[AMBER] cudaMemcpy GpuBuffer::Download failed - hardware or software?

From: Eugene Radchenko <genie.qsar.chem.msu.ru>
Date: Sat, 1 Aug 2015 23:30:14 +0300

Hi all,

I have some troubles with Amber14 pmemd.cuda on the ASUS Geforce GTX760 card.

I basically use the system (protein+membrane+water+ions) and protocol prepared by CHARMM-GUI for AMBER.
I guess this is not the ‘right’ thing to do but I was not yet able to get charmm2amber and tleap to process this system correctly.
Anyway, it works nicely using CHARMM parameters in the CPU/MPI mode.

In the GPU mode the performance is impressive and the minimization and equilibration phases also run ok.
However, in the production phase I get seemingly random CUDA errors (i.e. at random and not reproducible points during simulation):
    cudaMemcpy GpuBuffer::Download failed an illegal memory access was encountered
Right up to the error, the energy and volume/density seem pretty stable and similar to those for the CPU run. I tried running short strides with increased skinnb value (along the lines explained in Lipid14 tutorial) but it did not help.

So, the question is: how is it possible to check if this is some GPU card defect or some subtle bug in the AMBER code?

Thank you in advance
AMBER mailing list
Received on Sat Aug 01 2015 - 14:00:03 PDT
Custom Search