NVIDIA GeForce GTX 1080 PCI-Express Scaling

[axiumone]

What about sli? My own testing revealed to me that there's a substantial enough difference to go with x16/x16 if you're investing heavily already.

http://www.overclock.net/t/1616578/...8-w-titan-xp-sli-benchmarks-and-results/0_100

 

[EarthDog]

What about sli? My own testing revealed to me that there's a substantial enough difference to go with x16/x16 if you're investing heavily already.

http://www.overclock.net/t/1616578/...8-w-titan-xp-sli-benchmarks-and-results/0_100

I'm floored at those results...

 

[ThomasS31]

Same color = similar bandwidth.

Would have made more sense to color the same versions... ie.: 3.0 = color 1, 2.0 = color2... etc.

Then it would be clear that how they compare to each other.
At least for me.

 

[R0H1T]

Didn't anyone notice that the difference between successive generations of PCIe was less @4K than say 1080p, I figured it'd be the opposite?

 

[etienne31]

What about stuttering ?

 

[Vayra86]

Don't you all feel like being cheated by the industry?

They sell each "next step" as revolutionary. Well, 5 fps difference between PCI 2 or 3, x8 or x16 at 4K is not revolutionary at all!!!

Not really. The fact is, because there is a very real difference, even though it is small, the bandwidth has been growing along with the power of graphics cards. Back in the PCIE 1.1 days we had significantly less powerful cards. If anything this bench shows the need of PCIE 3.0 on any serious gaming rig. And that standard really isn't that old - the 1080 also isn't the most powerful card today, so a Titan will show us a larger gap.

I have trouble with the conclusion of this review because of that. The picture I saw on Toms' back in 2011 or 12 was completely different, with powerful cards not even showing a single % difference. Today, as FPS goes up, the bus gets actually saturated. I see some significant differences at 1080p and let's not forget we don't run most of the games in this bench suite at high fps, it is mostly eye candy and console port gaming and we are pushing ultra instead of high, which generally accounts for a major fps drop.

Also keep in mind that VR likes to run at high FPS.

 

[kn00tcn]

What about stuttering ?

i forgot to make this comment, this is exactly what matters!

seeing these really fast cards have fps drops on lower pcie standards by themselves is one thing, but seeing similar results on older slower cards (past pcie articles) means something else is reducing the fps that's NOT peak bandwidth (the old cards wouldnt have saturated if the new cards are much faster)

 

[a_ump]

thanks W1z. and my 2 cents. its rather cool to see that it does make a difference. However, i also see in some of those results, what looks to me like there's no indication that 3.0 isn't already maxed and that more bus bandwidth could improve performance further. I mean there's a decent crescendo on the FPS in many titles....guess we won't know if even PCI 3.0 is bottlenecking until 4.0/3.X is released.

 

[Vayra86]

thanks W1z. and my 2 cents. its rather cool to see that it does make a difference. However, i also see in some of those results, what looks to me like there's no indication that 3.0 isn't already maxed and that more bus bandwidth could improve performance further. I mean there's a decent crescendo on the FPS in many titles....guess we won't know if even PCI 3.0 is bottlenecking until 4.0/3.X is released.

CPU/GPU bottlenecking is what I'm attributing that to. CPU for 1080p, GPU takes over at higher resolutions.

 

[Bender]

Nice PCI-E scaling review. Thanks

 

[cheddle]

sigh... why bother testing so many games!

You should do a test that ACTUALLY MATTERS like testing SLI and Crossfire scaling on high end cards across different PCIe bus speeds.

 

[_larry]

I got one for you guys that is not tested here!! Well...sort of.

I just added another R9 290 to my rig. My mobo is an Asrock z75 pro3. It has 1xPCIE x16 Gen 3.0 slot and the other is 1xPCIE running at 4x Gen 2.0....

I have Crossfired AMD 7770s before and the performance was not that bad considering the small memory bandwidth of those cards.
HOWEVER, with the 2x R9 290's that have MASSIVE amounts of memory bandwidth, I am having stuttering issues and game crashes on a LOT of games...

I kind of knew what I was getting into...But I found the second card that can be BIOS modded into a 290X for cheap on Ebay...
I could just sell my old card and mod this one up if I can't tweak my drivers/mobo to cooperate with these beasts of cards.

 

[AquaeAtrae]

Thanks for the article. I'm seeing references to this in an attempt to understand the impact that new Thunderbolt 3 eGPU connections might have. Some laptops are wired with all x4 PCIe lanes while others only have x2 lanes. We have yet to see any tests for these.

More importantly, we should keep in mind and test for the fact that PCIe bandwidth has little to do with the GPU's output speeds (FPS). Once the textures, 3D models (map, characters), cameras, and lighting are all loaded into those 8GB of VRAM, very little PCIe bandwidth is required for the CPU to request an updated render with small changes in angles in positions.

However, loading a new map or region into VRAM may temporarily utilize all available bandwidth and take much longer. When traversing open worlds (map regions) like Far Cry, one might see this in-game. In games like Battlefield 1, this means you'll never start a match fast enough to jump in that plane. Certain games and maps may even timeout (e.g. Call of Duty: Advanced Warfare) before starting.

Any tests comparing PCIe lane scaling should really identify and report these differences by comparing map load times and by testing games in different video modes (DX11 vs DX12 vs Vulkan). With Thunderbolt 3 solutions, we should also expect an additional small loss from its own overhead.

 

[Dippyskoodlez]

Thanks for the article. I'm seeing references to this in an attempt to understand the impact that new Thunderbolt 3 eGPU connections might have. Some laptops are wired with all x4 PCIe lanes while others only have x2 lanes. We have yet to see any tests for these.

Plenty of information is available, thunderbolt 2 gpus have been working for a long time.