Hi Jinfeng,
I've seen similar issues with the latest cuda although didn't get a chance to properly repro it.
I would suggest going back to CUDA 11.1 and see how it performs with that. I'd also suggest downloading the benchmark suite (here <
http://ambermd.org/Amber18_Benchmark_Suite_RCW.tar.bz2>) and see how that performs.
It's also worth checking the card temperature and power consumption with 'nvidia-smi' while the benchmark is running to make sure things aren't thermally throttling.
All the best
Ross
> On Jul 16, 2021, at 00:40, jinfeng liu <jinfengliu1119.gmail.com> wrote:
>
> Dear Amber developers,
>
> I run MD simulations on our new RTX-3070 gpu cards with cuda-11.3. However,
> I find that it is much slower than that on our previous RTX-2080ti cards.
> For the same system, I can run 2.5 ns simulation per day on RTX-2080ti, but
> only ~1.2 ns per day on RTX-3070. In addition, I got a warning in the out
> file for the simulation on RTX-3070 which is like the following,
>
> "WARNING: Stack usage limited by a hard resource limit of 12582912 bytes!
> | If segment violations occur, get your sysadmin to increase the
> limit."
>
> Could anyone help to figure this out? Thanks very much !
>
> Best,
> Jinfeng
> _______________________________________________
> 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 Fri Jul 16 2021 - 06:00:02 PDT
