[gmx-users] gpu k20x vs k40

[Mirco Wahab]

On 20.05.2014 08:13, Andrei Neamtu wrote:
> We are currently in the process of buying several servers which will
> include GPU accelerato.
> Because of the price / performance balance in case of our limited funds we
> have to choose between K40 and K20X tesla accelerators.
> Initially, we wanted 2 GPU's/node (dual Xeon E5-2680v2 2.8GHz) but as
> expected the funds were  enough for only 1 K40 per node. Instead, it may be
> feasible to be able to put 2 K20x in each node.

The difference in performance between K20X and K40 is rather
marginal. The main advantage of the K40 is RAM (which you
won't probably ever need with GROMACS).

> We tried to make a test through NVidia TestDrive but the things take too
> long. (still :) )
> My question is if anybody can make a suggestion of which should be more
> convenient for GROMACS performance: 2 x K20x  versus 1 K40 per node. In the
> absence of any testing opportunity any suggestion is extremely valuable to
> us.

What kind of "box" is this? If it isn't server rack based, you
could even give GTX-780Tis a try (for workstation setups).
These will be, with single precision GROMACS, much faster
than even the K40s (while possessing less RAM than the K20X).

The answer really depends on your circumstances. What type of
servers, what type of planned simulations, etc.

Regards

M.