GTX580 Classified Still King in new Barefeats Multi-GPU tests

Discussion in 'Mac Pro' started by MacVidCards, Jul 15, 2013.

  1. MacVidCards Suspended

    Joined:
    Nov 17, 2008
    Location:
    Hollywood, CA
    #1
    http://barefeats.com/gpu680v6.html

    Nvidia took a knife to parts of GPGPU paths in new Kepler cards to keep Quadro and Tesla cards at top.

    But the GTX570 and GTX580 never got that treatment.

    Note that Dual Classified 580s wins every test except one where our GTX770 takes the lead.

    Interesting results, one does wonder what will happen to all of these CUDA based apps in 6 months.

    In the "truly embarrassing" category, a SINGLE Classified 580 beat 2 (TWO) new Quadro K5000s in the OpenCl test. $3600 of video card brought to it's knees by a single 580C. Ouch !!!
     
  2. DJenkins macrumors 6502

    DJenkins

    Joined:
    Apr 22, 2012
    Location:
    Sydney, Australia
    #2
    Cool thanks MVC.

    So with your insight into the kepler cards being restricted in performance, where would that place the Titan? I wish they tested that card as well!

    Just to confirm, you are saying that the Kepler GTX series have been restricted in some way to make the Quadro and Tesla cards look like they perform better... where in actual fact they are still outperformed by the Fermi series?

    Of course I'm talking strictly in regards to these tests performed, they may perform better at other things... but as in a sense of video production I would view these tests as the most relevant to a typical workflow.
     
  3. GermanyChris macrumors 601

    GermanyChris

    Joined:
    Jul 3, 2011
    Location:
    Here
  4. fhenry macrumors regular

    Joined:
    Feb 28, 2012
  5. Tutor macrumors 65816

    Tutor

    Joined:
    Jun 25, 2009
    Location:
    Home of the Birmingham Civil Rights Institute
    #5
    Within the next week or so, I'll run it with 1 through 8 Titans.
     
  6. linuxcooldude macrumors 68020

    Joined:
    Mar 1, 2010
    #6
    How does a GTX 570 compare to a single Quattro K5000?
     
  7. MacVidCards thread starter Suspended

    Joined:
    Nov 17, 2008
    Location:
    Hollywood, CA
    #7
    Click on the link and see. He also has 2 of our GTX570 in the "sweet spot" 2.5 GB size.
     
  8. Tutor, Jul 17, 2013
    Last edited: Jul 17, 2013

    Tutor macrumors 65816

    Tutor

    Joined:
    Jun 25, 2009
    Location:
    Home of the Birmingham Civil Rights Institute
    #8
    The real king of GTX GPUs - for now.

    I. Here's one of my Titan's performance of that AE CUDA test used by Barefeats: 9 min. 16 secs (faster than test run with 3 GTX 580 Classified). Here's a link to Danny's comparison chart [ https://docs.google.com/spreadsheet...GamhiUkIySTUteGlzeG9xMEE&oid=6&zx=4b87cbodocm ].

    II. Concerning the Octane Render CUDA test- Benchmark.ocs test used by Barefeats, here is how my Titans perform [(all but 1 only in seconds) / Barefeats uses minutes { http://barefeats.com/gpu680v6.html }]:
    8 GPUs - 10 secs,
    7 GPUs - 11 secs,
    6 GPUs - 13 secs,
    5 GPUs - 15 secs,
    4 GPUs - 19 secs,
    3 GPUs - 26 secs,
    2 GPUs - 39 secs and
    1 GPU - 76 secs (1 min. 16 secs - almost as fast as test run with two GTX 580 Classifieds + Quadro 4000).

    These results show a significant amount of linearity with the addition of GPUs, i.e., the render time for 2 GPUs is about 1/2 of the time for 1 GPU, the time for 4 GPUs is about 1/4 of the time for 1 GPU, and the time for 8 GPUs is about 1/8 of the time for 1 GPU. The same applies to the intervening fractional amounts. Below is a pic of what that scene looks like (test in pic run with RenderTarget DL with 8 GPUs -took 3 secs).


    III. The Luxmark Open CL scores for my Titans are on the following pages, both under the user name of TheRealTutor and all of the Titan scores without a user name (regardless of the number of Titans being tested):
    A. Overall Top 20 Complex Benchmark ("Room" scene) - http://www.luxrender.net/luxmark/top/top20/Room ;
    B. Overall Top 20 Medium Benchmark ("Sala" scene) - http://www.luxrender.net/luxmark/top/top20/Sala ;
    C. Overall Top 20 Simple Benchmark ("LuxBall HDR" scene) - http://www.luxrender.net/luxmark/top/top20/LuxBall HDR . My "room" scene score with one Titan GPU is 1760 [ http://www.luxrender.net/luxmark/se...tion_SEPARETOR_GeForce GTX TITAN_SEPARETOR_14 ] vs. 1345 for two GeForce GTX 580 Classifieds + Quadro 4000 [ http://barefeats.com/gpu680v6.html ]. Of course, high end ATI cards are better performers in Luxmark tests.

    IV. I don't have any test results yet for Da Vinci Resolve.
     

    Attached Files:

  9. fhenry macrumors regular

    Joined:
    Feb 28, 2012
  10. i make movies macrumors regular

    Joined:
    Aug 9, 2007
    #10
    Do you have Resolve? Like the paid version? There is a free version, which is great, but one of the limitations of the free version is only using one GUI card and one CUDA card tops.

    Also, the problem I have with the "Candle Test" is at a certain point you hit a CPU bottleneck (or a disk speed bottleneck), especially if you're looping Quicktime ProRes files. Also, not all codecs utilizes GPU acceleration

    The only way to truly test a GPU(s) in Resolve to render a clip that's at least 2 minutes. But, you need to test disk speed of target and destination drive to make sure they're fast enough.
     
  11. Photovore, Jul 19, 2013
    Last edited: Jul 19, 2013

    Photovore macrumors regular

    Joined:
    Dec 28, 2011
    #11
    my tuppence -- Moses on a moped!

    I just replaced a $500 Mac Edition AMD 5870 with a used nVidia 570 (1573AR) offa ebay for ~$137 including shipping and a dollar donation to something that sounded reasonable....

    [ Background: I have an OpenCL kernel that takes 155ms to run on the nV 330M in the MBP. I had hoped/estimated (from luxmark benchmarks more than stated max FLOPs) that the 5870 might give me a 10x speed bump over the laptop, i.e. 15.5ms frames, which might perhaps then display at 60fps. Nope; no matter how hard I optimized the code, vectorized, whatever, I only got about a 4x speed bump (that kernel on the MP/5870 took 45ms to calculate). To render fast enough for a show, I cut the resolution from 1024x640 to 800x500, to get it to <30ms (displaying 30fps), which is juuust adequate.... ]

    In the luxmark tests, IIRC, the 570 and 5870 scored about the same. BUT, given that the 570 uses nV architecture, I thought I might get a better percentage of theoretical maximum on the nV card. Specifically, I hoped/half-expected to get the full 10x bump over the laptop, or in other terms, a 2.5x throughput bump compared to the 5870. I.e., that the kernel might in fact calc in under 16ms and yield 60fps at 1024x640....

    Incorrect assumption! The kernel (at 1024x640) calculates in UNDER 5 milliseconds on the 570, which is ~30x the performance on the laptop's 330M, or >9x the performance of the 5870!!!!

    This means that it can now do FULL HD (1920x1080) in <15ms calc time, i.e. at 60fps!!!!

    This is stunning.
     
  12. Tutor macrumors 65816

    Tutor

    Joined:
    Jun 25, 2009
    Location:
    Home of the Birmingham Civil Rights Institute
    #12
    I don't have that application yet. I was thinking about getting the free version, but the limitations that you describe are leading me to wait until I can get the full version. Thanks for the heads up.
     
  13. juanm macrumors 65816

    juanm

    Joined:
    May 1, 2006
    Location:
    Fury 161
    #13
    Hi Tutor. I'm thinking of building a small rendering rig for Arion/Max Live. Unfortunately, I don't have the budget for a multi-Titan setup. A GTX760 4GB costs about 1/4 of the Titan. How would 4x GTX760 4GB fare against 1x Titan ? Any experience or advice?
     
  14. i make movies macrumors regular

    Joined:
    Aug 9, 2007
    #14
    No worries. The full version comes with a USB key (and costs $999).

    Do you have any friends in the coloring or post production who have the usb key to lend?
     
  15. Tutor macrumors 65816

    Tutor

    Joined:
    Jun 25, 2009
    Location:
    Home of the Birmingham Civil Rights Institute
    #15
    Nope. I'll just wait.
     
  16. 666sheep, Jul 20, 2013
    Last edited: Jul 20, 2013

    666sheep macrumors 68040

    666sheep

    Joined:
    Dec 7, 2009
    Location:
    Poland
    #16
    Check this out: http://www.tomshardware.com/reviews/geforce-gtx-760-review-gk104,3542-19.html

    In general, 760 is about a half performance of Titan, so 4 in linearly scaling app should give computing power of 2 Titans. Their estimated power consumption will be about 2.6 Titans. Not bad.

    But if you want best bang for buck, consider 2 or 3 used 580s. About ~$200 (or less) per one.
     
  17. juanm macrumors 65816

    juanm

    Joined:
    May 1, 2006
    Location:
    Fury 161
    #17
    Thanks, that's pretty much the conclusion I had come to.
    The drawbacks of going for the 580 are the RAM (3GB vs 4GB) and the TDP of 244W/card (vs 170W/card for the GTX 760). for 50€ less/card, I don't know whether it's a good deal (I'm in Europe, and 2nd hand cards are not abundant, apparently, and I don't have much time to spend looking).
     
  18. Tutor, Jul 22, 2013
    Last edited: Jul 22, 2013

    Tutor macrumors 65816

    Tutor

    Joined:
    Jun 25, 2009
    Location:
    Home of the Birmingham Civil Rights Institute
    #18

    Juanm,

    I agree with 666sheep that for the more cost conscious, the 760s and 580s would be good choices. The downside that I've had with the 580s is that they consume lots of power. Also, be aware that the fastest and most powwerful 580s each require 2 - eight pin plus 1 - 6 pin power connectors. If you go with the 580's, you'd be better off getting them refurbished from somewhere like EVGA's B-Stock where you'd at least get some kind of warranty.
     
  19. registudio macrumors member

    registudio

    Joined:
    Aug 8, 2013
    Location:
    Montreal, QC
    #19

    Or it is possible to get such card from macvidcards with a limited warranty.

    @Tutor, I'm in the process to set up my Mac Pro (Rev5,1) for octane render and I would really appreciate your advice.
    Indeed I read one of your very interesting post (http://forums.macrumors.com/showthread.php?t=1565650&page=5) but I was wondering if the GTX Titan in a cubix elite will work or at least 1 in the slot 2 will work on Mac OS X and not Windows?
    If not, I was thinking/planning to get 2xGTX580c and 2xGTX770 (2or4GB).
    What do you think?

    Thanks and Regards
    Regis
     
  20. DJenkins macrumors 6502

    DJenkins

    Joined:
    Apr 22, 2012
    Location:
    Sydney, Australia
    #20
    The cubix boxes work fine with mac pros, a lot of people have been using them for davinci resolve for ages. As far as the best cards, your choice will actually be made easier by the ease of power availability in the cubix, compared to stuffing them inside the mac pro case as most people do. If your only purpose is octane render, get as many mac friendly cards with the most cuda cores possible. Although GTX580 cards are still better for some tasks, they have significantly fewer cuda cores than the newer cards.

    Octane render is definitely Tutor's thing though, so I'm sure he'll guide you more than my rambling can.
     
  21. MacVidCards thread starter Suspended

    Joined:
    Nov 17, 2008
    Location:
    Hollywood, CA
    #21
    A certain benchmark site might have some interesting info soon.

    And BTW, we will start offering some exciting EFI cards in next few weeks.
     
  22. registudio macrumors member

    registudio

    Joined:
    Aug 8, 2013
    Location:
    Montreal, QC
    #22
    wow looking forward to this MacVidCards as I will have to purchase very very soon.

    Cheers
    Regis
     
  23. registudio macrumors member

    registudio

    Joined:
    Aug 8, 2013
    Location:
    Montreal, QC
    #23
    Thank you DJenkins for your input.
     
  24. fhenry macrumors regular

    Joined:
    Feb 28, 2012
  25. spudfan macrumors newbie

    Joined:
    Aug 13, 2013
    #25
    GPU for resolve

    hi i cant figure out what would be the best solution for me.
    Curerrently using a MP 2012 hex with a quadro4000 + GT120 gui, this machine will only be using resolve

    can some one please point me in the right direction,

    one option with cubix and one with internal cards

    /thanks
     

Share This Page