[gmx-users] gromacs+CHARMM36+GPUs induces protein-membrane system collapse?

[Justin Lemkul]

On 10/31/18 10:47 AM, Ramon Guixà wrote:
> Hi Justin,
>
> I do not have access to GPUs with ECC, but I have actually run the same tpr
> in different machines with different GPUs, including GTX 980, 1060 and
> 1080. In all cases, I get the same problem. Although they are all  non-ECC
> cards, it is quite unlikely that a bit flip is behind this, isn't it?

Certainly is. It wasn't clear if you were just running over and over 
again on the same card or what.

> In fact, I had simulated before CHARMM-GUI systems using GTX cards and
> never had a problem. Since this has only benn happening recently to me, so
> I am wondering if I should blame the new force field (CHARMM36m) or whether
> CHARMM-GUI changed something in the way they generate gromacs topologies?

I sincerely doubt it's the force field. We validate that extensively 
across many compounds to verify that GROMACS and CHARMM compute the 
forces the same (I do this myself when we produce a new force field 
port). We don't (can't) do that easily on GPU, but if the forces are 
different, that points to a software bug, because everything about the 
CPU implementations is identical. Given that you have a sudden change in 
system behavior, that further points to a software bug, in my opinion, 
because if something was fundamentally wrong with the force field, (1) 
you would see it on CPU runs and (2) it would almost certainly occur 
much earlier in the simulation.

-Justin

-- 
==================================================

Justin A. Lemkul, Ph.D.
Assistant Professor
Virginia Tech Department of Biochemistry

303 Engel Hall
340 West Campus Dr.
Blacksburg, VA 24061

jalemkul at vt.edu | (540) 231-3129
http://www.thelemkullab.com

==================================================