Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Frequently Asked Questions About NVIDIA PC (non-EFI) Graphics Cards
- Thread starter Asgorath
- WikiPost WikiPost
- Start date
- Sort by reaction score
- Status
- Status
This is what I was thinking too. I was aware of the PCIe 1.0 issue and I remember MacVidCards saying something about audio not going out over the mDP connection or something similar.
I am finding that Skyrim at 2560x1440 0xAA 8xAF is using all my vram and pushing the 5870 1Gb to about 85% load. I know the proposed Mac 7950 is a 3Gb card but I am concerned that the actual horsepower improvement over the 5870 does not really appear hugely significant so I was still looking at a 680 4Gb card (the 2x6-pin Twin Frozr one).
I am finding that Skyrim at 2560x1440 0xAA 8xAF is using all my vram and pushing the 5870 1Gb to about 85% load. I know the proposed Mac 7950 is a 3Gb card but I am concerned that the actual horsepower improvement over the 5870 does not really appear hugely significant so I was still looking at a 680 4Gb card (the 2x6-pin Twin Frozr one).
I benched LuxMark on the windows 7 partition and it scored half as much as the same test on the Mac version. It probably doesn't affect gaming much ,especially if they are CPU bound, but it does seem to halve performance at least for OpenCL.
Keep in mind nVidia's OS X drivers are far behind AMD's. At the moment nVidia doesn't even support hardware instancing.
Umm, this is just flat-out incorrect. I'm seeing accelerated instancing support with the latest NVIDIA drivers, what makes you think they are not supported? Last time I checked AMD didn't even support geometry shaders, while NVIDIA has for a long time.
It's been well documented by X-Plane's developers: nVidia's OS X drivers do not support hardware instancing. You can even see it if you look at the X-Plane log file it creates with each launch. After polling the GPU it says there's no hardware instancing. It does support software instancing, but in something like X-Plane, which has an object budget per frame of up to 25,000+ objects, software instancing dies quickly. You can also see the difference if you run X-Plane with the same nVidia GPU on Windows and OS X. On Windows you will see, at a minimum, double to frame rate.
You can see in posts like this that forcing instancing gets you lower frame rate.
edit: From the X-Plane log.txt file:
Code:
Disabling instancing for DX10 NV hw - it is software emulated.
Last edited:
Okay, this is something totally different. NVIDIA's 8-year-old DX10 architecture doesn't support hardware instancing, so both the Windows and Mac drivers emulate it in the driver (i.e. manually unrolling the instanced loops). The DX11 GPUs starting with the Fermi do have hardware instancing, from what I've seen. You make it sound like the NVIDIA driver for OSX is deficient, but as far as I can tell, it's behaving the same way as the Windows driver w.r.t. instancing support.
Now, X-Plane performance between OSX and Windows is a whole other ball game. My guess is that this has more to do with the different driver architectures between the two OSes than simply lack of HW instancing support (which again, is the same for both OSes -- it's not like Windows magically gets HW instancing support for NVIDIA's DX10 GPUs).
The X-Plane developers have found otherwise: instancing doesn't work with nVidia cards in OS X, and it's a driver issue. This doesn't surprise me. Historically ATi/AMD's OS X drivers have been much better than nVidia's, and it's only in the past year that nVidia has started to catch up. I remember a blog post from Ben Supnik, X-Plane's lead graphics developer, in which he was seeing 3x the frame rate on AMD versus nVidia on the same OS X hardware. It's still an issue, too. With the 10.8.3 drivers an AMD 6780 will outperform a 670 in some cases.
The good news is that nVidia knows about the instancing issue, which is why I'm hopeful it will be fixed in the next release of nVidia's web drivers. It's frustrating to fire up X-Plane with a 4GB 670 SC and watch my frame rate drop to 15 in object-dense areas.
edit: And DX10 most definitely supports instancing.
As I can say, my MP has 15-20% performance loss compared to modern PC hardware with PCI-E 2.0 or 3.0 and same GTX 690 video card.
etc - I wonder if much of that is down to faster memory, i/o or cpu rather than the PCI speeds on your other computer?
http://www.anandtech.com/show/5458/the-radeon-hd-7970-reprise-pcie-bandwidth-overclocking-and-msaa
----------
but you have one of the very few cards that really needs pcie 3.0
for a 7850 or less, probably not a big deal
All I can offer is a benchmark comparison from the new Tomb Raider:
Original ATI 5870: 1920x1080, Ultra Settings (except Tress FX Hair): 23 fps
GTX 670 I bought last Friday: 2560x1440, Ultra Settings (except Tress FX Hair): 31 fps
That exactly what Anandtech did NOT say. In their performance charts there was little difference (<5%) between PCIe 3.0 x2 (2GB/sec) vs. PCIe 3.0 x16 (16GB/sec), in most gaming at least.
To quote them: "The good news is that even at 2GB/sec the bottlenecking is rather limited, and based on our selection of benchmarks it looks like a handful of games will be bottlenecked."
A PCIe ver. 1.0 x16 is running at 4GB/sec, which is equal to the PCIe ver. 2.0 x8, or the PCIe ver. 3.0 x4.
Of course the more demanding the graphics card is, the more the effect will be notised. For OSX though, in the MP4,1 and later, it seems you have PCIe 2.0, and the graphics cards are x16 cards, so 8GB/sec.
We with MP3,1s will hopefully get ver. 2.0 when NVIDEA updates their reference drivers...
In Bootcamp it seems we are stuck
Weird Specs in Luxmark after 10.8.3 update
On 10.8.2 with NVIDIA driver specs for my GTX 570 gave me
Compute Units: 60
The same card with 10.8.3 Apple driver gives me
Compute Units: 15
Damn Apple give me back those 45 compute unit you have stolen with the update. ;-)
On 10.8.2 with NVIDIA driver specs for my GTX 570 gave me
Compute Units: 60
The same card with 10.8.3 Apple driver gives me
Compute Units: 15
Damn Apple give me back those 45 compute unit you have stolen with the update. ;-)
Some sort of cosmetic glitch, I have noticed this as well but actual CUDA & OpenCl are only down 2-5% while Core Count down 75%
You have to be putting an enormous amount of info through the bus to saturate PCI. As most games 1) use a lot of tricks to reduce the real size of their worlds while making them look big and 2) load all of the textures at level load, the odds of saturating the PCI bus while gaming are minimal.
X-Plane is one of the few games which has the potential to saturate the PCI bus, and that's only at 1) the highest graphics settings and 2) assuming you're not wither CPU- or GPU-bound. Even with instancing fixed I think I would still be CPU-bound with a 3.33 GHz six-core machine, as there is a limit to parallelization given X-Plane's logical constraints.
PCIE 1.0 x 16 lanes = 2.5 Gt/s
PCIE 2.0 x 16 lanes = 5.0+ Gt/s
These numbers can be confirmed by CUDAZ or the newer OpenCl waves bench
PCIE 2.0 x 16 lanes = 5.0+ Gt/s
These numbers can be confirmed by CUDAZ or the newer OpenCl waves bench
- Status