Re: [AMBER] Inconsistent results on TI with cpu vs gpu code

From: David A Case <>
Date: Sun, 9 May 2021 15:56:04 -0400

On Sun, May 09, 2021, Giorgos Lambrinidis wrote:

>I am running a TI simulation using the method of Thomas Steinbrecher and
>Hannes H Loeffler
>My system has a complex of protein+ligand embedded to lipid bilayer.
>I ran the production of TI with CPU and GPU separately.
>Using pmemd.cuda the results were:
>dG sum for Complex: 6.4 dG sum for Ligand: 6.2 and
>When I run the simulation using pmemd.MPI the results were
>dG sum for Complex: 5.6 dG sum for Ligand: 6.3 and

I am not a TI person, so I hope others will chime in. My guess is that
everyone will need more information. To start:

1. Since not everyone will be familiar with your cited reference, you might
briefly explain what is meant by "dG sum" and "ΔΔG". Are you trying to get
relative binding free energies for two different ligands, or something else?

2. How well-converged are the results on either CPU or GPU? If you run
a second simulation, indentical except for random numbers or starting
velocities, how close are the results to the first one? It's also probably
worth looking at the results for each lambda value (comparing GPU with CPU):
is the overall difference dominated by values near one end or the other of
the lambda series?

>I used on cuda code the -AllowSmallBox flag which was not the default choice.

My understanding is that this flag is dangerous, but (if I'm interpreting
things correctly), you only used this for the ligand, where the results seem
to be the same as for the CPU.


AMBER mailing list
Received on Sun May 09 2021 - 13:00:02 PDT
Custom Search