,
>>
>> --
>> Luc.
>>
>> renambot_at_gmail.com
>>
>
> --
> =
> ======================================================================
> Axel Kohlmeyer akohlmey_at_cmm.chem.upenn.edu http://
> www.cmm.upenn.edu
> Center for Molecular Modeling -- University of Pennsylvania
> Department of Chemistry, 231 S.34th Street, Philadelphia, PA
> 19104-6323
> tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425
> =
> ======================================================================
> If you make something idiot-proof, the universe creates a better
> idiot.
>