GeForce GTX 980 PCI-Express Scaling

[birdie]

Sandy Bridge users have nothing to worry about: PCIe 2.0 @16x is more than enough for the GTX980.

 

[Aquinus]

Sandy Bridge users have nothing to worry about: PCIe 2.0 @16x is more than enough for the GTX980.

It seems to me that in most cases 2.0 @ 8 lanes is enough as well. I find that if you /really/ want more PCI-E lanes, the real reason is not to drive more GPUs but to add other components such as a network card, RAID controller, or a PCI-E SSD which is much more common need on workstation and servers than general consumer devices.

 

[Champ]

Some people believe and preach everything some would post on a forum. I'm one of those people that need scientific backing. This was a great test and we need more myth busting.

 

[msimax]

i would also like to see multi gpu testing since I've been pondering moving to x99 for my 3 290x to see if it will offer better peformance than 8x 4x 4x in games when using max AA and a move to 4k

 

[stren]

Good work but not unexpected given that single GPU is not expected to really be harsh on pcie bw. Another +1 on SLI testing, that's where PCIE bandwidth suddenly becomes more limited. The only problem is that the works suddenly gets out of control if you try and be as thorough. Personally I'd assume anyone running the latest GPUs is also running something recent e.g. pcie3 and so only run

2 way SLI - 4790K vs 4790K + PLX vs 5960X e.g. 8x/8x vs pseudo 16x/x16x vs real 16x/16x
4 way sli - 4790K PLX vs 5960X e.g. pseudo 8x8x8x8x vs real 8x8x8x8x

Of course particularly once you get to 4 way, you're going to need a good OC on the CPU and a really demanding monitor setup to avoid being cpu throttled.

 

[Brusfantomet]

Great test, but it does make me wonder how a CF setup with a XDMA engine (290 or 285) will handle the loss of bandwidth.

My expectation is that in titles where there is no difference between 16x gen 3 and 16x gen 2 for the 980 there will be no/minuscule difference for CF 290X, but in titles where the bandwidth mattered it will have a bigger impact. a preliminary test with two games that were bandwidth sensitive and two that were not would give a interesting indication.

 

[Wark0]

Great review. Do you plan to do the same in SLI ?

 

[Aquinus]

Great review. Do you plan to do the same in SLI ?

I would expect the results to be the same if both cards are running at the same speed for any given speed but, I'm curious as well.

 

[jihadjoe]

Its always funny when people cite high resolutions like 3x1080p, 1440p, or 4k as a reason to get more PCIe lanes, when it's actually the lowest resolutions that see the most scaling.

 

[Ubersonic]

Nice review still no real difference.

Wow big boost from WoW MoP to WoW WoD. 161FPS to 231FPS at 1080p

It's not actually a good thing, the extra FPS at high settings is mostly due to Blizzard dropping MSAA in favour of FXAA or CMAA, so basically the game runs faster but looks a lot worse, the only way to get decent AA currently is to set the game to DX9 mode then force MSAA through drivers, which results in worse FPS than MoP as it's using the DX9 path instead of the more efficient DX11 one, hopefully they will end up adding SSAA as a lot of people are raging about this on the forums

 

[Delta6326]

It's not actually a good thing, the extra FPS at high settings is mostly due to Blizzard dropping MSAA in favour of FXAA or CMAA, so basically the game runs faster but looks a lot worse, the only way to get decent AA currently is to set the game to DX9 mode then force MSAA through drivers, which results in worse FPS than MoP as it's using the DX9 path instead of the more efficient DX11 one, hopefully they will end up adding SSAA as a lot of people are raging about this on the forums

Oh interesting, I haven't played since WoD.

 

[Deleted member 138597]

It's funny that even today people will prefer a 5930K over a 5820K in a dual-GPU system over basically nothing.

Correct, even I set out for 5930K and was like "OOO! It has 40 lanes, so it can run dual GPUs in x16. You know, x16 is da bestest for GPUs". But seeing this review, I now even consider 4790K in the list (and probably will cross out the 5930K).

 

[Deleted member 138597]

@W1zzard

You didn't do any synthetic benchmarks though.

Is this result limited to Maxwell only or other architectures too, like Kepler, Fermi and AMD GPU architectures?

And as other people are asking, does SLI have a impact on the result? Why Nvidia chose to limit SLI to at least PCI-E 3.0 x8 and AMD on the other hand PCI-E 3.0 x4?

 

[msimax]

Correct, even I set out for 5930K and was like "OOO! It has 40 lanes, so it can run dual GPUs in x16. You know, x16 is da bestest for GPUs". But seeing this review, I now even consider 4790K in the list (and probably will cross out the 5930K).

the 5930k comes in handy if you have multiple add on cards instead of a dead 4th slot

 

[CrAsHnBuRnXp]

Nice review still no real difference.

Wow big boost from WoW MoP to WoW WoD. 161FPS to 231FPS at 1080p

And yet people still believe that WoW is CPU limited. But its no longer the case which is clearly proven here.

 

[Zkanuck]

+1 on SLI & Crossfire testing.

I don't think you'd have to go to all the trouble to test all different modes - just the only ones that matter between Z97 vs X99 - SLI or crossfire 2-way at 8x8, 16x8, 16x16 and 3-way or 4-way configs with Z97(PLX) vs X99 28/40 lane CPUs.

 

[cerealkeller]

It would be awesome if you could re-do this benchmark with a set of 1080s. The word is that there is actually a measurable difference with the GTX 1080s on PCI-E 3.0 x16 compared to slower configs

This is a great thread though, I've kept it in my bookmarks for a long time and have referenced it on occasion.

 

[W1zzard]

It would be awesome if you could re-do this benchmark with a set of 1080s. The word is that there is actually a measurable difference with the GTX 1080s on PCI-E 3.0 x16 compared to slower configs

This is a great thread though, I've kept it in my bookmarks for a long time and have referenced it on occasion.

Thanks for reminding, I'll try to find time, won't be that soon though, doing a full rebench right now with new games and drivers, but after that maybe.

 

[Aquinus]

Thanks for reminding, I'll try to find time, won't be that soon though, doing a full rebench right now with new games and drivers, but after that maybe.

Definitely something to consider with modern mainstream platforms given the number of PCI-E lanes but, even my 3820 will do PCI-E 3.0 and 40 lanes, which is honestly a lot of I/O capacity (even more so if you consider how old my machine is now.) I honestly don't think that PCI-E 3.0 @ 8x is any better than 16x and if 16 is more than 5% faster than 8, I'll be surprised. Once things like textures have already been loaded, what GPUs crave from the CPU is latency until bandwidth becomes the limiting factor. Of course, the application itself and how it's implemented plays a big part in this but, generally speaking, it seems to be the case that PCI-E demand doesn't scale at the same rate as performance itself does.

Either way, I look forward to an updated PCI-E scaling review. You've always done a great job with this particular review, @W1zzard... not to set the bar high or anything.

 

[stinzza]

thanks for as always a lovly read

just wondering if now programs and games will start to multithread more since amd finally enterd the
scene, and that the 512mb/s bus of the 16x 2.0 will not be enough to handle more info over the bus handeled by the new cpus and gpu?

im sitting on an old 2500k z68 chipset with 1 16x 2.0 and 1 "electric" 4x pcie lane that goes through the chipset then to the cpu instead of pcie16x lanes straight to the cpu,
hence lag.. it says that in the rev 2.0 of my motherboard version, if one grafics card in populated with an i7 3770k cpu, it conforms
to 16x 3.0.. is that real 16x 3.0 or will latancy be way higher since the pcie 4x lane goes throuh the chipset then the cpu?

should i then bye a 2700k that is higher binned, easyer with heat and overclock and 2.0, or opt for ivy 3770k and pcie 3.0(?) and a bit faster clock for clock? and say a titan xp...