GTX480 in Pro with Q4000 driver and 10.6.5

[10THzMac]

So we now have new drivers for the Fermi, available from

http://www.nvidia.com/object/quadro-macosx-256.01.00f03-driver.html

and this installs a newer NVDAGF100Hal.kext dated October 11th that replaces the dodgy June one we have been playing with. CUDA 3.2 seems to have move from various release candidates to a proper release (November) and it is at:

http://developer.nvidia.com/object/cuda_3_2_downloads.html

So this lot was installed in a 2008 Pro with 10.6.5 installed a few days ago with a version of netkas' injector from some time ago. My 480 is powered via an external PSU supplying a 6-pin and 8-pin supply to the 480. The Mac boots of a std Mac version 285. The news is mostly good, but not perfect.

1. The 480 drives my ACD23 again with no problem (10.6.5 had previously stopped it working).
2. CUDA works on the GTX 480 under Mac OS X (which is my main reason for using the card).
3. The OpenGL performance of the 480 is still weaker than it should be under OS X. It is now slightly below the 285 in most tests, rather then dire as it was with the June kext. Only 1.4 is better on the 480 (which will make it clear which shot is which).

With regard to 3, I have no way of knowing whether this is because

(a) the Fermi driver is still immature
(b) the Q4000 driver is just not that well-matched to the older 480 Fermi chip
(c) the injector is compromising bandwidth somehow

One place where the 480 is MUCH better is double precision CUDA work. Despite being crippled compared to the Quadro and Tesla, it is still over twice as fast as a 285. In Single precision I have got over 700Gflops on the 480, but only on larger problems.

I attach some screen shots showing the still lacklustre OpenGL, but very good DP CUDA. Note that the OpenGL folk seem to have changed their test structure, so these numbers cannot be compared directly with some of the older ones on the forum. Also it looks like OpenGL 3 support is kinda half-there...

 

[Jocelyn84]

Weird! Here are my scores with a 480 and the same settings, though I'm running osX86.
I could be wrong, but you may have to edit your AppleGraphicsPowerManagement.kext/Contents/Info.plist depending on which Mac Pro version you have.
For example, I'm using Mac Pro 4,1 and most cards, even 2xx series, have been running at level 0 instead of 2 since 10.6.4. For a GTX 285 Vendor10deDevice05e2 needed to be changed to Vendor10deDevice05e3. For my 480, I changed Vendor10deDevice05e2 to Vendor10deDevice06c0. I'm not sure if this applies to a real Mac Pro, but figured I'd at least mention it. Oh and if you have a Mac Pro 3,1, this shouldn't be effecting your scores/results.

 

[J the Ninja]

Care to take SmallLuxGPU out for a spin with it?

@Jocelyn84.....weeeiirrddd....the only thing that comes to mind is bandwidth, but that shouldn't be a problem unless the OP put it in the x4 slot by accident.

 

[10THzMac]

Wow - those are much higher.

I have a Pro 3,1 (aka 2008) and had thought that was immune to the power management problem. The 480 is above the Mac 285, where I had previously put a second 2G PC 285. By the way, I put that back this afternoon and OpenGL viewer suggests the 2G bug might be gone (shot), which would make sense for the Q4000.

Meanwhile, Mathematica 8 talks to both cards (see shot) under CUDA. OpenCL is a bit iffy - the 480 is not seen and the GPU does not support double precision.

 

[SQUALL8765]

Questions:

1) What 480 card is being used?
2) Does it matter?
3) How do you install all this? Step-by-step?
4) Does this work in any Mac Pro? (ie. 2006 Mac Pro in particular)
5) Any abnormal fan/heat issues to worry about?
6) All ports working?

 

[voyagerd]

Jocelyn84, I followed your suggestion to edit that Info.plist and it really helped my performance. I'm still not getting as high of FPS as you are though.

 

[voyagerd]

Care to take SmallLuxGPU out for a spin with it?

I downloaded SmallLuxGPU, what do you want to know?

 

[10THzMac]

So I made a few checks:

1. Both cards are in the x16 slots, with the Mac edition in the lower one and the 480 in the upper one.
2. I tried swapping the slots and the Mac would not boot.
3. I am in a 2008 (3,1) Mac Pro. I had a look at the power mgmt kext and there is still no entry for 3,1, just 4,5 and 5,1, so I am guessing this is not the problem.
4. I am wondering if with the OpenGL benchmark there is a part of the one I posted that is dependent on the CPU speed and not really testing the GPU in isolation. After all, my 2008 Pro has the Harpertown processors which are rather left behind by Nehalem, Core i7 etc.

Thanks for the two sets of Hackintosh numbers - they are both interesting and annoying! Anyone with a Mac Pro got some numbers?

Questions:

1) What 480 card is being used?
2) Does it matter?
3) How do you install all this? Step-by-step?
4) Does this work in any Mac Pro? (ie. 2006 Mac Pro in particular)
5) Any abnormal fan/heat issues to worry about?
6) All ports working?

1. I have a PNY 1.5G 480 card - one of the 1st ones
2. It can do, based on experience with 2xx series, but I cannot comment with only one card.
3. (i) 10.6.5, Nvidia drivers, as in my post
(ii) find ATY_Init.kext and install (netkas.org)
(iii) power the 480 with a 6- and 8-pin PCI power somehow (I have external toughpower)
(iv) Boot off a proper Mac card in the lower slot

People may have other routes (which might be the reason they are getting better numbers...)

4. Cannot say for sure. Also note that when we got PC 285s working some people could not get 275 up.

5. I power up the 480 for a few hours when I am doing big CUDA problems, otherwise it is off. So I have had no problems but more cts usage might. Fan not doing funny antics at boot nor at other times.

6. Both DVI ports drive an ACD23 OK.

Look - given the cost of these cards this is really a project for those who want this card mainly for Bootcamp windows use and are happy to just try an experiment under OS X, and perhaps take the time to edit kexts etc. The hackintosh folk probably have an easier time of it once they have got through the driver issues.

 

[voyagerd]

The hackintosh folk probably have an easier time of it once they have got through the driver issues.

I'm just using the Nvidia Q4000 drivers with an unmodified 10.6.5 installation. I don't use any custom drivers at all. I wonder if your framerates have something to do with having to use that injector or the ATY_Init.kext.

 

[10THzMac]

The injector=ATI_init

Who knows - I am puzzled right now. The one thing we do know is that the link speed is usually 2.5 GT/s for these injected cards vs 5 for the Mac native ones. This is the number at the bottom the PCI cards entry in the System profiler. Maybe I have hit a bandwidth limit, not a CPU or GPU limit.

netkas, Rominator - if you are listening - any ideas?

Now I know the 2G bug looks like it is fixed I have put the 2G 285 back and will play with that for a bit. OpenCL recognizes that as well and Mathematica 8 is very happy seeing 2 CUDA and 3 OpenCL devs (2GPU + CPU).

 

[Jocelyn84]

The one thing we do know is that the link speed is usually 2.5 GT/s for these injected cards vs 5 for the Mac native ones.

Could this be the reason GPU Flame Fractal Renderer is reporting only 240 CUDA Cores instead of 480?

 

[10THzMac]

Could this be the reason GPU Flame Fractal Renderer is reporting only 240 CUDA Cores instead of 480?

I suspect not. CUDA deviceQuery reports 480, as does Math 8 (see shot above). This looks like a different problem.

Anyone with an Apple '09 or '10 Pro able to report scores? Over on netkas.org there are some other comments that 480s are not yet delivering full speed.

 

[voyagerd]

Has anybody tried two GTX 480s in SLI under Mac OS? I know that there is no SLI capability, but I wonder how Mac OS would see the cards. I tried booting with both my GTX 480 and my GTX 260 enabled and I got a kernel panic while booting into Mac OS.

 

[jeanlain]

Do NOT use the multiple cubes scene to measure GPU performance in GLView. This scene is heavily CPU-Bound. You are in fact measuring CPU performance.
Just render the single Cube/Face scene.

 

[voyagerd]

Do NOT use the multiple cubes scene to measure GPU performance in GLView. This scene is heavily CPU-Bound.
Just render the single Cube/Face scene.

I agree that the cubes are a really poor benchmark for a GPU. We should probably be comparing CINEBENCH scores.

 

[berkut]

If you install the CUDA drivers, that installs NVIDIA's version as opposed to Apple's, which do support OpenGL 3.0 fully (GLSL 1.3).

 

[Jocelyn84]

Has anybody tried two GTX 480s in SLI under Mac OS? I know that there is no SLI capability, but I wonder how Mac OS would see the cards. I tried booting with both my GTX 480 and my GTX 260 enabled and I got a kernel panic while booting into Mac OS.

Yes, you need to edit your DSDT or use a string in your com.apple.boot.plist
Method 1 Kdawg's DSDT method (Post 29) - http://www.insanelymac.com/forum/index.php?showtopic=183349&st=20
Obviously your info needs to be different than his for your card(s)

Method 2 - Aquamac's GFX string method - http://aquamac.proboards.com/index.cgi?board=hack1&action=display&thread=569
I've done both methods with two GTX 260 cards, so either of these should work for dual 4xx cards.

Do NOT use the multiple cubes scene to measure GPU performance in GLView. This scene is heavily CPU-Bound. You are in fact measuring CPU performance.
Just render the single Cube/Face scene.

Interesting. I'm only running @ 4.0Ghz and my score was still higher than voyagerd who is running @ 4.2Ghz. I'm sure it has something the EX58-UD5 DSDT probably being the most optimized X58 board on the planet, thanks to dOOd over at Insanely.

If you install the CUDA drivers, that installs NVIDIA's version as opposed to Apple's, which do support OpenGL 3.0 fully (GLSL 1.3).

I started using those before running GPU Flame Fractal Renderer. I installed Install Nvidia Cuda 3.1.17 then it semi-auto updated CUDA Toolkit 3.2, which also added a preference panel.


PS - I hope no one's getting upset with all this osX86 info. I'm only trying to update people on things. I own a MBP, but haven't posted on these forums in quite sometime. Anyway, I hope Mac Pro and osX86 people can help one another with this like we have in the past.

 

[jeanlain]

I agree that the cubes are a really poor benchmark for a GPU. We should probably be comparing CINEBENCH scores.

Cinebench ins't much better actually. Even the openGL benchmark is CPU bound on my Mac Pro (which admittedly has a slow CPU by today's standard). Just watch the CPU usage during a test. If it's capped at 100%, it's not good as a GPU benchmark.

^ Apple only supports openGL 2.1 fully.

 

[barefeats]

It would be interesting to see the results for SpecViewPerf 11 and redway3D benchmarks running under Windows 7.

Those tests and others were run on the Quadro 4000 by this site:
http://www.techeye.net/hardware/we-take-a-dekko-at-nvidias-fermi-quadro-4000

Also, run 3DMark Vantage:
http://www.futuremark.com/benchmarks/3dmarkvantage/introduction/

 

[sgunes]

As a comparison:
These are the numbers I am getting with my GTX480.

My Cinebench 11.5 numbers: 31.86 fps with 19.25 CPU score.

10.6.5 Cuda Drivers 3.2.17. Fermi Drivers 256.01.00f03.

 

[10THzMac]

sgunes - please would you tell us the host Mac or hackintosh spec - while we figure out how best to test the card separately from the CPU it helps to know the host chip and clock speed.