On Sun, Oct 19, 2025, Tomás Benjamín Cáceres Madariaga via AMBER wrote:
>I'm experiencing a strange issue when running pmemd.cuda.
>The error is:
> Error: an illegal memory access was encountered launching kernel kNLSkinTest
>
>I thought this was related to some problem in the simulation, i.e.,
>something exploded or there was a very bad contact. However, when I ran
>the same simulation on another machine, it worked without a problem. Also,
>I repeated a simulation that worked fine before, and now it's also
>failing. So I'm starting to think it's related to the GPU or nvidia
>driver. I tried recompiling amber but the error still happen.
>
>I'm using Amber24, with a 3060 nvidia gpu, with cuda 12.4 and driver
>580. This happened in Ubuntu 22.04, then I updated to 24.04 and recompiled
>amber again, and nothing changed. I tried to use an old nvidia-driver (550,
>560) with no success, so I couldn't test if this is something related to
>nvidia driver.
This sounds most like a GPU hardware problem, but that is obviously just a
guess. Have you run "make test.cuda.serial" on the GPU that may be a
problem? That might help narrow things down.
A second thought is to use barostat=2 for your NPT runs, if you have not yet
done that. There have been some recent reports that sound similar to yours,
and for which setting barostat=2 helped. Again, this would help narrow down
the problem. (Barostat=2 probably should be the default setting anyway...)
...good luck...dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Oct 24 2025 - 21:00:02 PDT
