Hi Ying,
To add to this. That configuration is TERRIBLE for raw AMBER performance.
Now if you want to run lots of CPU runs or have lots of students
interactively using etc then you need that many CPU cores but if you plan
to run GPU jobs using AMBER you are wasting a lot of money on the CPUs. We
recommend going with the cheapest CPUs you can. That's the great thing
about AMBER's GPU design, as Scott says it is independent of the CPU so if
you only plan to run GPU jobs then you can buy very cheap (<$150 cpus) and
just fill the box with GPUs. Attached is an example spec for building your
own machine through Amazon. You can put 4 GPUs in this and run 4
simulations all at once at full speed. Have 4 GTX-Titan's in there and you
can run something like DHFR NVE at 4 lots of 115ns/day for an aggregate
throughput of almost half a microsecond of MD per day!
NAMD and Gromacs take a very different approach and rely on the CPU for a
lot of the work. Thus for comparable performance with those codes you have
to spend a fortuned ($1500+ per CPU!) on rocket fuel powered top end CPUs.
AMBER couldn't care less and we specifically designed it that way since
the underlying driving force was always to bring supercomputer or better
performance to an individual grad students desktop and if one had to spend
$10K+ per desktop to get such performance it would be a non-starter.
So my advice if you haven't bought the machine yet would be to consider
building your own 8 core AMD box with 4 GPUs in as shown. Alternatively
you can talk to Exxact about the MD Simclusters and MD SimWorkstations
that we have designed specifically with AMBER performance in mind. A 2 U x
4 GTX680 config is shown as an example and more details are on
http://ambermd.org/gpus/recommended_hardware.htm#exxact_machines
Hope this helps.
All the best
Ross
On 5/8/13 12:51 PM, "Scott Le Grand" <varelse2005.gmail.com> wrote:
>With AMBER, performance is entirely determined by the GPU (well 98+% to be
>absolutely utterly 100% honest)
>With GROMACS, performance is a complex brew of the GPU(s) and the CPU(s)
>(if you spend enough to get it right, you're up to 30% faster than AMBER,
>if you get it wrong you can be <<1/2 the speed)
>
>I'd suggest plugging more GPUs into this box. For independent runs, it'll
>scale AMBER throughput linearly...
>
>
>
>On Wed, May 8, 2013 at 12:19 PM, Sergio R Aragon <aragons.sfsu.edu> wrote:
>
>> Hello Ying,
>>
>> We have a machine of this type where the GPU is a GTX-580. To
>>demonstrate
>> the much larger efficiency of using a GPU compared to CPU
>>multiprocessing
>> with amber, I asked my students to compare both methods for the protein
>> 1AO6. Using the GTX-580 you could get about 12 ns/day for an explicit
>> water simulation with TIP3P waters in an 8 A octahedral box using
>>ff99SB,
>> and the SPDP model in Amber 12. Using 12 CPU cores the performance was
>>a
>> factor of 100 slower. Thus, even if you use all 64 cores, the CPU
>> performance is unuseable.
>>
>> I built this machine for teaching computational chemistry methods - you
>> need many CPU processors to give many students a chance to go
>>simultaneous
>> computations with g09, and other packages that don't parallelize well.
>>But
>> for doing MD, it is much better to have even a cheap GTX-560 card in
>>many
>> PC's to have many students work at the same time.
>>
>> Hope that helps somewhat.
>>
>> Cheers,
>>
>> Sergio
>> SFSU
>>
>> -----Original Message-----
>> From: sun [mailto:sun.ntnu.edu.tw]
>> Sent: Wednesday, May 08, 2013 9:40 AM
>> To: amber.ambermd.org
>> Subject: [AMBER] pmemd performance on AMD 16cores x 4 (total 64 CPU
>>cores)
>> + 1 GPU
>>
>> Hi,
>>
>> does anyone have experience on the machine of this configuration? and
>> benchmarks?
>>
>> Thanks very much.
>>
>> Ying-chieh
>>
>> --
>> Open WebMail Project (http://openwebmail.org)
>>
>>
>> _______________________________________________
>> 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
>>
>_______________________________________________
>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 Wed May 08 2013 - 13:30:03 PDT
