[gmx-users] Gromacs GPU system question

Szilárd Páll szilard.pall at cbr.su.se
Thu Jul 4 19:14:11 CEST 2013

On Mon, Jun 24, 2013 at 4:43 PM, Szilárd Páll <szilard.pall at cbr.su.se> wrote:
> On Sat, Jun 22, 2013 at 5:55 PM, Mirco Wahab
> <mirco.wahab at chemie.tu-freiberg.de> wrote:
>> On 22.06.2013 17:31, Mare Libero wrote:
>>> I am assembling a GPU workstation to run MD simulations, and I was
>>> wondering if anyone has any recommendation regarding the GPU/CPU
>>> combination.
>>>  From what I can see, the GTX690 could be the best bang for my buck in
>>> terms of number of cores, memory, clock rate. But being a dual GPU card, I
>>> was wondering if  there is any latency issue that could make its
>>> performances less favorable with respect to a GTX Titan.
>>> Also, which motherboard, CPU is recommendable for this system.
>> The most important aspect to consider (by far) is, in my humble
>> opinion, *your specific workload*:
>>  - Size of the simulation box / number of atoms,
>>  - Specific force field/required integrator (verlet?),
>>  - Handling of long range electrostatics (pme/RF/coulomb).
>> Furthermore, the effect of the CPU is, imho, much more
>> pronounced. Remember, mdrun-gpu doesn't 'run' on the
>> GPU (as, e.g., HOOMD does) but loads work-sets up to
>> the GPU, runs them, and loads them back. For example: in
>> one box, I  have an AMD FX-8350 and a GTX-660Ti available
>> for tests, and I didn't see the GPU load going much over 60%,
>> even with millons of atoms. Here, small differences in the
> Without DD you should be able to get up to 75-80% CPU-GPU overlap
> (slighly less with vsites).

Note that you can increase GPU utilization by running two or more
*independent* runs (as part of a multi-run or not) on the same GPU
(for details see: http://goo.gl/dv1iK).

So in practice, if you have to simulate e.g. two confirmations of a
protein, running the two on half of the CPU cores, both using the same
GPU(s) can give substantial aggregate performance.


>> potential/force field used by the model will probably change
>> the performance of the GPU-related parts significantly (due
>> to cu-offs and buffering schemes).
>> Regards
>> M.
>> --
>> gmx-users mailing list    gmx-users at gromacs.org
>> http://lists.gromacs.org/mailman/listinfo/gmx-users
>> * Please search the archive at
>> http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
>> * Please don't post (un)subscribe requests to the list. Use the www
>> interface or send it to gmx-users-request at gromacs.org.
>> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists

More information about the gromacs.org_gmx-users mailing list