AMD Radeon HD 5870 PCI-Express Scaling

[W1zzard]

AMD's latest generation of graphics cards offers unprecedented single GPU performance levels. Such performance requires an optimum environment, especially important is PCI-Express bandwidth. We investigate if these cards can still deliver on a PCI-E x8, x4 or even x1 link.

Show full review

 

[newtekie1]

Very interesting. I wasn't surprised by the x8 performance, I expected next to no performance loss. However, the x4 performance was surprising, I expected a much bigger drop there, but 5% wouldn't even be noticeable.

 

[dir_d]

Gives me confidence in a P55 board with 2 5850s. 1 5850 now then when they drop in price after nvidia answers pick up another

Edit.. Itchy trigger finger made me buy a 5870 lol

 

[Athlon2K15]

this review makes me more comfortable with my new build thanks w1zzard!

 

[RaPiDo987]

nice!!

 

[human_error]

Wow i was expecting to see a bigger range between the different pci-e speeds - good info to know so thanks for the review w1zz

 

[A Cheese Danish]

Very nice! Definitely putting the 5870 x2 on my to get list. If/when they release it.

 

[Sihastru]

This means 2 x 58x0 CF setups are viable even on the old P45 motherboards (2 x 8x PCIe 2.0). I am very surprised about the 4x PCIe 2.0 performance... it's too good.

 

[stupido]

nice review...
10x

 

[Tuvok]

some assumptions are wrong, imho...
if single card 8x 2.0 is 99% of 16x 2.0 you can't say the same for dual cards, since you got much more overhead in crossfire

 

[tzitzibp]

great stuff... thanks W1z!

 

[tkpenalty]

hey w1zz mind running 1x and 4x with an overclocked PCI-E bus speed of like 110?

 

[jagd]

This is insane ,how could you manage publish both tests same day ?Thanks ,anymore tests on the way ?

 

[W1zzard]

yes 5870 crossfire, we'd had that too but customs took the card and wanted to play crysis with it

 

[newtekie1]

I assume the HD4870 corssfire review will be on the current testbed. I'm wondering if you could do a follow up review to this one, that shows crossfire PCI-E scaling, by taping off both cards and limitting them to x8, x4, and x1 speeds. It would give a good indication of how much more important the PCI-E bus is in multi-card setups. And really give a good indication of what people should expect when running two of these cards in a P55 or P45 board.

 

[btarunr]

Yes, that's part of the CFX review. Both cards will be taped off at PCI-E 2.0 x8 for one set of scores across all tests.

 

[mdm-adph]

Interesting proof of concept! Nice to see some scientific reviews in the community.

 

[PVTCaboose1337]

Very cool way to do it without cutting up the card like that idiot did with a 7800gs! (yeah it was photoshopped but still...)

 

[newtekie1]

Yes, that's part of the CFX review. Both cards will be taped off at PCI-E 2.0 x8 for one set of scores across all tests.

Excellent! I can't wait for the review!

 

[Sasqui]

W1z - you rock - what a great article with suprising results.

 

[lemonadesoda]

Extremely useful article and analysis. Thanks. This will make good reference material for the future.

I was also surprised that you get 95% of the performance with x4 until I remembered that x4 v2.0 is the same as x8 v1.x...But good to know, and also good to know that the figures still hold true even on a top end i7 system.

(And that dual CPU server boards with "just" x8 connectors will still make good workstations)

***

Just to kick some dust, if an x4 is good enough (x8 v1.x) then so is AGP! Yes, PCIe is a better general scalable format, but, from a technical standpoint, AGP and it's x8 speed equivalence and lower latency would still be in the running if high end products were still available. This thought really surprised me!

 

[5ilvgearX]

Thank you for the awesome scaling review

 

[mdm-adph]

ust to kick some dust, if an x4 is good enough (x8 v1.x) then so is AGP! Yes, PCIe is a better general scalable format, but, from a technical standpoint, AGP and it's x8 speed equivalence and lower latency would still be in the running if high end products were still available. This thought really surprised me!

As the people with Q6600's and 3850's on AGP getting 3dmark06 scores of 12000 have been able to prove.

 

[e6600]

but whats to say a future dx 11 game will not be able to thrash the x8 pcie bandwidth?
is it possible that at a now common res of 1920x1080, with some aa/16af applied, the next gen of games may prove p55 with electrical x8 pcie a bottleneck?

 

[lemonadesoda]

Remember what the AGP/PCIe interface is doing... transferring data. I think PCIex8 is going to be fine for a while.

Then remember what kind of data it is transferring

1./ textures
2./ coordinates
3./ calculated renderings

So long as there is sufficient memory on the card, then textures are loaded once or pre-loaded.

So long as the number of coordinates is a reasonable order of magnitude (hundreds of thousands, not hundreds of millions) then we are OK

So long as the GPU does the graphics work, not the CPU, then this doesnt happen.

So actually, the more memory on the GPU, the less bandwidth you need, because assets are preloaded and remain on the card.

Why are x16 benchmarks better than x8 better than x4 etc, even if the difference is only small. I bet the MOST OF IT is actually uploading new textures etc. If we looked as the MODE distribution frames-per-second http://en.wikipedia.org/wiki/Mode_(statistics) then I bet the results would be even closed. Remember that the "average" gets hit by just a few of the worst frames (loading new textures), whereas the "mode", the framerate 95% of the time, is probably the same.

***

An analysis by w1z of http://en.wikipedia.org/wiki/Mode_(statistics) FPS would be extremely enlightening and would put the discussion to bed.