GTX 4xx Fermi Cards & 10.6.8 - QE, OpenGL/CL & CUDA Benches

[Thireus]

Re: GTX 480 & 10.6.8 - QE, OpenGL/CL & CUDA Benches

Thanks for a wonderful sum up Thi, that was really a job well done...

Concerning the OpenGL issue (which I believe the most important) ... I started a Linux in dual boot and i have to say that it really works like a charm. OpenGL performance with Nvidia Linux drivers is WAY BETTER than in Mac OS. If Lion wont be able to do the job as people say (and I dont mean just getting some performance, I mean the same amount as Linux) I feel it might come to me abandoning Mac OS for Fedora again.

So just to sum up, if you dont want to wait or complicate your life now just go and download any Linux distro and enjoy much better Fermi performance

Hrm...I still don't understand the whole "OpenGL doesn't work well on Fermi cards in OS X" thing...I run Starcraft II on this machine and get BETTER performance under OS X than I do under Windows (with all details cranked to max, which would supposedly be GPU-limited as well as everything turned all the way down, which would supposedly be CPU-limited). Same deal with my other current addiction, X3: Terran Conflict. Am I missing something?

(Side note: I don't put much stock in synthetic benchmarking - it's way too easy to tweak benchmarking code for various OS eccentricities or optimize it for a particular architecture/OS combination, and often doesn't even reflect the performance of the component supposedly being benchmarked.)

Ok, I agree, you get some good results ... even with the tests (I ran Luxball Ultra test with result 50.5 secs on GTX 480 - you ran the same test in 17s.... there is something wrong there).

Can you tell me what did you actually do after you installed the system? It just might be your dsdt settings that did the trick (I used multibeast and a custom board for a Core2Duo build). But it might be something else... I just want to make sure what is it that causes such a big difference here...

Cause clearly your performance is way better than anybody's here

CPU makes the difference.

OpenCL is not well supported under MacOS 10.6.x for Fermi cards, so the OS uses CPU to compute. If you have a powerful CPU then results will be better.

 

[koszta15]

Re: GTX 480 & 10.6.8 - QE, OpenGL/CL & CUDA Benches

Aaaaaaa... there we go... Thanks for clearing that up Thi

 

[Thireus]

Re: GTX 480 & 10.6.8 - QE, OpenGL/CL & CUDA Benches

Do you have any actual evidence of this claim? I'm not saying you're wrong, just that other than you saying it's true, I haven't seen that result in reality. It's not like a have a monster of a CPU, it's an older non-SB i5 running at 100% stock speeds. It also doesn't explain why I get better performance under OS X than I do under Win7.

Sorry, this is all about OpenCL and not OpenGL.

Smalllux benchmark OpenCL.
OpenGL is fine, but OpenGL 3.0 is not supported entirely yet, gonna edit my posts.

 

[Thireus]

Re: GTX 480 & 10.6.8 - QE, OpenGL/CL & CUDA Benches

Yes alright, my mistake, OpenGL is good, but does not support 3.0 yet.

OpenCL is NOT good!

 

[koszta15]

Re: GTX 480 & 10.6.8 - QE, OpenGL/CL & CUDA Benches

Do you have any actual evidence of this claim? I'm not saying you're wrong, just that other than you saying it's true, I haven't seen that result in reality. It's not like a have a monster of a CPU, it's an older non-SB i5 running at 100% stock speeds. It also doesn't explain why I get better performance under OS X than I do under Win7.

Sorry, this is all about OpenCL and not OpenGL.

Smalllux benchmark OpenCL.
OpenGL is fine, but OpenGL 3.0 is not supported entirely yet, gonna edit my posts.

Well to me OpenGL is still crap on GTX480 no matter what.. I will try to do some fixes day after tmrw... do you guys think that my relatively (well really) poor performance could be caused by the latest Nvdia drivers? I have the latest and I tried installing a Windows game under Wine (playonmac) and running it. Oblivion (which was the game ) ran SOOOOO SLOW (25-30fps) concering that I have a high-end card

 

[Thireus]

Re: GTX 480 & 10.6.8 - QE, OpenGL/CL & CUDA Benches

So basically, Apple fixed OpenCL for Fermi cards with their "Software Update 1.4" for 10.6.7 (Have a look here: http://netkas.org/?p=764). But when we upgraded to 10.6.8, we lost support for Fermi cards and had to install NVidia Drivers, and with the latest NVIDIA drivers NVIDIA did not ported support for OpenCL maybe because they think OpenCL sucks and CUDA rocks... So, now we have to wait for a new Apple update to erase the latests kexts of NVIDIA and bring back OpenCL OR wait for Lion ^^

The fight:

Apple guys want OpenCL, so they don't care about CUDA.
Nvidia guys want CUDA, so they don't care about OpenCL.

So both are providing kexts barely working xD.

Actually, under 10.6.8 Nvidia won, because Apple kexts doesn't work for Fermi cards, so we do not have FULL support for OpenCL.

 

[koszta15]

Re: GTX 480 & 10.6.8 - QE, OpenGL/CL & CUDA Benches

I'd definitely test native games and OpenGL apps before deciding one way or the other. Wine introduces a middleware layer that interprets the DX system calls and translates them to OpenGL calls, at a decent performance penalty. Crappily ported games (mostly using Transgaming's Cider tech) suffer the same issues. In fact, one of my present addictions, X3: Terran Conflict, was originally a Cider port - the performance was so utterly terrible that the company responsible apologized publicly and released a patch that made it a native OpenGL app rather than a Cider-wrapped port =)

That being said, I had nothing but trouble with the newest Nvidia drivers. That's one of the reasons I mentioned that I use the previous revision of them =)

Well I know that it is a crap and damage on performance but STILL come on, it is GTX480 we are talking about. Wine should run fine... even with penalty it should be WAY faster... I am away from my station on a business trip for now but will be back home to try games in Wine on Linux and see the difference

 

[Thireus]

Re: Fermi Cards & 10.6.8 - QE, OpenGL/CL & CUDA Benches

Please correct me if I'm wrong:

The little explanation:

Apple and NVidia are fighting for OpenCL/CUDA.

Apple by default barely support Fermi cards. These cards are not their priority! Also they do not care about CUDA.

Nvidia wants to use the CUDA technology on the market, so Nvidia provides support for CUDA/OpenGL with their BETA drivers for OS X. They do NOT want to bring OpenCL support for their card. The kexts they provide into their installer overwrite Apple kexts for NVidia cards and kill OpenCL support.

So as you can see, our game is to bring both OpenCL and CUDA support for Fermi cards under OSX.

Under 10.6.7, Apple released an update with kexts to support Fermi cards. So you can have OpenCL working with Apple kexts AND CUDA working with NVidia drivers. Thus, you have full support for OpenCL/CUDA/OpenGL/QE under 10.6.7, read more here: http://netkas.org/?p=764. Basically, what you have to do is to install NVidia drivers for CUDA/OpenGL, and then erase some specified kexts with the Apple Update.

Note: NVidia Driver Installers do not only install kexts, but also Frameworks Frameworks are the key for OpenGL/OpenCL/CUDA. Kexts only tell the kernel how to use the video card when these Frameworks are used by apps. That's the reason why you need Frameworks from NVidia (with CUDA/OpenCL/OpenGL) but kexts from Apple (because they unlock the OpenCL functionality for Fermi cards).

Under 10.6.8, Apple kexts do NOT support Fermi cards anymore , like I said, this is not their priority. So NVidia jumped into the arena and quickly provided working kexts. Actually if you want to have your Fermi card working under 10.6.8 you have to install all kexts and Frameworks provided by NVidia (into their BETA driver installers).
This time they won the game, and we do not have OpenCL working anymore. We either have to wait for Apple to release updated kexts to bring back OpenCL support, or wait for Lion.

Under 10.7, Apple did a big effort and brought OpenCL support for Fermi cards

 

[Thireus]

PCIRootUID=1 never worked for me. PCIRootUID=0 works for me (but not needed in Kernel flags because it is set to this value by default).

EFI string not needed. Chameleon RC5 does the injection trick.

And of course kexts and other settings will be related to your own config.

 

[Thireus]

My goodness, so much stuff to remember. So what you're saying about Khan's install process, it won't take advantage of all the performance needed for the GTX 480 card? Let me know what you can.

Again, like I said before when I get to that point of upgrading everything next week. I will contact you Thi for your help in getting this thing going step by step. Later...

Yes it does ^^ but remember that custom kexts need to be related to your own configuration.

EFI string is for Chameleon <= RC4, to inject information about your card, so the OS will recognize it correctly. Chameleon RC5 automatically detects your card, and injects the EFI string, so you don't have to worry about the EFI string you are using and you can freely boot with different card without making any new changes in your com.apple.Boot.plist.

PCIRootUID has to be set to basically tell the kernel where the card is located, on which PCI port. So it won't boot if this value is incorrect.

Under 10.6.8, only CUDA/OpenGL works. OpenCL and graphic acceleration does not work properly, OpenCL is used in a lot of MacOS apps. Read my first post, I have updated it with new info