Re: [AMBER] GPU memory error during ABMD run

From: Owen O'Reilly via AMBER <>
Date: Mon, 6 Feb 2023 10:52:25 -0500

The input files are identical and the job was run on the same hardware, so
I'm leaning towards it either being an issue with the libraries having been
recompiled or possibly a result of my home directory being relocated.
Probably should have mentioned that in the original posting.

In addition, a "normal" pmemd.cuda run without the ABMD parameters runs
with no issues.


On Mon, Feb 6, 2023 at 8:33 AM David A Case <> wrote:

> On Sun, Feb 05, 2023, Owen O'Reilly via AMBER wrote:
> >
> >I've recently been encountering an error when I try to run a single-walker
> >flooding-type adaptively biased MD run using pmemd.cuda. The run will
> >proceed as expected until around 1/10 of the expected duration, at which
> >point it will quit and write out a single error message:
> >
> >cudaMemcpy GpuBuffer::Download failed an illegal memory access was
> >encountered
> >
> >This has occurred in both AMBER18 and AMBER20, and has occurred with
> >identical simulations that ran properly a few months ago.
> Can you say what changed in the past few months? Are "identical
> simulations" truly identical? Did you recompile the codes? Run them on
> different hardware? (etc.)
> Do the cuda tests pass now? It's not impossible that you are seeing a
> hardware problem. Of course, maybe someone on the list will recognize
> the symptom (the error message itself is of almost no use.)
> ....dac
AMBER mailing list
Received on Mon Feb 06 2023 - 08:00:02 PST
Custom Search