From: Axel Kohlmeyer (akohlmey_at_cmm.chem.upenn.edu)
Date: Mon Jul 06 2009 - 16:46:40 CDT
On Mon, 2009-07-06 at 12:39 -0600, Olve Peersen wrote:
> From what John said, it seems to make more sense to just jump to a
> Tesla rather than a GeForce card if we really wanted to do GPU
> computing. So what do you think of starting out with more basic video
> card such as a GeForce 9800GT, 250GTS, or 260GTX for non-stereo
> graphics rendering work? Or put another way, if VMD is just rendering
i would suggest going with a recent variant GTX265. this is basically
an underclocked GTX285 with a couple of disabled shaders. they have
in my opinion the by far best price/performance ratio, and don't
need too much power, but are still very close to the top performing
cards in both OpenGL and GPGPU compute performance. that should give
you a good insight into what GPU performance you can expect before
investing into additional GPUs.
> trajectories, how much GPU power does it need? This configuration
> should give me enough residual electrical power in the system to add a
> C1060 or a headless 285GTX/295GTX in the future as a dedicated GPU
> node, and if we go beyond that then we'll be dealing with a dedicated
> server configuration.
if you want to go totally GPU crazy, you could even buy a tesla
S1070 with two 16xPCIe riser cards and then add 4 additional GPUs
to your box without and additional PCIe power requirements (the
tesla node has its own power supply). but those boxes are not
cheap on an absolute scale (they rock if you have an application
that can run on them 30-50x faster on a single GPU than on a single
CPU core).
cheers,
axel.
>
> Cheers,
>
> Olve
>
>
>
> On Jul 6, 2009, at 10:04 AM, John Stone wrote:
>
> >
> > Hi,
> > Axel's suggestions are good. The main comment I have is
> > that the primary benefit of the Quadro GPUs is larger memory
> > capacity and support for stereoscopic rendering. If you need
> > support for quad buffered stereoscopic rendering, then there's
> > not presently much of an alternative to the Quadro cards.
> > If you just need the large memory capacity in order to use CUDA
> > for electrostatics calculations or the like, then a less costly
> > alternative is to buy a Tesla C1060. If you don't need either of
> > these features, then the consumer video game GPUs such as the
> > GeForce GTX 285 may work fine for you.
> >
> > You should consider whether you'll be using the GPUs for
> > long-running CUDA calculations as well, e.g. NAMD, and
> > if so, then getting Tesla C1060s for that purpose would be
> > advised since they undergo much more reliability testing than
> > the the consumer video game GPUs do.
> >
> > Cheers,
> > John Stone
> > vmd_at_ks.uiuc.edu
> >
> > On Fri, Jul 03, 2009 at 12:26:11AM -0600, Olve Peersen wrote:
> >> ========= Note: This message has been sent to both NAMD and VMD
> >> mailing lists ==============
> >>
> >> We're looking into buying a new Linux box for structural biology
> >> refinement and some molecular dynamics work (i.e. equilibration runs
> >> and making sure input files are right before handing jobs off to
> >> bigger clusters) and I wanting to get some input on where it is worth
> >> spending money on the system from a NAMD/VMD perspective. Our
> >> experimental systems are proteins of 150-500 residues in size, and we
> >> already have some experience running NAMD/VMD on an older Pentium 4D
> >> system with a Quadro FX1300 and hardware stereo.
> >>
> >> At the moment I am leaning toward a dual Xeon 55XX (quad-core) system
> >> with 12 GB of RAM - this is the Xeon version of the quad-core Core i7
> >> Nehalem processor. My choice will probably end up being the midrange
> >> 5520 or 5530 processors (2.26 vs 2.40 GHz) unless there is a really
> >> good computational reason to go to the faster X series, where I
> >> suspect I'll be mostly paying a "bleeding-edge premium".
> >>
> >> For NAMD it seems that more and faster processors, more RAM, and fast
> >> disks are the key elements. I'm not up for doing full cluster
> >> architecture, so any inter-machine connections would be gigabit for
> >> the time being. I also suspect I'd be better off compiling with the
> >> Intel compilers and certain Xeon specific flags as some processor
> >> specific tweaking is apparently worthwhile for the new Nehalem
> >> architecture, but that is something I have not seriously looked into
> >> yet.
> >>
> >> The question I've had a harder time nailing down is how dependent VMD
> >> performance is on the graphics card versus the core system itself and
> >> how powerful a 3D graphics card I should invest in. In our
> >> experience
> >> thus far, rendering of our systems seems fine with our old Quadro
> >> FX1300 and I was thinking of going with a 768 MB Quadro FX1800 for
> >> this new system, but I could spend another $400 and get a 1GB FX3800,
> >> or get really carried away with ~$1700 for a FX4600/4800 card. Is
> >> there a compelling performance reason for doing so?
> >>
> >> Thanks in advance for any input,
> >>
> >> Olve
> >>
> >> -------------------------------------------------------
> >> Olve Peersen
> >> Associate Professor
> >> Dept. of Biochemistry & Molecular Biology
> >> 1870 Campus Delivery
> >> Colorado State University
> >> Ft. Collins, CO 80523-1870
> >> -------------------------------------------------------
> >> 970.491-0433 Office
> >> 970.491-0271 Lab
> >> 970.491-0494 Fax
> >> -------------------------------------------------------
> >>
> >
> > --
> > NIH Resource for Macromolecular Modeling and Bioinformatics
> > Beckman Institute for Advanced Science and Technology
> > University of Illinois, 405 N. Mathews Ave, Urbana, IL 61801
> > Email: johns_at_ks.uiuc.edu Phone: 217-244-3349
> > WWW: http://www.ks.uiuc.edu/~johns/ Fax: 217-244-6078
-- ======================================================================= Axel Kohlmeyer akohlmey_at_cmm.chem.upenn.edu http://www.cmm.upenn.edu Center for Molecular Modeling -- University of Pennsylvania Department of Chemistry, 231 S.34th Street, Philadelphia, PA 19104-6323 tel: 1-215-898-1582, fax: 1-215-573-6233, office-tel: 1-215-898-5425 ======================================================================= If you make something idiot-proof, the universe creates a better idiot.
This archive was generated by hypermail 2.1.6 : Wed Feb 29 2012 - 15:53:00 CST