PhysX SLI Broken?

Discussion in 'Videocards vs General Purpose - NVIDIA Ageia PhysX' started by mazty, May 22, 2011.

  1. mazty

    mazty New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    NGTX560ti Twin Frozer II
    I've just installed a GTX 560 Ti OC in my PC, replacing a GT 240 OC. As my motherboard is SLI capable (x8 x8) I thought I'd use the GT240 as a dedicated PhysX card instead of letting it gather dust in a box somewhere.

    However after running a few benchmarks, I wonder if SLI Physx is broken or just really badly written...:

    - On the Dark Void PhysX benchmark, other than only being able to select the low PhysX setting, I get a Min/Avg FPS result of 59.91/78.91 with the GTX560 doing everything, and the best I can get in PhysX SLI, when pushing the OC of the GT240, is a Min/Avg FPS result of 60.05/74.80.

    - In the Darkest of Days benchmark, the single GPU VS SLI set up results in an average result of 38.8 vs 30 FPS. Also overclocking does not affect the result. Considering I'm increasing the core by 30hz, shader by 100hz and memory by 80hz, and these results give me a few FPS better in the Dark Void benchmark...what the hell? :3eyes:

    - When I managed to run Kombustor it give me just slightly better PhysX results when in SLI, but degraded graphics. Since then Kombustor and Fluidmark always crash just before the results screen.

    - The GT240 isn't ever being used more than 45% by the Dark Void or Darkest of Days demo. If it was being used 100% and producing less FPS than the GTX560 when in SLI, I'd understand if it was considering not powerful enough, but it's not even being used at 50% load. Is this an issue with architecture or poor optimisation of PhysX when in SLI?

    - I am however running it all off a 500W PSU...But with the GT240 taking no more than 75W, and the average full system load of the MSI GTX560 being 310~, surely I should have a headway of ~100 Watts?

    Anyone else had this issue or suggest better benchmarks to run? Any reply greatly appreciated :)
     
  2. Xzeonx

    Xzeonx New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    2x GTS 250 1.0GB SLI
    I've actually had a similar issue when trying to use a dedicated Physx Card. If the Physx Card is too weak, it can actually cause a slow down when the Physx effects are on.

    One of the things you should look at is in the Nvidia control panel, if you set the Physx to auto select and it defaults to the GTX 560 Ti, then even Nvidia thinks that it would be better handling both the Rendering and the Physx Processing.

    Also if you go to Tom's Hardware and look at the Graphics Card Hierarchy Charts at the end of each months Best Graphics for the Money articles you'll see that the GT 240 GDDR5 is 6 tiers below the GTX 560 Ti. I've found that the ideal balance (depending on how high you turn up the Physx and how the game implements it) is to have a Physx Card that is either 3 or 4 tiers below the Primary Display Card.In your case to run Physx well on a separate card, I'd suggest a 9800 GTX or GTS 250 if you can find one. Also the 512MB vs 1.0GB models for physx don't make too much difference unless you're gaming at High Res with Physx at maximum.
     
  3. mazty

    mazty New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    NGTX560ti Twin Frozer II
    Thanks for the reply.
    The Nvidia control panel selects the GT240 when on autoselect for PhysX which makes it even more confusing. Also PhysX only seems to use around 40% of the GT240 at any one time. If that is 40% is due to architectural limits of how much of the card can process PhysX, fair enough, if not, surely a better card would just be compensating for something that seems horribly inefficient?

    If anyone can share the loads on their SLI PhysX cards when running PhysX benchmarks/intensive scenes etc, that'd be great.

    UPDATE: Just ran Cryostatis and got a higher avg and max FPS and same Min FPS when in PhysX SLI when compared to just running PhysX on the GTX560.
    I think it looks like some games simply aren't made/optimised for PhysX SLI. The Unreal Engine and a few others seem good, but Mafia II, Darkest of Days, Dark Void etc seem poorly optimised, if at all, for dealing with PhysX SLI, unless given a card which should be considered as overkill...
     
    Last edited: May 22, 2011
  4. Xzeonx

    Xzeonx New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    2x GTS 250 1.0GB SLI
    It really does depend on how the game engine implements the Physx API, some games have more that can be offloaded to the Physx Card than others, if the game engine just doesn't have enough Physx data to crunch, then you're not going to see high loads.

    I've also seen on one or two games where the load on the PPU drops in the really intensive scenes (I think one of them was Batman Arkham Asylum) which almost makes me think that there might be some driver bug to work out with getting the data to and from the PPU Card, as all the data has to be sent First from the Main memory to the GT 240 then back to the main memory, and then back to the GTX 560 Ti to have the scene rendered for each frame.

    From the Dark Void Benchmark it sounds like your system is working as it should. Having a Dedicated PPU is not really supposed to increase your frame rates, it's really only supposed to help you turn the Physx eye-candy up higher without taking as much of a hit in FPS.
     

  5. teleguy

    teleguy Maha Guru

    Messages:
    1,337
    Likes Received:
    160
    GPU:
    GTX 1070/Vega 56
    The demo is limited to low Physx.


    Besides if you install the GT 240 you halve the bandwidth for the 560. This might be the reason for lower FPS.
     
    Last edited: May 23, 2011
  6. sykozis

    sykozis Ancient Guru

    Messages:
    21,798
    Likes Received:
    1,056
    GPU:
    MSI RX5700
    1 Correction here....you're not running any form of SLI at all. SLI Explicitly requires 2 graphics cards with the same graphics processor.

    You're wrong.... PCIe 2.0 x16 provides 16GB/s bandwidth for the graphics card to use. PCIe 2.0 x8 provides 8GB/s bandwidth for the graphics card to use. Since even the GTX560 can't use 50% of the bandwidth provided by a PCIe 2.0 x16 slot, reducing the available bandwidth to 8GB/s does NOT "halve the bandwidth for the 560"....it simply halves the available bandwidth. The GTX560Ti still has plenty of bandwidth to do it's job at full speed. You're not in any way halving the utilized bandwidth, as it's less than 8GB/s. It's a simply matter of the GT240 not being able to keep up with the considerably faster GTX560Ti.

    GT240, is basically a GeForce 9600GSO with DX10.1 support and slightly slower Shader clock (1340 vs 1375 on the 9600GSO)

    GTS240 would provide better performance as it's essentially an 8800GT/9800GT and has 112 shaders as opposed to the 96 Shaders of the GT240 and 9600GSO.
     
  7. Xzeonx

    Xzeonx New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    2x GTS 250 1.0GB SLI
    I hate to nitpick here, but the GT 240 is not related to the 9600 GSO at all. The GT 240 is based off of the GT215 Core built on a 40nm process, while the 9600 GSO is based off of the the G92 Core based on either 65nm or 55nm. The GT215 core has a whole load of other differences, like it has a built in PureVideo HD VP4 processing vs the VP2 processing of the G92, it also has a sound processor built on die for audio over HDMI, and finally the Core configuration is totally different. the G92 in the 9600GSO has 128 CUDA cores with 96 Active, 64 Texture Units with 48 Active, and 16 ROPs with 12 Active. The GT215 Die only has 96 CUDA cores with 32 Texture Units, and only 8 ROPs, all of them Active.

    In Short the GT 240 is as related to the 9600 GSO only as much as the 9800 GTX is related to the 8800 GTX. Its an updated version of the same basic design, but based off a totally different core with updated features.

    (End of GT215 Rant, I really really like those cards, great for HTPCs and low power Folding cards)
     
  8. Xzeonx

    Xzeonx New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    2x GTS 250 1.0GB SLI
    Also, A minor (re)correction: When using the GTX 560 Ti at x8 instead of x16, it very well could cause a slight drop in FPS, if you read any of the articles from Tom's Hardware or the like such as: 9800 GX2 PCIe Scaling you'll see that depending on the game or benchmark going from PCIe 2.0 x16 to 2.0 x8 can have as much as a 15% hit (as in MS FSX).

    Also if he's using an SLI board, and a Phenom II CPU, then he's using either a nForce 750a, 780a, or 980a Chipset (or modded Drivers) in which case, when the Chipset is in multi GPU mode, it acts like the nForce 200 PCIe Bridge and sends the same data to both GPUs which further reduces the bandwidth available to the GTX 560 Ti because it's also getting the same raw instruction data as the GT 240.

    That being said I don't think that the 20% drop (from 38FPS to 30FPS) is right, and the OP should look at doing a clean sweep of all the drivers in Safe Mode and starting with clean settings. Every once in a while game profiles can become dorrupted and lead to poor performance.
     
  9. mazty

    mazty New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    NGTX560ti Twin Frozer II

    I think I'll do the clean sweep of drivers when I get a free moment to do so. Having the increase in FPS and >50% of the GT240 being used in Cryostatis, and yet a 20% drop in FPS in the other demos, and only about 40% of the GT240 being used, a driver issue could be the source. Will update when I wipe it all.
     
  10. mazty

    mazty New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    NGTX560ti Twin Frozer II
    'Back in my day' it was called PhysX SLI - not sure what the term is now for when you have one card running just PhysX and the other doing the graphics.
     

  11. jbmcmillan

    jbmcmillan Ancient Guru

    Messages:
    2,764
    Likes Received:
    276
    GPU:
    Gigabyte G1 GTX970
    That article from Tom's is from 2008 and in no way is a 560 and a 240 saturating 8x. so that shouldn't enter into the equation here.I went from a 790i with 16x,16x to P67 with 8x,8x and took no performance hit whatsoever with sli 460's.
     
  12. sykozis

    sykozis Ancient Guru

    Messages:
    21,798
    Likes Received:
    1,056
    GPU:
    MSI RX5700
    There's actually 2 9600GSO's....so, stating that it uses the G92 graphics processor isn't quite accurate as the 9600GSO 512 uses the G94 graphics processor, as does the 9600GT, but with 48 "Cuda Cores" and a 256bit memory bus.

    For Comparison:
    9600GSO: http://www.nvidia.com/object/product_geforce_9600_gso_us.html
    9600GSO 512: http://www.nvidia.com/object/product_geforce_9600_gso_512_us.html
    GT240: http://www.nvidia.com/object/product_geforce_gt_240_us.html

    nVidia has only supported PhysX since 2008, and at no point was the method for using 1 card for graphics and 1 card for physx ever considered "PhysX SLI" as this was possible prior to nVidia's buyout of Ageia, with their dedicated PhysX Processing Unit (PPU for short). The Ageia PPU allowed for hardware level processing of PhysX regardless of the primary graphics card in use.

    SLI, or Scalable Link Interface, is a method for connecting 2 graphics cards to split the graphics load.
    Source: http://en.wikipedia.org/wiki/Scalable_Link_Interface
     
    Last edited: May 25, 2011
  13. Xzeonx

    Xzeonx New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    2x GTS 250 1.0GB SLI
    That was only the first example that came up in Google, they've since done the same comparison with a GTX 570. And even if they didn't the GTX 560 being at least on par with, if not faster than the 9800 GX2, it would be more susceptible to losing bandwidth. I'll post the other article when I can find it again. Second the reason for there not being a hit when switching to the Intel platform is the way that the PCIe traffic is actually being handled. On the Nvidia chipsets (750a in this case) it has the core logic of the nForce 200 Bridge built into the chipset so it acts just like it would on any other platform with an nForce 200 Chip, it will send the same data to both GPUs.

    It the difference between sending Data pack A on the first set of 8 Lanes and sending Data B on the second set of 8 Lanes on the Intel platform, while on the Nvidia 750a it is sending both Data A and B on both sets of 8 lanes. Just like any other chipset with an nForce 200 Bridge sends the same data to both GPUs. Now newer Drivers have tried to work around this because it presents problems when you're trying to run anything other than a GPU in the second PCIe x16 slot, just look up a bunch of the forum posts about problems with the earlier SLI boards and RAID Cards.

    And onto the note of the term Physx SLI, its true thats its not correct to call it Physx SLI, but when the technology was being released and covered by the review sites, there were a few that called using a GPU as a PPU Physx SLI. Example: Firingsquad Physx Overview But again, thats just nitpicking.
     
    Last edited: May 25, 2011
  14. sykozis

    sykozis Ancient Guru

    Messages:
    21,798
    Likes Received:
    1,056
    GPU:
    MSI RX5700
    If you look, Toms Hardware did an article using the GTX480 and came to the conclusion that PCIe 2.0 x8 does not hinder the performance of that card, which outperforms the GTX560.
     
  15. teleguy

    teleguy Maha Guru

    Messages:
    1,337
    Likes Received:
    160
    GPU:
    GTX 1070/Vega 56
    Which one? The only one I can find primarily focuses on SLI performance and has this to say:

     

  16. sykozis

    sykozis Ancient Guru

    Messages:
    21,798
    Likes Received:
    1,056
    GPU:
    MSI RX5700
  17. teleguy

    teleguy Maha Guru

    Messages:
    1,337
    Likes Received:
    160
    GPU:
    GTX 1070/Vega 56
    No, that article comes to exactly the same conclusion: In a SLI Setup you won't notice much of a difference. Unlike tomshardware they didn't even test single cards.
     
  18. Xzeonx

    Xzeonx New Member

    Messages:
    6
    Likes Received:
    0
    GPU:
    2x GTS 250 1.0GB SLI
    True, but another thing to keep in mind when comparing those setups to the OPs is that the true SLI configurations have the SLI bridges to use for the inter-card communication, thereby removing the need to use the PCIe Pathways for card to card communication, when using a mixed GPU setup, that option is not possible, so comparing an SLI config at x8/x8 to a Primary GPU + PPU combo at x8/x8 is almost an apples to oranges comparison because the x8/x8 SLI configs have the additional bandwidth of the SLI bridge available for use. (I believe that the original spec for the SLI communication was 1GB/s which is equivalent to an extra 4 PCIe lanes available)

    When you consider that those 4 lanes can take up to 4 lanes up and downstream load off of the main PCIe connectors for the inter-card communications , it really doesn't seem like quite a fair fight to compare them, almost as if you'd be comparing it to 20 lanes between the two cards instead of just 16.

    What would be a little more accurate and relevant to this thread, and I will try to set up some benchmarks for bandwidth would be to test the primary GPU + PPU vs two Matching GPUs with SLI disabled to get a more accurate picture of how the limited PCIe traffic would affect the OP setup.
     
  19. StatykiVi

    StatykiVi Master Guru

    Messages:
    209
    Likes Received:
    0
    GPU:
    eVGA GTX 570 SC
     
  20. elkosith

    elkosith Maha Guru

    Messages:
    1,421
    Likes Received:
    15
    GPU:
    AsRock RX 570
    Same here, GTX 460 SE with 9600GT. The 9600GT somehow limited to 45% - 50% usage. 9600GT as dedicated Phsyx didn't give me any performance increment, so I pulled it out from the slot and gave it to my nephew
     

Share This Page