On Wed, Apr 13, 2016 at 8:39 PM, Leonardo Lazzaro <lazzaroleonardo.gmail.com
> wrote:
> Dear experts,
>
> I am building a setup for amber14 with two gtx980.
> However I have low budget and using Intel drive us to buy a cpu with enough
> pci lanes to have both gpu using using the pci-e at 16x.
>
> Since the chipset for AMD 990fx allows us to have both cards at x16, do you
> think that using an AMD setup with the chipset 990fx will have any
> performance impact?
>
Since pmemd.cuda runs entirely on the GPU, there will be little to no
noticeable performance hit to using slower CPUs. I use a slow, budget AMD
CPU with 6 cores and see identical performance with my card compared to
higher-end machines (at least back when the GTX 680 was considered "l33t").
> Also how bad is the impact when both card run on x8? the initial idea was
> to have an Intel with 16 pci lanes like a i5 4690k.
>
Depends. For parallel runs, bad. For serial runs, fine (as long as you
don't try to write trajectory frames too frequently). Serial runs avoid
much communication over the PCIe bus, so dropping to x8 isn't much of a
hit. But parallel runs communicate through it, so the reduced bandwidth
will hurt.
HTH,
Jason
--
Jason M. Swails
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Wed Apr 13 2016 - 19:30:04 PDT
