PCI Express 2.0 vs 3.0 GPU Gaming Performance Review @ [H]

This article must have taken a cubic shit ton of time to get both the data and make it all purty for us. I think it answers a lot of questions but also opens some others. Thanks again.
 
Quartz-1 said:
Thank you for taking the extra time to ensure you got things right.

This really begs the question though: do we need a faster Crossfire / SLI bridge technology, or, how can we make multi-GPU solutions faster?
Click to expand...

Change the whole way multi-GPU is currently done, though AFR is the most efficient way right now, it's entirely inefficient in the grand scheme of things, the framebuffers for one are not shared. You want to make multi-GPU better, figure out a way to combine both video cards framebuffer into one usable combined virtual space of memory and negate the need to copy the framebuffers.
 
As another user mentioned, a more accurate test would be to just get a UEFI that supports changing from PCIe 3.0 to 2.0 so the same CPU can be used. Also there are proven advantages for GPU Compute applications for PCIe v3.0 vs v2.0.
 
Brent_Justice said:
Change the whole way multi-GPU is currently done, though AFR is the most efficient way right now, it's entirely inefficient in the grand scheme of things, the framebuffers for one are not shared. You want to make multi-GPU better, figure out a way to combine both video cards framebuffer into one usable combined virtual space of memory and negate the need to copy the framebuffers.
Click to expand...

I remember back when Crossfire and SLi were resurgent and there was a lot of talk about the different multi-GPU rendering modes. AFR was the least favored because of things like micro-stuttering (then not clearly defined), and the obvious lack of synchronization with rendering engines. Split-frame rendering or alternate line rendering (a la 3Dfx Scan Line Interleave) makes more sense as all GPUs would be rendering the current frame, but average FPS numbers and game compatibility win, so AFR it is.

If multi-GPU wasn't still such a niche we might already have good solutions, and until we get them, we need to keep clamoring for support!

And we need to keep supporting guys like our team at the [H], as they are our representatives to the OEMs :).
 
WOW!

Kyle's a fucking asshole, and the performance difference between PCI-E modes is very little more than the ipc difference, and pretty much no net difference for AMD cards.

Why isn't Captain Rectal Headgear chiding Nividia about a clear CPU dependence. Pretty obvious to me that NV cards are leaning on CPU for something that ATI covers on card.

Go ahead and ban me. You're a biased asshole. Oh, and you act completely unprofessional, given as you have official capacity here.

Clean the green jizz of your chin.
 
Solomutt said:
WOW!

Kyle's a fucking asshole, and the performance difference between PCI-E modes is very little more than the ipc difference, and pretty much no net difference for AMD cards.

Why isn't Captain Rectal Headgear chiding Nividia about a clear CPU dependence. Pretty obvious to me that NV cards are leaning on CPU for something that ATI covers on card.

Go ahead and ban me. You're a biased asshole. Oh, and you act completely unprofessional, given as you have official capacity here.

Clean the green jizz of your chin.
Click to expand...
wtf is this rant about?
 
Just a thought, but does PCIe 3 affect level loading times? I'm guessing but maybe at the start of each level the game loads as many textures as possible into the video memory, and PCIe 3 might make that much faster.
 
Quartz-1 said:
Just a thought, but does PCIe 3 affect level loading times? I'm guessing but maybe at the start of each level the game loads as many textures as possible into the video memory, and PCIe 3 might make that much faster.
Click to expand...

didn't notice any differences
 
Quartz-1 said:
Just a thought, but does PCIe 3 affect level loading times? I'm guessing but maybe at the start of each level the game loads as many textures as possible into the video memory, and PCIe 3 might make that much faster.
Click to expand...

Brent_Justice said:
didn't notice any differences
Click to expand...

To add to that, consider that the game typically has to load stuff into main memory before piping it to the GPU(s). Main memory is considerably slower than any PCIe 16x interface.
 
Solomutt said:
WOW!

Kyle's a fucking asshole, and the performance difference between PCI-E modes is very little more than the ipc difference, and pretty much no net difference for AMD cards.

Why isn't Captain Rectal Headgear chiding Nividia about a clear CPU dependence. Pretty obvious to me that NV cards are leaning on CPU for something that ATI covers on card.

Go ahead and ban me. You're a biased asshole. Oh, and you act completely unprofessional, given as you have official capacity here.

Clean the green jizz of your chin.
Click to expand...

Without calling names, it's obvious you have an AMD bias here, and that is not appreciated, and certainly won't garner you any respect. That said, I think we all picked up that Nvidia cards are a little more CPU dependent than AMD cards while reading this article. Thing is, it doesn't matter- even if the Nvidia cards are a little more CPU dependent, we have such a relative excess of CPU performance that their CPU dependence just won't make a difference. GPUs will have to get much faster and games much more demanding before a slight difference in CPU utilization will come into play, if at all.
 
IdiotInCharge said:
Without calling names, it's obvious you have an AMD bias here, and that is not appreciated, and certainly won't garner you any respect.
Click to expand...

You bias guys are funny. It seems someone always has their panties in a wad about what companies we are biased towards. One day, AMD, then NVIDIA, then Intel, and so on and so on...you get the picture.

It comes down to this. If you don't respect our content, then don't read it, don't use our forums, don't complain about something you made up in your head and convinced your self that you have not one single shred of evidence about. That little red X up in the corner of your screen will cure every single bias concern you have about me or any of the HardOCP staff.
 
IdiotInCharge said:
To add to that, consider that the game typically has to load stuff into main memory before piping it to the GPU(s). Main memory is considerably slower than any PCIe 16x interface.
Click to expand...

You sure about that? PCIe 3.0 x16 is 16 GB/sec one-way, while main memory (on Sandy, YMMV) is 20+ GB/sec depending on speed. I guess it depends how you define it, but main memory isn't considerably slower, I don't think.
 
Forceman said:
You sure about that? PCIe 3.0 x16 is 16 GB/sec one-way, while main memory (on Sandy, YMMV) is 20+ GB/sec depending on speed. I guess it depends how you define it, but main memory isn't considerably slower, I don't think.
Click to expand...

Not in raw bandwidth necessarily, but CPUs are pretty heavily loaded when a level is loading and thus eating into that bandwidth with additional data transfers, latency, and overhead.

It's the same basic idea as SSDs present when boot and application load times increase substantially, but game level load times are barely affected as compared to an HDD. It's also why it's perfectly fine to use a slower (even a green!) drive for game installs and and a smaller ~120GB SSD instead of one large SSD, where you just use links to put the most used/slowest loading games on the SSD. Just like PCIe 2.0 vs. 3.0.
 
Thing is, it doesn't matter- even if the Nvidia cards are a little more CPU dependent, we have such a relative excess of CPU performance that their CPU dependence just won't make a difference
Click to expand...

The noticed "CPU dependency" is possibly related to the legacy interrupts instead of the message signal interrupts (MSI); which nVidia cards still do not do. Legacy interrupts can add a little more CPU processing, and can be a little more latent in handling.
 
Forceman said:
You sure about that? PCIe 3.0 x16 is 16 GB/sec one-way, while main memory (on Sandy, YMMV) is 20+ GB/sec depending on speed.
Click to expand...

Let's run with that. Assume that a game has 1 GB (nice round figure, pulled out of my hat) of textures and other data to load, and it's all in main memory. Assume 3 cards, running at x8/x8/x4. That will take half a second for all three cards if they're loaded sequentially on PCIe 3, and a second on PCIe 2. A trivial and barely noticeable difference.
 
good article and rigorous methodical testing!

glad to know that 3.0 has no real benefit at this point in time. no need for early adoption, it'll be standard soon enough anyways (and probably still won't matter :p )
 
Very interesting stuff. Great article. Thanks a lot, guys.

I think X79 would have made a better comparison test bed just because it could be apples to apples on everything (except PCIe 2.0 to PCIe 3.0) due to the amount of native PCIe lanes versus Z77 (for example) and their reliance on PLX. Although I completely understand the logic behind helping people make purchase decisions (P67 to Z77 in this case).

Early results of PCIe 2.0 versus 3.0 showed that there was a HUGE gain going to PCIe 3.0. But I think NVIDIA has made some strides in drivers since that time (it was April - so pretty early in Kepler's life). Reference: http://www.evga.com/forums/tm.aspx?m=1537816.
 
Great review! This reminds me of the old days of DDR speeds and everyone pushing the highest speeds with little to no real-world performance difference.

I'm also happy to hear this, since I was considering a mobo/cpu update along with a new higher-end video card, but it really won't make the amount of difference to make up for the money spent.

Thanks again!
 
trugate said:
Great review! This reminds me of the old days of DDR speeds and everyone pushing the highest speeds with little to no real-world performance difference.

I'm also happy to hear this, since I was considering a mobo/cpu update along with a new higher-end video card, but it really won't make the amount of difference to make up for the money spent.

Thanks again!
Click to expand...

DDR marked a point where memory bandwidth exceeded what was needed by the CPUs (except Netburst, but faster RAM wouldn't have saved it); as did AGP 4x. Since both we've been ahead of the curve.
 
In order to take away the 4-7% difference in clock cores, why not test the 3770k on a z68 motherboard that only has pcix 2.0 slots and a z77 motherboard with pcix 3.0 slots? Of course, it could be that the margin between the chipset performance could ruin this, but I remember reading that it is even less of a performance difference. I think it was 2-3%. I think that would be a better test, since it would be the same cpu. Just a thought
 
so I guess updating the BIOS on the P67 Revolution to run the Ivy Bridge chip and then adding another determining factor is/was too much to ask for.?
this way instead of having the 4-7% difference between Sandy and Ivy would not factor in at all.

just saying..
 
Has anyone run more recent tests with either of the suggested changes in the above two posts?
Shouldn't be too hard to get an Asus P8z68-V Pro non gen 3 vs an Asus P8z68-V Pro gen3 (or z77) for a more accurate comparison where the cpu itself is a non-factor.
 
A little late to the party but I want to say thank you for the review too. I recently got two 680s for my P67 gaming machine and I was wondering if I'm not using them to their full potential. The review put my fears to rest:) Not worth upgrading to a new platform yet. I have i7 2600 and 8GB of 1600 RAM and it looks like I'll stick to this setup for a while. Thanks!
 
You must log in or register to reply here.
Back
Top