Motherboards All you need to know about running multiple GPU's in x16 x8 x4 slots and the outcome.

The truth about running Multiple GPU's at x16 lanes vs dual X16/X8 - X16/X4 - x8/x8

This setup has the first GPU @ x16 Lanes and the second GPU in x4 lanes. The out come is he is only getting a 50% increase in FPS, where he should be nearly 90% if it was in a x16/x8 situation, and 100% Scaling if it was in a x16/x16.
Dirt3 @ 1080p, Single card = 122.1702128 average fps
Dirt3 @ 1080p, CrossfireX = 182.3529412 average fps
= around 50% increase in fps.

This is where it gets more interesting, as the more bandwidth you use when running an extreme resolution. The bigger the performance hit, and in this case with a x16/x4 configuration, he is losing FPS. Now if this was in a x16/x8 configuration, he would be expecting about a 70% Scaling increase, and at x16/x16 will be 100% scaling
And the high-res bench results were quite interesting.
The single card had a lower score and average, but a (slightly) higher max fps, where as the dual cards had a slightly higher average, but lower max. They are pretty close though

The left is a single HD6970 and the right is 2x HD6970 Crossfire


Here's another example of dual GTX 580's in a x16/x8 and x16/x16 configuration.

So at x16/x8


And at x16/x16



Know what motherboard to go for, for your desire set up, as buying something cheap, and then wanting to add a second graphics card, when running 1080p resolution +. Will leave you penalised and spending more money on a motherboard to give you the extra scaling. x16 PCIe and x16 PCIe slot with 4x lanes (x16/x4) is a scaling killer, and should be avoided. A minimum would be x16/x8 if you plan on running SLI or Crossfire. But expect a average of 10% lost at 1080p and more as your resolution increases.

I will be adding more to this as time goes on...

 

Good thread - Worth getting this pinned. Hopefully my new board on Monday will prove 100% that this is an issue that people should be more aware of. Obviously there is no mention of the potential performance hit when using multiple cards anywhere on the mobo box!

 

Yep! Make it a sticky please.

As soon as you get your benchmark results, I will be posting your finding in here, with hopefully near 100% scaling.

 

Interesting stuff.

Would be good to know, out of the current crop of boards recommended by people WHICH
ones actually do the full x16/x16 or are limited to x16/x8, x16/x4 in SLI and/or Crossfire and
how to ensure you are running in the best scaling for your motherboard.

Looking forward to my build in April i'm probably looking to SLI two 570's, maybe 580's
depending on price at the time. This sort of info would be really useful.

 

ah seems this would be similar if I crossfired my board, my slots only run at x8 when pairing up. Def +1 for sticky.

 

Hello all,
I run 2 5850's at 1080p, what are the best settings in ccc to have it set-up? my monitor is 60hertz, its the filtering etc what confuses me- triple buffering and trilinear really get me stumped, i'd like to play me games with most quality too.
Many thanks.
ps, ive seen a few conflicting views on v-sync, whats best for me there?

 

What about running 2 cards on x8 ? This is what most mid-range boards are capable of. How does that scale?

 

+1. This is what I'd be interested in seeing.

 

Just a point to make - having two cards on x16 lanes does not necessarily mean 100% scaling. It's got a lot to do with driver and game optimisation, which rendering method you're using. Alternate Frame Rendering, the method which tends to give the best returns is known to be prone to a phenomena known as 'microstuttering'. Scissoring/Checkerboarding relies on both GPUs and in somecases has been known to give worse frame rates (this is fairly rare though).

It's got just as much to do with how the application itself is rendered than how much bandwidth is available to the cards. Implying that having two x16 lanes means you will get 100% scaling isn't entirely right.

Well, here's what these terms mean:

Triple Buffering means that your GPU will pre-render up to 3 frames for the purposes of smoothing the graphical output. So, the frame you are seeing on the screen possibly isn't the one the GPU just rendered. This means that in a particularly graphically complex moment (say, an explosion happens or something) the GPU has those 3 pre-rendered frames to fall back on whilst it tries to render the more complex one. NOTE: This will only work with V-Sync enabled, otherwise your GPU will just spit out frames as soon as they’re rendered.

Pros: Smoother FPS.
Cons: Can introduce horrible input lag.
Do I use It? No, too much input lag.

Tri-linear filtering is a technology used to reduce aliasing and improve fidelity on textures. This means you get crisper, clearer textures when looking at objects from an angle. However, it's old hat and has been completely superseded by Anisotropic Filtering, meaning nothing uses it anymore. So I don't think you even need to worry about this unless you play a lot of really legacy games.

VSync (short for Vertical Synchronisation) means that the GPU will output frames in time with the vertical blanking interval of your monitor (basically, when your monitor starts to output a new frame). So it will output frames at the same speed as the refresh rate of the monitor or a product of that (So a 60Hz monitor would cause a GPU to output frames at 60, 30 or 15 FPS). This eliminates any screen-tearing artefacts that can be caused by your GPU outputting frames higher than the monitor’s refresh rate. Basically, your monitor will start drawing one frame, before receiving data about a new frame and finishing the drawing cycle by outputting the rest of the new frame instead. This results in an ugly ‘tear’ somewhere in the middle of the frame due to it actually being comprised of two frames.

Pros: Eliminates screen tearing
Cons: Comes with a performance hit.
Do I use it? Yes, if the game I’m playing consistently outputs 70FPS+ with VSync off. If it doesn’t, then I’ll disable VSync as the chances of any screen tearing happening are minimal and it increases overall FPS.

 

@TheKrumpet -You are correct in many aspects, but most modern games are now designed to run with multiple cards, and therefore you can expect more than 100% scaling in some. But this again depends on what components you use, to what resolution.

More to the point, this thread was designed to show people that mixed PCIe slots do make a differences, when you come to pushing the limits of PCIe slots that run slower than x16.

Anyway, thanks for your input.

 

I will get around to these test, and post some results up when I get a chance.

 

Enjoy...

 

Did you just come in here without reading the thread with our hard evidance of proof with some clips that don't show you nothing?...

 

I dunno I quite enjoyed the videos, good way of visualising the kind of testing and configurations that goes on. Also kept waiting for the bugger to touch the wrong part of his mobo and go flying across the room now that'd be interesting FPS wise. Still think this should be made a sticky.

 

Those benchmark for the 580 in the OP are mine. I've tested this myself, there is definitely a marked FPS drop when using 16x 8x over 16x 16x. That was at 1080P resolutions as well. Later tonight I'll do the same tests at 2560x1600 with Heaven again, and I bet the FPS drop is more pronounced still.

 

Those test results are interesting. It wasn't too long ago that TechPowerUp did some testing on PCie 2 lanes using a HD5780 showed that PCIe 2 coped well. Shows how much graphics cards have moved forward. Bring on PCIe 3.

 

"you can expect more than 100% scaling in some. But this again depends on what components you use, to what resolution."

'Tis witchcraft I tell yee !



dunx

P.S. I have seen a drop in BOINC performance myself... not a gamer sorry !

 

+1 for this too, I actually game at 2560x1600 on a 30" dell with 2 x 580's on a evga classified x58 e760 so I'm using x16 x16 but I'd like to upgrade my CPU mobo and ram but alot of the z68 boards don't have 16x 16x due to the chipset only providing 16 lanes so I don't know what to do!

 

I'm benchmarking now, but unfortunately, I can't run x8 x8 because the slots are too close together. I can do 16x 16x, 16x 8x only. I'm also doing single card 16x and 8x.

Will report back soon.

 

So is there an easy way to know what SB boards do 16x 16x and can accommodate two gfx cards
in the relevant slots?

For example, reading specs on scan;
ASUS Maximus IV Extreme R3
Expansion Slots:
1 x PCIe 2.0 x1
4 x PCIe 2.0 x16
1 x PCIe 2.0 x4

ASUS P8P67 Deluxe R3
2 x PCI (32 bit)
2 x PCIe 2.0 x1
2 x PCIe 2.0 x16
1 x PCIe 2.0 x16 (x4 Bandwidth)

So this is saying MIV has 4 full x16 slots, put your 2xGFX in any of the big slots they will be
x16 x16 definitely.
but the P8P67 only 2 will do that? (I will assume top two big slots are the full x16).

Or does something else say whether they run at x16 x16?