Unfortunately with vsites GPU-resident mode (where integration is on the GPU) is not supported, so try removing -udpate gpu (or pass -update cpu).
With -update cpu, the results are in.
gmx mdrun -ntmpi 1 -ntomp 24 -bonded gpu -pme gpu
351.4 ns/day
gmx mdrun -ntmpi 2 -npme 1 -ntomp 23 -ntomp_pme 1 -bonded gpu -pme gpu
429.4 ns/day – I did set the env vars as instructed within the submission script.
These numbers are to be compared with the 2 fs timestep data I posted before (nearly 600 ns/day), so my 4-GPU runs are still winning. :)
Interesting, thanks for testing. One last check you could do is using three of the four GPUs.
Are you able to share you input, I’d like to have such a case as a reference.
We will soon have the release candidate of the 2022 release out which contains multiple multi-GPU improvements. It would be great if you could test it and give feedback on the new release.
I think I tested on 3 GPUs earlier – don’t remember the exact performance, but it was scaling as well. Happy to share the input privately, if you could provide an email address you prefer, because it looks like there’s no way to send a private message to you here.
You can send private messages here, e.g. by clicking on the icon next to the post and selecting “Message”:
