Here are some more data points on the latest cuda code drop by Scott:
1. The jac benchmark crashes with the DPDP version with the error:
Error: invalid argument launching kernel kNLBuildNeighborList
(I reported this earlier, but on my system; here it's on one of the standard
benchmarks, so should(?) be easier of others to reproduce.)
2. The new SPDP code is indeed 60-70% faster than before for NVE calculations
of solvated proteins, with the default cutoff of 8 Ang. (2 separate tests
of this, including the jac benchmark). But it is about 3 times *slower*
than the old code for our crystal systems, which are NPT, have 8,000 to
12,000 atoms, and use a cutoff (vdw and electrostatic) of 12 Ang.
I'm in the process of interpolating conditions between these extremes to
try to pinpoint the symptoms better. For a small (8000 atom) system
with the default cutoff of 8 Ang., here are the timings for a 15000-step
run: (non-setup wall time):
old new
NVT 63 47
NPT 73 59
So: based on this very small sample, the problem with the new code shows up
when one uses big cutoffs, doesn't seem to have much to do with NPT vs NVT.
Time for bed...more later...
....dac
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Thu Feb 10 2011 - 20:30:05 PST