Re: [AMBER] Performance of RuBisCo (74, 916) atoms on NVIDIA Tesla K40 GPU

From: David A Case <>
Date: Tue, 12 Aug 2014 07:37:30 -0700

On Tue, Aug 12, 2014, Amit Harode wrote:
> I was using DPFP precision model because of its higher accuracy. I am using the following "mdin" file:
> &cntrl
> irest=0,
> ntx=1,
> igb=2,
> rgbmax=14.0,
> cut=9999.0,
> ntpr=10,
> ntwx=100,
> nstlim = 500,
> dt=0.001,
> ntt=3,
> tempi=300.0,
> temp0=300.0,
> ntb=0,
> nscm = 0,
> &end

OK...I think this is what you should expect. A GB calculation with 74,000
atoms will indeed be very slow, espcecially on DPFP. We don't support cutoffs
with GB on the GPUs, so the number of nonbonded interactions is enormous.

One alternative would be to use the Heirarchical Charge Partioning method
on a CPU (see section 37.6 of the Amber14 Manual). Or, using a periodic
system with explicit solvent (on the GPU) should also be faster than GB
for a system this big. Or, run GB on the CPU version of pmemd.MPI, but use a
cutoff of 25 Ang. or so.

...hope this helps....dac

AMBER mailing list
Received on Tue Aug 12 2014 - 08:00:03 PDT
Custom Search