Re: [AMBER] pmemd24 problems with RTX5090

From: David A Case via AMBER <amber.ambermd.org>
Date: Wed, 16 Jul 2025 08:32:57 -0600

On Tue, Jul 15, 2025, Oscar Conchillo-Solé via AMBER wrote:
>
>We've just acquired an Nvidia RTX5090 which we plan to use mainly with
>amber.
>
>#### Summary of the following long mail ######
>Amber24 crashes when running pmemd.cuda (pmemd.cuda_SPFP) with this
>error in stderr:
>of length = 42Failed an illegal memory access was encountered
>#### END of summary #####
>
>However, In the computer with the Nvidia RTX5090:
>
>I decided to keep going to see what happened and I managed to compile
>the code, this was my cmake command:
>cmake ../pmemd24_src
>-DCMAKE_INSTALL_PREFIX=/sharelab/labapps/AMBER/AMBER25/Amber25-GPU_MPI
>-DCOMPILER=GNU -DMPI=TRUE -DOPENMP=TRUE -DCUDA=TRUE
>-DCUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-12.8 -DINSTALL_TESTS=TRUE
>-DDOWNLOAD_MINICONDA=FALSE -DBUILD_PYTHON=FALSE -DBUILD_PERL=FALSE
>-DBUILD_GUI=FALSE -DPMEMD_ONLY=TRUE -DCHECK_UPDATES=FALSE

Just an aside: you don't need the final two lines in pmemd24_src: that code
automatically builds just pmemd.

>273 file comparisons passed
>15 file comparisons failed (8 of which can be ignored)
>4 tests experienced errors

This looks pretty good, although the 4 errors are unexpected.

>
>However, since the pmemd.cuda that the software has installed is a
>link to bin/pmemd.cuda_SPFP (like in the other previous versions) I
>also executed inside the test directory:
>./test_amber_cuda_serial.sh SPFP
>
>In this case it reported many errors:
>Finished CUDA test suite for Amber 24 at Fri Jul  4 12:11:47 CEST 2025.
>
>146 file comparisons passed
>44 file comparisons failed (9 of which can be ignored)
>101 tests experienced errors

We expect lots of "file comparisons failed" messages, since the mixed
precision GPU code does not match the CPU code to the full precision that is
printed. But the "errors' look quite serious.

We are working on one patch for Blackwell cards, and I am trying to get more
information about exactly what error messages are involved.

Thanks for the report....dac


_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Jul 16 2025 - 08:00:02 PDT
Custom Search