Massive FPS gains in rf2 using PCI-e 3.0 x16 with higher end cards!

Discussion in 'General Discussion' started by DrR1pper, Sep 30, 2014.

  1. MJP

    MJP Registered

    Joined:
    Oct 5, 2010
    Messages:
    988
    Likes Received:
    21
    One only supports up to 16 PCI-E lanes whilst the other up to 40 but that wasn't the question.
     
  2. Ari Antero

    Ari Antero Registered

    Joined:
    Jul 27, 2012
    Messages:
    1,882
    Likes Received:
    829
    I don`t think I am going to use my time to explain How PCI e devices talk, I am 100% sure that if I answer to him with 10 pages he still finds something to ask.
     
  3. DrR1pper

    DrR1pper Registered

    Joined:
    Apr 29, 2012
    Messages:
    3,294
    Likes Received:
    36
    A bit condescending but okay.

    Ari, i'm not alone on this. You seem to be the only one who understands what you're saying.

    So far i haven't seen you give a complete explanation. You've only repeatedly said "x16/x16 can't be done" or something like that. To be honest you're not very clear with your thoughts but i won't hold that against you since your English is a million times better than my Norwegian. Having said that....doesn't change the fact that you haven't given a proper explanation.
     
    Last edited by a moderator: Oct 12, 2014
  4. DrR1pper

    DrR1pper Registered

    Joined:
    Apr 29, 2012
    Messages:
    3,294
    Likes Received:
    36
    Like MJP said, 4770K only has 16 lanes of PCI-e 3.0 available whilst 4930K has 40 lanes available.

    From what i understand, the x16 lanes available on the 4770k is split into x8/x8 when using SLI/Crosfire meaning each card gets half the maximum bandwidth they are capable of receiving. Your saying this is not true and that a 4770k can have x16/x16 when using SLI/Crossfire, right, by switching the available 16 lanes from one card to the other as needed? Do you have any evidence that shows this to be equal in performance as simultaneous x16/x16? And if so it would need to be with software that makes a difference when using x16 vs x8 (such as rf2).

    You already mentioned the PLX chip that can switch the full x16 lanes from one card to another as needed but from what i've read, it seems to be a detriment to performance with 2-way SLI (due to the induced latency) but works well for 3/4-way SLI. Am i correct?

    Again, just to be sure, we are talking about games/software that have performance to be gained from using PCI-e 3.0 x16 vs 3.0 x8 or 2.0 x16, such as rf2.
     
    Last edited by a moderator: Oct 12, 2014
  5. DrR1pper

    DrR1pper Registered

    Joined:
    Apr 29, 2012
    Messages:
    3,294
    Likes Received:
    36
    I just realised we can actually test the theory because both you and Spinelli have the same cards in SLI. Spinelli's CPU has 40 lanes and your's only 16. So it can be tested if it works as you say and we can see if you're performance equals Spinelli's.

    Are you up for that Ari?...in the interest of science.

    I might add it would be really cool if what you say turns out to be true. It would also mean i can get sli without upgrading my cpu and ram. I can't seem to find any SLI/Crossfire simultaneous x16/x16 motherboards for socket 1150 however. Even your motherboard is only x8/x8 when using SLI according to ASRock.
     
    Last edited by a moderator: Oct 12, 2014
  6. Spinelli

    Spinelli Banned

    Joined:
    Jan 28, 2012
    Messages:
    5,290
    Likes Received:
    32
    Ari, i'm sorry but you are massively not understanding this.

    This has nothing to do with CPU cores. The 3820 and 4820k only have 4 cores yet they still have 40 PCI-lanes, and the 5960X has 8 cores, yet it does not have more than 40 PCI-E lanes just because it hase 8 cores instead of 6, it still has "only" 40 lanes.

    DrR, you can't find any true 16x/16x LGA 1150 boards because there is no such thing, again, they only support 16 PCI-E lanes. However, again, you can get boards with the PLX chip. The boards with the PLX chip usually happen to be very high end boards. I had a PLX chip EVGA P67 FTW motherboard for my 2500K a couple years back. It wss around $300 or so brand new, I got it used for $175 though. I used it to run 3 HD 6950s + a soundcard. The new PLX chips are better and have less latency than the ones on the older motherboards. I think there have been 3 or 4 different PLX versions over the years (as of Sandy Bridge / P67 chipset motherboards).
     
  7. DrR1pper

    DrR1pper Registered

    Joined:
    Apr 29, 2012
    Messages:
    3,294
    Likes Received:
    36
    Well that certainly makes sense since the most number of pic-e lanes any socket 1150 CPU has is 16.
     
  8. alpha-bravo

    alpha-bravo Registered

    Joined:
    Dec 22, 2012
    Messages:
    1,374
    Likes Received:
    31
    interesting discussion. I own a Asus Maximus Extreme VI with a 4770k. ATM I use native 2 Slot x8 with my dual crossfire x config.
    With all new ISI tracks I have a very good gfx performance (stable 64 fps with multiview) for example Indy 2014 road with 29 AI I can run on high track detail with x8 multisampling all other settings on full except shadow blur which I set to fast.
    This is "why ever" not possible with ISI Portugal.
    With NVI you can measure the bus usage. As I understand the PCIe lane topic you need the additional lanes when the utilization of your PCIe bus is to high.
    Always when I measured the bus utilization on my old GTX670 dual SLI config the usage peaks was not higher then 60% and the constant stream was around 15 - 30%.

    All in all I'm not sure if more lanes are in any situation = better performance like you have it with for example a higher cpu clock.

    I see it more as a bandwith topic that can bring some advantages in some specific enviroments.

    Not 100% sure about it. It is only my opinion based on my actual knowledge.
     
  9. Saabjock

    Saabjock Registered

    Joined:
    Nov 13, 2011
    Messages:
    906
    Likes Received:
    19
    That GTX670 is still quite capable in the right PCI-e configured system.
    It's already been proven that even today's cards, can't saturate those 40 lanes...even in an SLI installation.
    As shown by these test, the PCI-e 3.0 setting does offer a substantial boost over 2.0 in RF2.
    Ripper was very skeptical when I gave him my result after a 200 MHz O/C and enabling PCI-e 3.0...'tripped his B/S alarm' ;) as he put it.
    I run one in a Biostar TP X79 which has (2x Gen3 x16) and (1x Gen3 x8) with an I7-3820 and 16GB of DDR1333 @ 7-7-7-20.
    It still churns out a very respectable framerate with all settings on and maxed out.
    It's important for guys to understand PCI-e 3.0 was not enabled by default in the early round of mb due to a potential timing issue.
    I still have to re-enable it after every driver update on my 3820.
     
    Last edited by a moderator: Oct 12, 2014
  10. DrR1pper

    DrR1pper Registered

    Joined:
    Apr 29, 2012
    Messages:
    3,294
    Likes Received:
    36
    ROFL! :D

    You about page doesn't show it but are you using 2x GTX670's in sli?
     
  11. Saabjock

    Saabjock Registered

    Joined:
    Nov 13, 2011
    Messages:
    906
    Likes Received:
    19
    No Ripper...just one. :D
     
  12. alpha-bravo

    alpha-bravo Registered

    Joined:
    Dec 22, 2012
    Messages:
    1,374
    Likes Received:
    31
    Ok did some research about the topic. The PCIe 3.0 bus has some really good improvements like less protocol overhead, a higher clock rate specification and per specification a doubled bandwidth.
    Further research in some OC Forums confirm that most of the people report a performance boost of around 10% some more some less.

    So yes activating 3.0 if it's available on gpu and systemboard is a good thing.

    But what I meant with my previous post was that I'm not sure if the count of lanes (not the bus specification) bring per se a performance gain.

    Edit: Forgot to say: Thx good observation and important to know the impact. :)
     
    Last edited by a moderator: Oct 12, 2014
  13. DrR1pper

    DrR1pper Registered

    Joined:
    Apr 29, 2012
    Messages:
    3,294
    Likes Received:
    36
    If i understand you correctly bravo, it should bring more performance with more lanes because PCI-e 2.0 x16 should equal PCI-e 3.0 x8 in graphics card performance.

    Unfortunately we've not had the chance to test to see if this is 100% true yet because no-one with SLI cards has done a test of PCI-e 2.0 x16/x16 vs PCI-e 3.0 x8/x8 with their system in rf2. And I can't figure out how to force my PCI-e 3.0 lane into x8 mode to test myself. Anyone with a PCI-e 3.0 x16 lane cpu with sli cards can perform this test. Ari is one of them.

    But regarding PCI-e 3.0 x16 vs PCI-e 2.0 x16, we've seen increases in performance on the highest end cards as much as 35% or more (if i recall correctly). It lessens with less powerful cards, for instance my GTX 970 by 16-18% and my GTX 770 improved by only 6-9%. I reckon the cut off point is somewhere around a GTX 660 but it's only a guesstimate.
     
    Last edited by a moderator: Oct 12, 2014
  14. alpha-bravo

    alpha-bravo Registered

    Joined:
    Dec 22, 2012
    Messages:
    1,374
    Likes Received:
    31
    Yes but :D (sry for my bad explanations) what I mean is for example a system with 16 Lanes and PCIe 3.0 (in this case I have this specs in my system).
    Concrete example: 2 x PCIe 3.0 capable GPU in a systemboard with PCIe 3.0 slots but this slots work in native x8 mode. What I would like to know is if the gfx performance would take advantage if the slots would working in x16 mode and if yes how big the impact would be and what are the pegged (hope this is correct comes from translator) parameters that such a system can scale better.

    I can not test this because my chipset support native only x16. With the PLX it can provide 40 lanes (with some overhead for the cpu) and I think I can mix native x16 with PLX x16 but I remember that then the PLX was running hot (around 80 degrees).
    Also during my research about this topic I saw posts that NVIDIA SLI take more advantage from this specs settings as the AMD 79xx series (not the R2xx series).

    Example in other words:
    If you have a car with 400 PS (40 lanes) but due the weather and road conditions (system and software environment) you can take only on a small part full advantage of the power how much faster are you ?
     
  15. DrR1pper

    DrR1pper Registered

    Joined:
    Apr 29, 2012
    Messages:
    3,294
    Likes Received:
    36
    Right, gotcha. The only test that has been done with working SLI in 2d mode (SimonV6's GTX 970's for example) have shown the same beneficial increase in performance from going PCI-e 2.0 x16 to PCI-e 3.0 x16 on both single and SLI mode. No one has done a PCI-e 3.0 x8/x8 vs PCI-e 3.0 x16/x16 comparison because i think if you're able to use full x16/x16 (which requires you have a cpu with at least 32 PCI-e 3.0 lanes available), the motherboard can't be manually set to x8/x8. Equally if you have a single card, unless unless you have a specific pci-e lane that is only x8 then you can't force into into x8 mode. That is however just a theory of mine since i can't seem to manually downgrade mine from x16 to x8 and i also can't find anywhere showing it's possible on the internet.

    Now the closest thing we have to PCI-e 3.0 x8 comparisons are the PCI-e 2.0 x16 results which should be identical to PCI-e 3.0 x8 results. And if so, then based on what we've seen, the performance gains from 2.0 x16 (or 3.0 x8) to 3.0 x16 is the same for both single and sli mode.
     
  16. Spinelli

    Spinelli Banned

    Joined:
    Jan 28, 2012
    Messages:
    5,290
    Likes Received:
    32
    Yes, 3.0 @ 8x has the same bandwidth as 2.0 @ 16x, but 3.0 - even at equivalent bandwidths - still holds an advantage due to less overhead and such as alpha bravo explained. So technically it's not an absolute 100% apples-to -apples comparison.

    Not a single person here has SLI on an LGA 1155/1150 CPU (excluding Sandy Bridge) with 3D vision?...

    There are tons and tons of PCI-E benchmarks online; try comparing the 3.0 @ 8x vs the 2.0 @ 16x. Then again, rFactor 2 - (and - I keep on forgetting to mention - possibly all other ISI-engine based games like Game Stock Car, GTR 2, Raceroom, etc.) seems to show wildy different results compared to just about every other game, therefore we need to test rF2 itself in order to make a proper conclusion regarding this matter.
     
    Last edited by a moderator: Oct 12, 2014
  17. DrR1pper

    DrR1pper Registered

    Joined:
    Apr 29, 2012
    Messages:
    3,294
    Likes Received:
    36
    Yep and Prodigy in the live benchmark thread has an SLI PCI-e 3.0 x8/x8 setup with what i suspect to be GTX 970's. When he replies and if it turns out to be true, we can check if at least for rf2 whether 3.0 x8 is equal to 2.0 x16 exactly or not.
     
  18. Spinelli

    Spinelli Banned

    Joined:
    Jan 28, 2012
    Messages:
    5,290
    Likes Received:
    32
    Great. How would we check this?...I guess he should do the test in single GPU mode.

    First test - simply disable SLI in the Nvidia control panel, (takes 5 seconds) leaving only 1 GPU working, and working at 3.0 8x.

    Second test - physically remove the second GPU which would give the first GPU access to all 16 lanes. Then go into BIOS and set PCI-E to 2.0.

    This wil give us a proper 3.0@8x VS 2.0@16x test.
     
  19. DrR1pper

    DrR1pper Registered

    Joined:
    Apr 29, 2012
    Messages:
    3,294
    Likes Received:
    36
    I don't think you can manually force you card to run in x8 mode if x16 is available or else you could have done the test.

    All he'd need to do is run the benchmark in sli mode on his cpu and mobo with pci-e 3.0 mode which will automatically be in x8/x8 mode (as he only has 16 lanes of 3.0 available to him). Then go into the bios and set pci-e gen to gen 2 mode and repeat the benchmark with sli enabled. His cards will be in pci-e 2.0 x16/x16 mode automatically. Then just compare results. If the same, then 3.0 x8 = 2.0 x16 performance exactly in rf2.
     
    Last edited by a moderator: Oct 13, 2014
  20. Spinelli

    Spinelli Banned

    Joined:
    Jan 28, 2012
    Messages:
    5,290
    Likes Received:
    32
    No, his cards will be in PCI-E 2.0 8x/8x mode then. He only has 16 lanes, doesn't matter if it's 3.0 or 2.0.

    Enabling PCI-E 2.0 on my 4930k doesn't turn my 16x/16x 3.0 setup into 32x/32x, it's still 16x/16x, just in 2.0 mode.

    There are only 16 lanes (or 40 for some others) it doesn't matter if those lanes are running in 1.0, 2.0, 3.0 or whatever mode, the amount of lanes still totals 16 (or 40 in some people's cases).
     
    Last edited by a moderator: Oct 13, 2014

Share This Page