[gmx-users] Re: CUDA with QUADRO GPUs

[Szilárd Páll]

Dear Ramon,

Thanks for the kind words!

On Tue, Jun 18, 2013 at 10:22 AM, Ramon Crehuet Simon
<ramon.crehuet at iqac.csic.es> wrote:
> Dear Szilard,
> Thanks for your message. Your help is priceless and helps advance science more than many publications. I extend that to many experts who kindly and promptly answer questions in the mailing list. I wish that could be valued when evaluating a CV. Sorry for the digression...
> The visualization I had in mind was the one most of us do: MD trajectories of biomolecules. In particular I use VMD. From you message, I deduce that a GTX card would work as good as a quadro for a much lower price. Right?

Molecular visualisation does not require Quadro cards, GTX works just
fine. In fact, when it comes to the redering itself even AMD GPUs
should be just fine (although vmd can use NVIDIA GPUs for accelerating
compute tasks).

> The problem is that most worksations come with quadro cards, not GTX. In particular I was looking at Dell precision workstation. If you want a GTX card you have to go to a gaming PC. Which is probably ok if the administration understands it... :-)

I'm pretty sure you should be able to get a workstation with e.g. only
Intel integrated graphics and add another discrete GTX GPU to be used
for compute. In fact, it can be advantageous having separate display
and compute GPUs - especially if you plan to run Windows. The only
drawback of using a weak GPU for driving the display is that
visualisation of large/complex molecular systems also requires a fast
GPU. If that turns out to be the case, one can always disable the
integrated graphics and use the discrete GPU for both display and
compute.

> The GTX cards lack ECC. Is that an issue for running gromacs? or the memory errors only result in adding a negligible stochastic term to the dynamics?

In our experience ECC is not really needed for MD - at least not to be
able to run simulations. Of course, error correction can always come
handy, but the question is at what cost. While data corruption e.g.
due to soft errors can (and does) happen, the chance of such an event
going unnoticed in the results of an MD simulation is IMHO quite low -
most of the time such an event will manifest as a crash, or corrupted
frame.

Cheers,
--
Szilárd

> Thanks again,
> Ramon
>
>
>
> Message: 5
> Date: Tue, 18 Jun 2013 00:20:00 +0200
> From: Szil?rd P?ll <szilard.pall at cbr.su.se>
> Subject: Re: [gmx-users] CUDA with QUADRO GPUs?
> To: Discussion list for GROMACS users <gmx-users at gromacs.org>
> Message-ID:
>         <CANnYEw7Y3Otq2zGG8G9E7+q=oa8tRZWLY31zYT6mAyE9u1f-sA at mail.gmail.com>
> Content-Type: text/plain; charset=ISO-8859-1
>
> Dear Ramon,
>
> Compute capability does not reflect the performance of a card, but it
> is an indicator of what functionalities does the GPU provide - more
> like a generation number or feature set version.
>
> Quadro cards are typically quite close in performance/$ to Teslas with
> roughly 5-8x *lower* "GROMACS bang for the buck" compared to consumer
> (GTX) cards. You can find a bit more detailed and more general
> explanation here: http://goo.gl/80cQQ
>
> More concretely, the K4000 is roughly as fast as a GTX 650 Ti, a
> low-end consumer GPU (~120 Eur), for ~6x the price. For the price of
> the K4000 one can get a high-end card, e.g the GTX780 which is among
> the fastest NVIDIA GPUs to date (close to the Tesla K20 in single
> precision). Hence, unless you really *need* the capabilities of a
> Quadro card, that is if the visualization you are referring to means
> CAD or similar *professional* graphics work, I suggest getting
> consumer cards.
>
> If you want to compare cards, from the point of view of GROMACS 4.6,
> pretty much the only thing that matters is the number of cores (or
> multiprocessors) times the GPU clock frequency. Note that comparing
> cards between different generations (different "major" compute
> capability) is not as straightforward.
>
> Cheers,
> --
> Szil�rd
>
> --
> gmx-users mailing list    gmx-users at gromacs.org
> http://lists.gromacs.org/mailman/listinfo/gmx-users
> * Please search the archive at http://www.gromacs.org/Support/Mailing_Lists/Search before posting!
> * Please don't post (un)subscribe requests to the list. Use the
> www interface or send it to gmx-users-request at gromacs.org.
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists