DAVINCI RESOLVE and ***HACKINTOSH***

[Peterson Mark]

How's the stability with the new Cuda 4.0?
And did it fix the problem that davinci only uses half cuda cores on Macos?

[luigivaltulini]

CIAO Peterson,
Half of cuda? all core cuda work in Macos, i hope not use Flame4cuda for test this, if you see my post 104 Cuda core is 480 for gtx 480.
If you have quadro 4000 you have 256 core , if you have gtx you have from 400 to 480 cuda core depended your gtx .In my case i use gtx 470 and i have 448 core x 4 video cards for total of 1792 core . Cuda driver 4.0 Rc2 improve stability and performance.

thanks
Luigi

[Dave P]

I think the "half the cores" rumor is either from seeing the Flam render window which did show like 240 cores for GTX480 cards or from the issue I have become aware of.

It turns out that many Hacks only run Nvidia cards at PCIE 1.0 spec instead of 2.0

This results in literally 1/2 bandwidth for transfers in and out of VRAM to CPU.

You can check your status by looking in System Profiler under "PCI Cards"

If you are running at PCIE 1 spec it will show "2.5 GT/s" and if you are running at PCIE 2.0 spec, it will show "5.0 GT/s"

You can run CUDA-Z to check your actual bandwidth as sometimes System Profiler will report incorrect number.

Download CUDA-Z and run the performance section. Pretty easy to spot the difference and the numbers correspond to the expected values.

I am attaching the bottom part of exported report. These were generated on Mac Pro using GTX480.

Apparently some means of injecting or enabling Nvidia cards will result in the lower setting.

[luigivaltulini]

If you are running at PCIE 1 spec it will show "2.5 GT/s" and if you are running at PCIE 2.0 spec, it will show "5.0 GT/s"

Ciao Dave,
some parameters can be changed in HACKINTOSH dsdt table for the system profile.

As far as I know PCI-E 1.0 x 16 is equivalent to PCI-E 2.0 x 8 only difference if your card run at pcie 2.0 16x
for speed to transfers in and out of VRAM to CPU is double , but the number of core is a problem of flame4cuda, if
you have 480 core is 480 not half , only for change speed of bandwith depend of pcie 2.0 16x or 1.0 16x . But the gtx no use total of 16x
pcie 2.0 and in this case for gpu computing 16x pcie 1.0 is perfect.

Excellent post Dave helps a lot to understand, thanks.

Luigi

[Jose Lomeña]

Ciao Dave,
some parameters can be changed in HACKINTOSH dsdt table for the system profile.

As far as I know PCI-E 1.0 x 16 is equivalent to PCI-E 2.0 x 8 only difference if your card run at pcie 2.0 16x
for speed to transfers in and out of VRAM to CPU is double , but the number of core is a problem of flame4cuda, if
you have 480 core is 480 not half , only for change speed of bandwith depend of pcie 2.0 16x or 1.0 16x . But the gtx no use total of 16x
pcie 2.0 and in this case for gpu computing 16x pcie 1.0 is perfect.

Excellent post Dave helps a lot to understand, thanks.

Luigi

Luigi are you sure that we are using 480 cuda cores?, I think we are using the half, because the performance of a gtx480 vs gtx285 is very similar, maybe flam4cuda are right...

Do you saw any improvement with cuda4.0, i have installed since a month and the performance is exactly the same...

[Dave P]

I am going to try to be careful about my wording as I am not sure if I can disclose everything.

I have been in contact with someone at Blackmagic who tried using a GTX460 and GTX470 in both a Mac and a Hack. I provided the cards for testing.

His findings where that these Fermi cards were unable to show a clear speed difference from GTX285. They were no faster in tests done at their facility.

I have been trying to figure out why superior cards showed no speed increase.

The cards were running at 2.5 GT/s. This is not important for most graphic rendering tasks, but GPU computing is all about transferring info from RAM into CPU and back.

PCIE 2.0 is important for this.

For some info:

http://code.cheesydesign.com/?p=113

Luigi's numbers indicate that he has 5.0 running..despite what system profiler says. His numbers for Pinned memory are in line with what I get for GTX470 in Mac Pro.

His numbers for Pageable are very high, I am not sure how they can be that high. He has a monster machine. But I can only report directly on what I get in a 4,1 Mac Pro.

I am about 90% certain that this is why the un-named gentleman at Blackmagic found these cards no faster.

When they return from NAB, I will provide cards that now run at 5.0 GT/s and I am hopeful that they will prove to finally "stretch their legs"

Oddly, Luigi's System Profiler indicates 2.5 but his perf figures show 5.0 (or better).

Another way to see which mode you are running in is to use IORegistry Explorer

Navigate to the card and look at "IOPCIExpressLinkStatus" if it reads 0x102 you have Full 2.0 Spec, if it is at ox101 you are stuck in 1.0 mode.

I am working with Netkas on way to enable in Hacks as well. We found a way together to do this in BIOS but I am hopeful that he can add it to an init someday.

I posted at Insanelymac to see what others were getting. Only got a couple responses, one guy at 2.5, another clearly at 5.0. Netkas hack came in at 2.5 by default. It would be good to figure out why some get 1.0 and others 2.0

Meanwhile, I will post here if and when I get Blackmagic guy to test cards again and let you know what they find.

EDIT: Ultimately, I think the CUDA-Z benchmark showing speeds above 5xxx in "pinned" is what matters. Speeds in the 2xxx-3xxx range in the "Pinned" memory section indicate 2.5 GT/s speed. Luigi has MUCH better scores in Pageable because he has a much faster system...i7 I imagine. They have faster RAM access.

[luigivaltulini]

Ciao JOSE,
yes i am sure my tools of nvidia see me 448 core and gtx 470 have this core, and for your question i dont no test why no have a gtx 285.
But i have teste 9800 gt and 460 and 8800 and flame4cuda reponse half of cuda and tools of nvida repose full cuda. A this point i hope flame4cuda is be wrong.

Ciao DAVE,
certainly be strange that the 2.5 profile in the system and I have ioreg 0x101 ((in my case i have 0x1101)) but I think they are
errors, probably the best solution is CPU-Z and control speed.
I am convinced that the cards go to a few PCIe 2.0 16x and other at 8x.
In my case I have 2 video cards in the blue slot at 16x, and two video cards in the slots to 8x blacks. In fact, two are above 5000 and 2
are above 3000.
I remain very interested in your idea Dave and waiting for your reply or edit.
I hope DAVINCI RESOLVE (blackmagic design) help the fermi card especially gtx


Many thanks
Luigi

[Paul Huynh]

finished building an air cooled asus p6t7 supercomputer hackintosh.

THANK YOU LUIGI and Everyone else. Couldn't have done it with you guys!

p.s. ill post pics later this week. been slammed at work

p.s.s.

anyone over clock their core i7 950? if so share your settings?



thanks again!

p-

[David Kellermann]

There's a pretty good how-to guide, filled with examples. Unfortunately it's in German, but google does a good job on translating it:

http://translate.google.com/translat...ing.html%23a02

[Paul Huynh]

thank you!