On Thu, Dec 24, 2020, Matthew Guberman-Pfeffer wrote:
>
>I noticed when minimizing my system with Sander or PMEMD on CPUs,
>the energies agreed very closely, and the structure relaxed
>smoothly. However, using the identical input files, but submitting to
>GPUs to use PMEMD.cuda, the structure did not relax, became horribly
>disfigured, and the minimization terminated early. I am new to the list
>(and Amber), and I do not know if attachments or figures are allowed in
>postings. But, I would be more than glad to share my files so others can
>reproduce and solve this issue.
You (probably) already have the solution: perform the initial minimization,
on a CPU. Generally, something like 1000 steps will remove the bad initial
forces that are likely to be causing your problem. You might also want to
check the initial structure: using the "check" action in cpptraj is a really
easy way to spot common problems.
The GPU code is written to be efficient and deterministic, but as a
consequence, it may fail if there are large forces. Many systems will
minimize fine on a GPU, because the forces are not too large. But many will
fail, and need some help from the CPU.
....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu Dec 24 2020 - 06:00:02 PST
