On Thu, May 14, 2020, Niransha Kumarachchi wrote:
>
>I am trying run verify energies calculated by sander, sander.MPI, pmemd,
>pmemd.MPI, and pmemd.cuda. What I am observing is differences in energies
>calculated when pmemd.cuda is utilized.
By default, pmemd.cuda is linked to pmemd.cuda_SPFP, which does some
(many) calculations in single precision, and will not exactly match the
other three, which are all fully double-precision codes. Generally, if
you run pmemd.cuda_DPFP (the slower, double-precision version) you will
get closer agreement with the CPU codes.
Amber developers (and other users), having runs tons and tons of SPFP
simulations, are pretty convinced that these precision difference do not
affect any statistical properties of the simulations.
....hope this helps....dac
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Thu May 14 2020 - 18:30:02 PDT
