Reduced 7xxx / R9 Performance in OS X on 2008 Mac Pro 3,1

[jasonsansone]

I am creating this thread to better consolidate this topic from various other posts. I have probably mucked up some other threads and gone off topic, so I feel it better to try and stay on point at least somewhere.

In short, I am experiencing substantially lower performance in OS X than Windows 7 when using 7xxx cards.  The issue does not appear to be limited to the 7950 as I have also tested a 270X.  The problem seems to only plague the MP 3.1 as others are experiencing much better results on other model Mac Pros.  I have seen links to results on 2.1 and 1.1 models that surpass my own.  Blacksheep and others have demonstrated the 7xxx cards are capable of coming close to, if not matching, Windows performance in OS X.  The argument has been made that the MP3.1 is hardware limited when compared to the 4.1 or 5.1, but hardware limitations should result in similar limits across any OS.  It also wouldn't explain how even older models from 2006 or 2007 can surpass the 2008.  I have attached various screen grabs showing the disparity.

I have performed a completely fresh install of OS X, removed APGM (replaced after no change seen), and tested 10.9.2 beta 1.  In all I have used 10.8.4, all 10.8.5 betas and final, every DP of Mavericks, 10.9.0, 10.9.1 betas and final, and 10.9.2 beta 1.  I have tried both Sapphire 7950 Mac Edition and an XFX 270X.  I have used stock bios, OC'd bios, EFI and non-EFI bios... the situation stays the same.  The 7950 Mac performed this way out of the box in EFI and Windows settings.  I don't believe it is defective cards as this now effects three different GPU's from different series lines.  The diminishment is nearly the same percent between the 270X and 7950's.  I don't know if the cards are throttling in OS X, but they aren't in Windows, nor should they be.  LuxMark results remain constant from start to end of testing, so there doesn't appear to be any degradation after loading.  The TDP limits on the 7950's have been raised and I have a secondary 750W PSU pushing the GPU's PCIe 4 x 6-pin power inputs.  The system is rock solid stable in Win7 under various torture tests and the GPU's never throttle down.  I own two different Mac Pro's from 2008 and have tested all three cards in both machines.  Results are consistent.  The problem does not appear to be isolated to one framework like OpenCL or OpenGL alone.

LuxMark Sala Scene (OpenCL):
XFX 270X: Stock OS X - 870; Stock Win 7 x64 - 1606.  54.2% OS X output vs Win7.
2 x 7950 OC: OS X - 2857; Win 7 x64 - 4732.  60.4% OS X output vs Win7.

Heaven Basic Preset Benchmark (OpenGL):
7950 OC: OS X - 62.8 fps, 1582 score; Win 7 x64 - 103 fps, 2595 score.  60.1% OS X output vs Win7.

Heaven Extreme Preset Benchmark (OpenGL):
7950 OC: OS X - 22.8 fps, 575 score; Win 7 x64 - 46 fps, 1159 score.  49.6% OS X output vs Win7.
Stock 7950 tested by Barefeats: 38.7 fps - http://www.barefeats.com/gpu7950.html.  58.9% vs Barefeats.

Valley Basic Preset Benchmark (OpenGL):
7950 OC: OS X - 41.3 fps, 1728 score; Win 7 x64 - 32 bit only

Valley Extreme Preset Benchmark (OpenGL):
7950 OC: OS X - 29 fps, 1213 score; Win 7 x64 - 32 bit only
Stock 7950 tested by Barefeats: 33.1 fps - http://www.barefeats.com/tube04.html

Valley Extreme HD Preset Benchmark (OpenGL):
7950 OC: OS X - 22.8 fps, 952 score; Win 7 x64 - 32 bit only

OceanWave (OpenCL):
7950 OC: OS X- 506.8 fps; Barefeats - 548 fps - http://www.barefeats.com/gpu7950.html
5770: OS X- 146.3 fps; Barefeats - 138 fps
270X: OS X- 353.9 fps; Barefeats - Not Tested

Luxmark Sala Scene with single stock 7950 Mac Edition on Mac Pro 3,1:
10.8.3 - 1449
10.8.5 - 1442
10.9.2 - 1242
10.9.3 13D12 - 1287
Win7 - 1792

[Heco]

My tests show similar results.

I did an OceanWave with 10.9.1 that scored around 61 FPS with a VRAM speed of 70 GB/s.
(With 10.8.2 and original Sapphire drivers the 15 second test would crash, but it showed a VRAM speed of 110 GB/s.)

Barefeats did the OceanWave test on a 2010 Mac Pro with a result of > 550 FPS. In comparison the Mac Pro 2008 is not only 'slower' - it seems the 7950 is not addressed properly ...

[jasonsansone]

I got 61.3 fps, almost the exact same result.  

EDIT: V-SYNC was enabled.

[jimj]

FYI-

There is a sync issue with the oceanwave test - it is no coincidence that all results are hovering around 60 Hz...

[bx5a]

I was following your data in other threads; hopefully the condensed information here brings us closer to a solution. From what you've said, it seems like the issue lies within the OS X drivers but somehow in a way that makes it an issue and results in performance degradation only for the MP 3,1.

What really surprised me was the part about the MP 1,1 and 2,1 performance. If that's true, then the issue isn't even in the OS X drivers in relation to the older architecture but solely in relation to the 3,1. If anything, this huge performance disparity explains why Sapphire decided not to support the card in 3,1 officially.

But the cause for the reduced performance appears really arcane. Even more so when one takes into the non-problematic Windows results. Hopefully someone here will figure it out.

[Heco]

There is a sync issue with the oceanwave test - it is no coincidence that all results are hovering around 60 Hz...

That's right ... switching the display to 75Hz gives me a score of 75. But what did Barefeat test?

[jasonsansone]

There is a sync issue with the oceanwave test - it is no coincidence that all results are hovering around 60 Hz...

That's right ... switching the display to 75Hz gives me a score of 75. But what did Barefeat test?

Looks like they ran it under ML.  Its a Mavericks issue with OpenWave.  So we solved one (tiny) problem.

[blacksheep]

Regarding Ocean Wave and 10.9, it's v-sync issue: DL Quartz Debug and tick Beam Sync "Disable" in settings (don't quit QD) and run test. You'll get real scores this way.

[jasonsansone]

Regarding Ocean Wave and 10.9, it's v-sync issue: DL Quartz Debug and tick Beam Sync "Disable" in settings (don't quit QD) and run test. You'll get real scores this way.

7950 OC - 506.8 fps
5770 - 146.3 fps
270X - 353.9 fps

Keep in my I am running 1.25V, 1200mhz core, 1500mhz memory.  My core speed is 50% higher than stock and still comes in under Barefeats tests.

[Fangio]

Regarding Ocean Wave and 10.9, it's v-sync issue: DL Quartz Debug and tick Beam Sync "Disable" in settings (don't quit QD) and run test. You'll get real scores this way.

7950 OC - 107.9 fps
5770 - 347.4 fps

Hmm? How do you do that with a 5770?

Edit: ah, edited.

[attachment deleted by admin]

[jasonsansone]

Regarding Ocean Wave and 10.9, it's v-sync issue: DL Quartz Debug and tick Beam Sync "Disable" in settings (don't quit QD) and run test. You'll get real scores this way.

7950 OC - 107.9 fps
5770 - 347.4 fps

Hmm? How do you do that with a 5770?

Edit: ah, edited.

It was reporting incorrectly.  It reported the combined power of both GPU's.  I didn't realize it at first.  Sorry.  Final tests are posted above.

[Heco]

Looks like they ran it under ML.  Its a Mavericks issue with OpenWave.  So we solved one (tiny) problem.

I tried running OceanWave with ML. It reported 'no OpenCL device' or sth like that and wouldn't run.
Tried with 10.8.2 and Sapphire drivers and with 10.8.5 and Apple drivers. Same error on the 2008 Mac Pro.

So it seems that the drivers are working under ML with a Mac Pro 2010.

[Heco]

Regarding Ocean Wave and 10.9, it's v-sync issue: DL Quartz Debug and tick Beam Sync "Disable" in settings (don't quit QD) and run test. You'll get real scores this way.

That's it! My results with stock speed HD 7950: 492.7 fps

[jasonsansone]

If anyone can figure out how to flash a 7xxx or R9 to a Dxxx, I would be curious to see if the D300/D500/D700 drivers perform differently. I tried but failed.  As Apple has never released or sold a 7xxx but now sell the Dxxx, it stands to reason they will spend much more effort optimizing those drivers.  Unlike prior official graphics card releases, the 7950 is not backed by Apple but by Sapphire, so we may not see a lot of support.

[Heco]

Test completed ...
(Mac Pro 3,1, HD 7950 Mac Edition (Stock), OS X 10.9.1)

LuxMark Sala Scene (OpenCL):
7950: OS X - 1339

Heaven Basic Preset Benchmark (OpenGL):
7950: OS X - 59.4 fps, 1497 score

Heaven Extreme Preset Benchmark (OpenGL):
7950: OS X - 18.3 fps, 461 score
Stock 7950 tested by Barefeats: 38.7 fps - http://www.barefeats.com/gpu7950.html.  47.2% vs Barefeats.

Valley Basic Preset Benchmark (OpenGL):
7950: OS X - 39.4 fps, 1649 score

Valley Extreme Preset Benchmark (OpenGL):
7950: OS X - 24.3 fps, 1015 score
Stock 7950 tested by Barefeats: 33.1 fps - http://www.barefeats.com/tube04.html.  73.4% vs Barefeats.

Valley Extreme HD Preset Benchmark (OpenGL):
7950: OS X - 18.6 fps, 780 score