GPU performance

Dear Axel, Steve, and other developers

Would you please let me know the conceptual difference of cutoff set for vdw and coul forces when using kspace?

I’ve gone through the mailing list and came across a discussion in which Axel claimed increasing the coul cutoff may enhance calculation speed while using gpu package. So how and why?

Any hint is welcome

The vdw cutoff and Coulomb cutoff apply to different terms, which have
different errors as a function of cutoff. So LAMMPS lets you apply
two differerent cutoffs if you want. Most people don’t.

If you are doing long-range Coulombics, there is no physical cutoff,
you are computing them out to infinity. in that case the Coulomb
cutoff is simply a way to partition work between the short-range (pair style)
and the long-range (Kspace style). Typically you want the
cost of both parts to be roughly equal. On a CPU versus a GPU
that optimal Coulomb cutoff may be different, since the short-range
part may run much faster on a GPU.