GeForce And Radeon On Intel's P67: PCIe Scaling Explored

Intel’s Sandy Bridge-based processors dramatically advance gaming value by increasing performance at lower prices than LGA 1366-based configurations. But is the platform it sits on worthy of that CPU? We test three slot configurations to find out.

Tom's Hardware's Three-Part, 3-Way Graphics Scaling Series

Part 1, The Cards: Triple-GPU Scaling: AMD CrossFire Vs. Nvidia SLI
Part 2, The Slots: GeForce And Radeon On Intel's P67: PCIe Scaling Explored
Part 3, The Chipsets: P67, X58, And NF200: The Best Platform For CrossFire And SLI

We’re constantly impressed by the performance of Intel’s latest processors, which offer both higher IPC (instructions per cycle) and better overclocking headroom than anything seen previously. This is particularly true in games, where the extra execution resources in Intel's Core i7-990X go unused anyway.

Yet, in spite of the Sandy Bridge architecture's superior gaming performance (or at least better value, since you don't need to spend $1000 on a six-core processor to get equivalent frame rates), many enthusiasts insist on the added expandability of the old LGA 1366 platform. This makes sense in light of the P67’s incredible handicap: namely its 16-lane processor-based PCIe graphics interface.

But do we really need more than 16 lanes? Aren’t eight lanes enough for a single card? If that’s true, expensive and power-robbing PCIe bridges certainly wouldn't be necessary to support multiple cards. We’ve even heard stories that the PCIe 2.0 x4 slot supported through the P67 chipset is adequate for some graphics cards.

Before we move on to see how the P67-based platform stands up against the X58- and NF200-enhanced versions of both, let's get to the basics of PCIe scaling using modern cards on the best-possible gaming processor. This is Part 2 of our multi-card scaling series, but will there be enough graphics performance difference between various slots to justify a Part 3?

This thread is closed for comments
71 comments
    Your comment
  • geofelt
    These tests were done with a single card, on X16/X8/X4 slots. Fine.
    But... Who would use anything other than a X16 slot if they had one?
    The only real use for a X8 slot would be for sli/crossfire where the addition of a second card should result in an Increase of performance, not a decrease.
  • carlhenry
    it would be nice if you included the GTX 570 in the x8/x8 and x16/x4 test. the 570 flies over the 6950 on the single card config but i was curious how it would do since i think the AMD's scale better than nvidia's. would the 570 still lead because of its advantage? or would AMD even it out because of its scaling (if any) "advantage"
  • classicaxe
    ^They already did an article on that man
  • joytech22
    Can you guy's do an article on how performance is affected if you SLI/Xfire using PCI-E 16x slots running @ 4x?

    3 way would be preferable because if performance is still adequately faster I'll consider it.
  • dalauder
    Good comments. Can we please see 8x/8x and 16x/4x since that comparison is relevant? I get the impression that somehow SLI/crossfire reduces the performance hit of x4 lanes but I'd like to see numbers.
  • Crashman
    joytech22Can you guy's do an article on how performance is affected if you SLI/Xfire using PCI-E 16x slots running @ 4x?
    I think you missed a page then!
    http://www.tomshardware.com/reviews/pci-express-scaling-p67-chipset-gaming-performance,2887-10.html
    The numbers were there all along!

    carlhenryit would be nice if you included the GTX 570 in the x8/x8 and x16/x4 test.
    Well, you should probably read the linked page too then. There's no point in artifically creating a configuration (by taping lanes or whatever) that doesn't exist in real life, is there?

    "While Nvidia prevents SLI from functioning on PCH-hosted lanes, x16/x4 configurations are completely possible in CrossFire. But should they be? We tested our motherboard in both x8/x8 and x16/x4 configurations to find out."
  • dalauder
    Yeah...my bad.
  • Crashman
    dalauderYeah...my bad.
    I didn't mean to call you out to that extent, here's a link to the forum part of this thread:
    http://www.tomshardware.com/forum/2887-56-geforce-radeon-intel-pcie-scaling-explored
    I'm going there to delete your quote from my response.
  • rolli59
    Looking forward to see the third article!
  • cats_Paw
    Im guessing that 8x lanes are mostly enought. I do belive that it would depend on how fast an actual gpu is, as well as how much ram it has, and how big is its bandwidth.
    I means, its logical, but mayb not true :D. Would be nice to see this test on a GTX560 Ti, since it has a lot of headroom for OC, then compare oced version vs non oced. Also this might be interesting in GPUs that have diffrent versions with more and less RAM.
    Just my 2 cents :D.
  • Crashman
    cats_pawIm guessing that 8x lanes are mostly enought. I do belive that it would depend on how fast an actual gpu is, as well as how much ram it has, and how big is its bandwidth.I means, its logical, but mayb not true . Would be nice to see this test on a GTX560 Ti, since it has a lot of headroom for OC, then compare oced version vs non oced. Also this might be interesting in GPUs that have diffrent versions with more and less RAM.Just my 2 cents .
    It's...different for sure. Just about anything that can "bottleneck" graphics performance will have a greater effect on AMD GPUs than on Nvidia GPUs. CPU too slow? DRAM? PCIe? It's going to hurt Radeons more than GeForces.

    Of course, Tom's Hardware has all the cool kit that helps editors overcome those bottlenecks.
  • feeddagoat
    Finally the x16/x4 question answered!! It got me thinking tho, since ATI allows "mismatched" cards do it matter which slot the faster card is in......
  • kilo_17
    Great article!
  • rpgplayer
    feeddagoatFinally the x16/x4 question answered!! It got me thinking tho, since ATI allows "mismatched" cards do it matter which slot the faster card is in......



    that's an interesting question, like a 6950 and a 6970, would it be faster with a 6970 in the first pcie slot or the second.
  • silverblue
    I'd love to know what AMD cards are doing that requires so much more bandwidth... faster memory speeds, perhaps?

    It might be worth conducting at least two way Crossfire/SLi tests more often in future with new card reviews or, at least, follow-up reviews, though I can imagine it'll delay such articles a little. Perhaps using AMD cards on P67 is limiting their scaling, even if it's already very good (with the 6000 series)? Would using two mid-range cards have the same effect?
  • ubercake
    AHHHH! Good info. Can't wait for the next part!!!
  • ubercake
    One thing though... No BFBC2 performance charts?

    This is a mighty popular game. Does EA make reviewers pay them to mention their game in reviews?
  • wishmaster12
    it depends on your chipset if you want awsome scaling get a motherboard with a better chipset
  • ta152h
    One nitpick are some of the conclusions, like x16/x4 don't work well for AMD, or NVIDIA. It could be correct, but you haven't proven anything in this article. Although, one could now make that case for the LGA 1155/1156 platforms, and this certainly goes a long way in showing those deficiencies, you'd still have to prove this isn't platform related, or extenuated. Your assumption is being four lanes wide is the problem, but we'd have to prove that it's not related to where the lanes are attached to before we could fully prove it's the width.

    For example, would this be the case on LGA 1366, where all the PCIe lanes are on the chipset, and not some going through the chipset and some being part of the CPU package? In real world situations, it's not so important because you'd have less cause to use an x4 in LGA 1366, but for academic purposes, and a better understanding of what's really going on, it could be interesting.
  • azduman
    So, i currently have a 5770 on w/a i5-750, obviously w/1156 chipset, would it be a good idea to get another 5770. Reason i ask is i have 2 pci express slots but second one is at 4x lanes. I have a asus p7p55d le motherboard. Do i have to buy same card or could it be a different brand at different speeds. Or better to get the 6950 2gb i want to buy.
  • nd_hunter
    silverblueI'd love to know what AMD cards are doing that requires so much more bandwidth... faster memory speeds, perhaps?


    I too am wondering why this is. My non-educated guess would be that the AMD card has more memory, requiring more bandwidth. I wonder if there would have been such a sizable hit on the 1GB version of the 6950 when in the x8 or x4 slot.

    And thank you Tom's for answering the x16/x4 Xfire question that a lot of us had. However, it would have been nice to see the single x16 card's numbers on the same chart to illustrate the moderate gains an additional card in the x4 slot gives us. Thanks again for another great article.
  • wolfram23
    Well, I would have thought the drop would be more on these cards compared to the last time we saw this with a 5870. Makes me wonder where I'll go with a future card upgrade...

    I would like to see 16x/16x vs 8x/8x done more thoroughly though. Say an X58 vs P67 (NF200) vs P67 (regular).
  • professorprofessorson
    Do really got to say, this kind of test should have been done on both a Intel and AMD based platform. Especially the x8/x8 and x16/x4 test. For that matter, would also have been good to see not just higher end cards used, but also cards like the GF 460 and HD 5770, so that GPU performance can be factored into how much bandwith is needed per card per pci-e slot.

    That way it could be shown whether or not the lower end cards also take just as much of a performance hit at x8 and x4 as the higher end cards do, or if they do not get held back nearly as band. Reason I say this is that prior test have been done on prior graphics cards before using x4 and also x16/x4 dual card modes, and while showing a performance impact, it was no where near as bad as what you have shown. One would definitely assume that the newer generation higher end cards are indeed going to demand more from their pci-slots, as they can do more. But is this ultimately the same case (as of now anyway) with the more mainstream cards from the Geforce 450, 460, 550 line, along with the Hd 5770, 5830, 6790, and 6850 line?

    Ultimately many factors are going to come to play. One is the games themselves. F1 2010 and Just Cause 2 for example took major performance hits in your test, even in Crossfire x16/x4. The other titles seemed to take less of a impact however, depending on AA settings, ect, with fps loss of maybe 4-7 fps depending on said title, showing that performance loss will definitely vary per title, with some titles still benefiting quite well from Crossfire regardless of the second card running at x4.


    Another factor, as stated, is what platform is being used. You did not include any AMD based platform in this test, nor any more commonly used mainstream cards. Due to this the scope of your test was far too controlled by limiting factors, and only provided quite limited results that honestly can not speak for every x16 through x4 configuration, let alone every x8/x8 and x16/x4 dual card configuration, even with a margin of error provided. In other words, this is hardly what I would consider the end all of articles on pci-e scaling explored. Sorry Tom, on this one, no dice.
  • ntrceptr
    I thought Tom's was better than this. The PCIe speeds only matter for data transfers to/from the card. Once a game/level is loaded there is much much less data going back and forth (usually), unless.....

    You choose a graphics card with alot less memory causing the game to fill the graphics card memory and rely on system memory for the overflow of textures and geometry. Then you should see much bigger variations between the 4x/8x/16x slots.

    My opinion is the tests you are doing are flawed as all the games textures and geometry fit within the graphics card memory and the only real traffic thru PCIe is the updates to geometry/shaders/etc...nothing to big.

    Maybe benchmark the difference in load times too.