VMD-L Mailing List

From: John Stone (johns_at_ks.uiuc.edu)
Date: Fri Oct 31 2014 - 14:45:04 CDT

Hi,
  A brief note about ECC specifically:

On Fri, Oct 31, 2014 at 01:35:02PM -0500, Josh Vermaas wrote:
[...]
> 3.) Error correcting memory. Useless for VMD purposes as far as I can
> tell. For running MD simulations themselves, useless as well, and may
> actually be harmful. See [1]http://dx.doi.org/10.1145/2484762.2484774 for
> a comparison study that some of the Amber guys did in collaboration with
> NVIDIA.

I would state that in general, when choosing a GPU for use in a
desktop workstation, ECC is not a primary consideration, and
many other factors rank much more highly.
For VMD in particular, running on a desktop workstation, ECC is not likely
to be necessary in a normal usage scenario, and the only case when it would
make sense to buy an ECC-enabled GPU would be when the host machine ALSO
has ECC protection. If the GPU is the only component in the system with
ECC, then you've spent extra money for zero benefit. The whole system
has to have ECC through-and-through for there to be a tangible benefit.

This reference gave results for the first GPU hardware generation that
had an incomplete ECC implementation that covered DRAM but not various
parts of on-chip shared memories, caches, etc. In my personal opinion,
the results that relate specifically to the ECC hardware of that time
have very limited value in the context of evaluating the use of ECC
on current generation GPU devices, and particularly from the perspective
of the use of ECC on large supercomputers such as the Cray XK7, XCx0, etc.

Cheers,
  John 

-- 
NIH Center for Macromolecular Modeling and Bioinformatics
Beckman Institute for Advanced Science and Technology
University of Illinois, 405 N. Mathews Ave, Urbana, IL 61801
http://www.ks.uiuc.edu/~johns/           Phone: 217-244-3349
http://www.ks.uiuc.edu/Research/vmd/