Dear Amber developers,
Our lab has recently bought your latest license. We tested it by using some
systems that are described in tutorials. For our calculations we use GPU.
It is much faster than the CPU.
We also prepared our own system and run minimization, NVT equilibration.
The system was stable and everything looked good. But then when we tried to
make an NPT calculation we had an error: "cudaMemcpy GpuBuffer: Download
failed an illegal memory access was encountered" - the calculation stopped.
We tried to restart it. Surprisingly, in some runs, everything was all
right, no errors during calculations. Then we run the production
calculation and the same error occurred. We also checked the size of our
system. It is smaller than the possible size (here you write some
limitations of possible calculations: http://ambermd.org/GPULogistics.php).
Also, we tried to change NTT and Barostat. We revealed that another NTT is
bad for calculations. Anyway, our calculation took less memory than is
available on our NVIDIA video cards (in the attachment you can find a
picture).
We also checked whether this error occurred because of the type of
parameters (we prepared systems with another type of water, systems without
ligands, systems without cations, etc). As a result, we didn't find any
correlation between the types of systems and the error emerging.
Our last assumption of the reason why some calculations have just
stopped was the number of NVIDIA video card. So we run our calculations on
different video cards. No correlation was found. Still, some runs are good
(without errors till the end) and some just stopped in the middle.
We run these calculations on CPU and everything was great. But we can't use
it as it is much slower.
So, we suppose there is a problem between Amber and GPU.
Could you help us, please? Looking forward to hearing from you.
Evgenia Elizarova
Moscow State University
The Faculty of Bioengineering and Bioinformatics
Laboratory of Structural Biology
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Nov 26 2020 - 10:30:02 PST