[gmx-users] GTX card vs Kepler card

[Mirco Wahab]

On 22.01.2014 15:03, Jianguo Li wrote:
> I have an doubt on the GPU hardware. Recently I heard from a nvidia person that unlike Kepler card, GTX card does not have the memory check and correction. It is no problem for gaming, but may lead to wrong results for scientific computing.

"Fermi", "Kepler", and "Maxwell" are the names of the (evolving)
core-architectures, whereas "Tesla" (professional) and "GeForce"
(consumer) designate the marketing segment of the product.

The products of the "professional" segment will generally have more
capabilities enabled by drivers and/or additional hardware options
(like memory correction, dynamic parallelism and load reporting)
for a significantly distinct price.

"GTX" means (imho): the enthusiast products within the consumer
market segment.

You'd therefore mean "the difference in reliability (e.g. through
memory error detection) between a Tesla card and a GTX enthusiast
card"?

> I am just wondering how this will affect the accuracy of gromacs simulations? Does the lack of memory check/correction lead to significant error?

I don't know for sure, but from own experiences over the years I'd think
the major source of "right" and "wrong" in Gromacs is the proper
application of the correct force field.

Memory errors on a device, either workstation (most group servers
use commodity hardware without any ECC) or graphics processors, have
been - as I remember, rather easily detected - either by crashing
programs (segmentation faults, catastrophic parameter drifts) or by
statistical deviations within a group of similar runs on order to
scan over an input parameter range.

Regards

M.