From: Norman Geist (norman.geist_at_uni-greifswald.de)
Date: Tue Aug 21 2012 - 01:22:49 CDT
as I see it, the 680s have 1/3 less clock rate, but therefore 3x more CUDA
cores. So for one GPU they should be faster by a factor of two.
In HPC the required bandwidth of any link usually scales with the absolute
computing power of the link using/sharing units (neglecting latency here).
That means, as the 680s are two times faster, they will scale worse than the
580s with the same bandwidth. And as we all know, the PCIE bandwidth is
already a bottleneck for namd and GPUs with PCIE2. Because of that reasons,
you should go for the PCIE3 when going with the Kepler and make sure you
have enough full occupied x16 slots with PCIE3 bandwidth and also keep an
eye on the power supply, one of these GPUs will take 195W I guess.
Von: owner-namd-l_at_ks.uiuc.edu [mailto:owner-namd-l_at_ks.uiuc.edu] Im Auftrag
von Aron Broom
Gesendet: Montag, 20. August 2012 18:21
An: Blake Mertz
Betreff: Re: namd-l: GPU workstation build
In terms of the bandwidth and the integrated PCI-e bridge, I think you might
get a sense from the motherboard manual, as they will generally say
something like "one at 16x or two at 8x each". I guess ideally you want to
be able to tos 4 cards in there and get 16x for all of them.
I'm also quite interested in the performance for the GTX680s. One thing to
note, although I'm not 100% certain of this, so something can correct me, I
think that without using the new PCI-e 3.0 (which the Keplers are capable
of) you'll actually have lower bandwidth with a 680 compared with a 580, and
you only get back to 580 bandwidth levels when you are using the 3.0.
On Mon, Aug 20, 2012 at 11:41 AM, Blake Mertz <blake.mertz_at_mail.wvu.edu>
I'm in the process of putting together a GPU workstation based off of
a Tyan S8232GM4NR motherboard -- it's a dual-socket AMD G34, with 4
PCI-e Gen2.0 slots that are spec'd to run at 16x
My two biggest concerns are:
1) the bandwidth on my motherboard with respect to the GPUs, and
2) which commercial Nvidia GPUs I should go with.
>From what I can tell by Axel Kohlmeyer's response to Matt Roark's
you have to look out for motherboards with integrated "PCIe bridges",
which I haven't been able to confirm on the S8232, either in the
manual or from looking online. Does anyone have experience using this
board for namd?
Also, I noticed Francesco Pietra has done some testing with GTX 680s
(which is what I happened to initially purchase), but I haven't seen
much else on namd-l discussing these cards. Has anyone else had a
chance to compare the 680 performance over the 580s? I've had a hard
time judging the advantages of the new Kepler architecture over Fermi,
especially in context of namd performance. Thanks for any help you can
-- Assistant Professor C. Eugene Bennett Department of Chemistry West Virginia University "Life is not easy for any of us. But what of that? We must have perseverance and above all confidence in ourselves. We must believe that we are gifted for something and that this thing must be attained." Marie Curie "Start by doing what's necessary; then do what's possible; and suddenly you are doing the impossible." St. Francis of Assissi -- Aron Broom M.Sc PhD Student Department of Chemistry University of Waterloo
This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:21:55 CST