[gmx-users] cudaStreamSynchronize failed in cu_blockwait_nb

[Alex]

Additional info:

The base system is a SuperMicro X8DTG-D motherboard with dual Xeon E5620
CPUs at 2.4 GHz, and 24 GB of DDR3 ECC RAM

The GPUs are two NVidia Tesla C2075s

Running under Ubuntu 14.04 LTS 64-bit, kernel 3.13.0-53, with CUDA 7.0,
installed via the NVIDIA-provided .deb file,
cuda-repo-ubuntu1404-7-0-local_7.0-28_amd64.deb, followed by the usual
apt-get update/upgrade.

Nvcc reports version 7.0.27.

GMX version is 5.0.5, locally built, with auto-detected options, including
"GMX_GPU ON", and "GMX_SIMD SSE4.1"

On Mon, Jun 8, 2015 at 2:07 PM, Alex <nedomacho at gmail.com> wrote:

> Hi all,
>
> I am running a GPU-accelerated test (GMX 5.0.5) and, after about 60K steps
> of what seems to be a huge speed-up, I get the error from the subject with
> the following comment: "unspecified launch failure."
>
> This is on an 8-core (16-thread) Intel CPU and two Tesla cards. The run
> command is:
>
> mdrun -ntomp 4 <file directives>
>
> I saw some previous messages on this error, but not sure there's a
> solution.
> Any thoughts? I will provide any additional info as needed. If there are
> additional options necessary to debug etc, could you provide the exact
> command line?
>
> Thanks,
>
> Alex
>