Hi Amr,
Then AMBER GPU REMD code is not really designed for running on a single GPU. It was designed as a way to utilize large HPC machines that have many hundreds of GPUs. As such it is designed to one replica per GPU. While you can run multiple replicas on a single GPU it will run them all at the same time which will need lots of GPU memory - at least Nreplica times what a single run will take. It will also cause lots of thrashing issues on the single GPU such that performance is likely to be a lot worse than single replica perf / nreplicas.
So while you can do it neither of the options below is idea. Do you have access to a large cluster or HPC resource with many GPUs?
Note, in the case of being hardware limited (single gpu / single desktop) vs hardware unlimited (a big supercomputer) then I am not sure if you actually gain anything from using replica exchange. In raw terms it would be:
Does running a single calculation for X time give you worse sampling than running an REMD calculation for X / nreplica time? (seems unlikely to me). I am not sure how well this has been addressed or answered in the literature - maybe others can chime in here. Certainly it is true that running replica exchange for X time gives you better sampling than running a single simulation for X time - so if you have computer time you need to blow it is certainly a social acceptable way of doing that. Although one also should ask the question in that case of whether running REMD x X time is better than running Nreplica individual simulations each of X time.
In terms of GPUs - The Titan-X is faster and more capable than the GTX980 but around $400 or so more expensive. So it really comes down to your budget and what you would like to do most - run 2 independent simulations most of the time = go with the 2 x GTX980s. Run 1 simulation as fast as you can within the budget (and have the flexibility of being able to handle more atoms) go with the Titan-X.
All the best
Ross
> On Mar 23, 2015, at 8:19 AM, amr hamed <schrodinger8383.gmail.com> wrote:
>
> Dear all,
> I would like to know the performance of running a replica exchange
> simulation on a single gpu (namely the latest nvidia gpu, the titan-x which
> has 12 GB of ram).
> What do you recommend for this type of simulation: one titan-x or 2 gtx 980?
>
>
>
> Best Regards
> _______________________________________________
> 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 Mon Mar 23 2015 - 10:30:02 PDT
