Thank you for the helpful information.
What I find is that using the same initial structural information generated
using tleap, the minimization runs okay on the CPU, whereas the same
minimization fails on the GPU even when using double-precision executable.
Is there any special care that needs to be taken while generating the AMBER
topology, parameter, initial models etc. if I want to run the minimization
exclusively on GPU (using double-precision) and then switch to
single-precision production run on the GPU?
On Sat, Nov 12, 2022 at 10:08 AM David A Case <david.case.rutgers.edu>
wrote:
> On Fri, Nov 11, 2022, Prithviraj Nandigrami via AMBER wrote:
>
> >Thank you for the helpful information. Looking into the mdout files, it
> >looks like the first step (minimization) does not run properly on the GPU
> >and the error propagates downstream - the energy at successive steps of
> the
> >minimization run remains the same. In contrast, on a CPU, the energy
> >decreases during the course of the minimization.
>
> The GPU uses fixed-point accumulation, which can overflow if there are bad
> initial forces. It is not all that uncommon for users to have to carry out
> a short minimization on the CPU prior to using the GPU.
>
> ....dac
>
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Nov 18 2022 - 12:00:03 PST
