netkas.org

Some gfx benches

Wanted to compare quality of nvidia/ati drivers in osx, here is results:

Test setup – 10.6.7 OSX, nvidia drivers – latest nvidia drivers for mac from nvidia site (256.01.00f03), C2Q @ 3GHz, 8GB ram

Used cs:source from steam, custom demo.

Settings – 1680×1050, all settings maxed + 4xMSAA + vsync off

Results:

GTX285 – 134.32 fps (something wrong with drivers in some moments of demo, really it has much better fps)
8800gt 1gb – 167.39 fps
4850 – 184.58 fps
GTX480 – 217.12 fps
5870 2gb – 228.75 fps

Comments

  1. miqlas
    April 5th, 2011 | 5:15 pm

    Hi Netkas!

    Thanks for the info! Do you make 10.7 bench too?

  2. npwski
    April 5th, 2011 | 6:13 pm

    Thank you, very useful info. I thought that my 5870 is too slow..

  3. arobasefr
    April 6th, 2011 | 8:50 am

    Hi Netkas,

    For the ATI4850 card in 10.6.7 did you use the original drivers or your QE/CI patch ?

  4. jeanlain
    April 6th, 2011 | 9:32 am

    I believe that at that limited resolution, you’re partially CPU bound. I’ve a 2.66 Xeon Mac Pro (C2D equivalent), and with a 4870 running at max settings (1680*1050), almost all the demo is CPU bound.

    GLview (single cube) isn’t CPU bound.

    ///you are very very wrong. its cpu bound when u can see 100% load on one of cpus cores made by an app, exactly case of glview single cube.

  5. joe75
    April 6th, 2011 | 12:59 pm

    Does anyone who knows graphics really care about benches in osx? 😛

  6. fc_candido
    April 6th, 2011 | 5:48 pm

    My 9800GTX on LION DP2 1680×1050 work are 170FPS
    http://www.macupdate.com/app/mac/34981/smallopengl-fps-bench/

  7. besweeet
    April 6th, 2011 | 7:55 pm

    Can you post an exact screenshot of your video settings for CS:S? And can you explain how your demo works? I want to compare this with my Windows machine, but I can’t, since I don’t know *exactly* what your test settings and map are.

  8. netkas
    April 7th, 2011 | 10:35 am

    besweeet, i said, settings maxed, except 4xMSAA.

    unpack and copy demo to cstrike folder, then run timedemo 1

  9. jeanlain
    April 7th, 2011 | 5:16 pm

    Netkas, I know what I’m talking about, I did the tests.

    On my Mac, the single GLView cube uses 30% of my CPU at 1600*1000 windowed mode (to check CPU usage), with 8x AF and 4x AA.
    On the same Mac, same resolution and AF/AA settings; the CSS demo uses more than 100%, mostly 120-140%, which apparently is the best the Source engine can do.

    You might be seeing something different on a faster CPU, but the fact remains that the GLView cube is less CPU bound than the CSS timedemo, at similar settings.

  10. jeanlain
    April 7th, 2011 | 5:46 pm

    Something’s fishy with that demo.
    I get 100 fps, and my Mac isn’t that much slower than your hack. It’s a 2.66GHz quad woodcerst with 8GB ram, pretty much the equivalent of a 2.6GHz C2Q, with perhaps a bit slower memory (due to the FB-DIMMs).
    I’ve check that the demo is CPU bound, and in fact I still get 100 fps at 800*600. That’s not even half your best result. My gfx card is irrelevant here, since my CPU is limiting.

    I know that my Mac has normal performances in Source games, as I have been regularly comparing results with others, namely a Valve developer who has the exact same config as mine. Plus, I ran this timedemo many times before (you posted it on Macrumors) and always got around 100 fps.

    Are you using specific cvars to boost your fps? If not, I suspect something wrong on your side, perhaps some shaders were disabled somehow. I sometime see surprisingly high results when running timedemos because of this.

    To be sure, are you using vanilla 10.6.7 or do your graphic drivers come from 2011 MacBook Pros?

  11. Andy
    April 7th, 2011 | 6:48 pm

    VSync perhaps?

  12. netkas
    April 7th, 2011 | 7:53 pm

    jeanlain, I meant vanilla glview, since u didn’t mention 4xaa and 8xaf.

    And, if css uses more than 100% cpu then its multithreaded and you can’t tell for sure if its cpu bound, since you can’t know cpu load of rendering thread.

    if it was cpu bound then I would be getting nearly identical results with same drivers (so same results on and cards and another results for nvidia cards)

  13. fdeldongo
    April 8th, 2011 | 6:07 am

    Did the timedemo (on 1920 x 1080 though) : 151 fps
    (Q9650@3.4 and ATI 5850)

  14. Erik Lindahl
    April 8th, 2011 | 11:15 am

    What would be interesting is also relative benchmarks to Windows – i.e are some boards more or less bound by drivers than others in OSX. Some games have this issue in OSX (StarCraft for example).

  15. jeanlain
    April 8th, 2011 | 6:34 pm

    Even if CSS is multithreaded, the same performance at much lower screen res tells me that it’s CPU bound on my Mac. But your results suggest it isn’t CPU bound on your hack, since you constantly get higher FPS on faster cards.
    Your hack has a slightly faster CPU yes, but again, it should not be that different from my Mac. Something I can’t explain.

    fdelongo’s results is more in line with what I expect.

    For those interested in a comparison with windows. Well OS X is considerably slower in CSS, mostly because the game isn’t truly multithreaded on OS X (it just uses the multithreaded openGL feature of OS X), while it’s heavily multithreaded on Windows. Better drivers also make a difference.
    http://www.macbenchmarks.com/css_high.html

  16. netkas
    April 8th, 2011 | 10:41 pm

    I just checked with system monitor and didn’t see its cpu bound, check it:

  17. fdeldongo
    April 9th, 2011 | 7:45 pm

    I tested CSS on Win 7 x64 : 160 fps without multi-core (in the ballpark of the Mac performance), and 230 fps with it.

  18. jeanlain
    April 9th, 2011 | 10:10 pm

    Hard do say whether it’s CPU bound based on the graphs, as the scheduler constantly switches the process between CPUs.
    In my experience, Source uses at most 140-150% of a CPU, even at lowest resolution. I believe 40-50% is used by the openGL driver thread (via Apple’s Multithreaded openGL introduced in tiger 10.4.7) and 100% used by the main engine thread. This thread would be CPU bound.
    If you disable the use of OS X mt GL via a launch option in Steam, HL2_0S X stalls at 100-102%, losing about 40% of its performance.
    OTOH, you can enhance the multicore support with the console command mat queue_mode 2. This is equivalent to the Windows multicore rendering, but much less efficient.

  19. fdeldongo
    April 10th, 2011 | 9:41 am

    With mat_queue_mode “2”, I have 167 fps on the Mac side (1920 * 1080, Netkas settings).
    It’s an improvement, but not as dramatic as on Win side.

    So Mac graphics drivers seem good, and on par with Win, it’s css multi-core handling on the Mac side that could be improved

  20. Hackinsack
    April 12th, 2011 | 5:44 pm

    Quick question:

    I have an MSI 8800GTS 640MB and this driver won’t work for this line. Can someone let me know why that would be if the GT is? Pci 1.0 vs 2.0 cards? I’m on 10.6.7

    TIA

Leave a reply