[gmx-developers] PME on GPUs time line

[Berk Hess]

Hi,

A basic version of PME on CUDA GPUs will be shipped with the 2017 
release. Most likely this will only be single GPU, possibly with the 
option to run pair interactions on one GPU and PME on another. The 
changes up in gerrit are all working, but not complete yet. Aleksei has 
a development tree with a complete implementation. The main question is 
how much performance optimization and feature completion we can do 
before the 2017 release.
PME on GPU will allow you to buy a CPU+GPU cluster with cheaper CPUs if 
you are targeting throughput. It is less clear what the best setup will 
be for the highest ns/day, since PME requires a lot of different kernels 
which results in higher overheads.

Cheers,

Berk

On 01/19/2017 08:55 AM, Jochen Hub wrote:
> Hi developers,
>
> I noticed that Aleksei has uploaded a whole bunch of commits for 
> implenting PME for GPUs. Since we about to buy a new cluster, I was 
> wondering if there is a rough time line when PME/CUDA code goes into 
> master and realease branches. If PME runs smoothly on GPUs, the 
> hardware with the best price/performance ratio obviously changes.
>
> Also, are the commits in the branch "master (pme)" already suitable 
> for some preliminary benchmarking?
>
> Many thanks,
> Jochen
>