[gmx-users] 2020.1: comm broken with -update gpu -bonded gpu ?

[Magnus Lundborg]

Hi Jochen,

Have you tested if this happens with -update cpu? Perhaps it's the 
bonded on GPU that's the problem, unrelated to updating. I didn't see a 
Redmine issue yet. Could you make one?

Cheers,
Magnus

On 2020-03-19 18:01, Jochen Hub wrote:
> Hi developers,
>
> I am running a simple DPPC membrane (Berger force field, PME, 1nm 
> cutoff, 4fs time step, all standard) with 2020.1 and
>
> comm-mode                = Linear
> nstcomm                  = 100
> comm-grps                = DPPC water  ; OR System
>
> but the membrane rapidly drifts along the z direction: approx. once 
> across the box per 100ps, and accelerating over time.
>
> This happens only with
>
> mdrun -update gpu -bonded gpu
>
> but not with
>
> mdrun -update gpu -bonded cpu   (no spelling mistake)
>
> (with a GTX 1070Ti).
>
> Also no problems with 2018.6 or 2019.6.
>
> Seems like the center of mass motion removal is broken when doing both 
> *bonded and updating* on the GPU. Is this issue known?
>
> Cheers,
> Jochen
>