Any SLI owners here use ShadowPlay/Share?

Discussion in 'Videocards - NVIDIA GeForce Drivers Section' started by shaqobe, Jul 19, 2017.

  1. shaqobe

    shaqobe Guest

    Messages:
    5
    Likes Received:
    0
    GPU:
    nvidia
    I want to know if just having ShadowPlay/Share switched on, not necessarily recording, causes a significant (20-30%) drop in GPU usage and FPS for you in fullscreen DX11 games. Again, this is only for SLI.

    Thanks.
     
  2. archie123

    archie123 Guest

    Messages:
    2,475
    Likes Received:
    17
    GPU:
    Aorus 1080ti Extreme
    I had performance issues with shadowplay in certain games both with 980ti in sli and now with a single card , but never anywhere near 30%! Do you have the creators edition of windows 10? Have you turned OFF the gaming recording and "optimisations"?
     
  3. shaqobe

    shaqobe Guest

    Messages:
    5
    Likes Received:
    0
    GPU:
    nvidia
    I'm using Windows 7. I have no ShadowPlay performance issues in single GPU, or in non-DX11 games or windowed games.
     
  4. Bloodred217

    Bloodred217 Master Guru

    Messages:
    356
    Likes Received:
    1
    GPU:
    2x GTX 1080 G1 Gaming 8GB
    I never really use Shadowplay/Share, but I did a quick test using the TWWH benchmark at 4K, pretty much maxed settings:

    Share disabled: 76.8FPS, 99%/95-98% GPU load
    Share enabled, Instant Replay OFF, NOT recording: 77.0FPS, 99%/95-98% GPU load
    Share enabled, Instant Replay ON, NOT recording: 71.3FPS, 99%/95-98% GPU load
    Share enabled, Instant Replay ON, recording: 71.1FPS, 99%/95-99% GPU load

    So there is a small performance drop, but nothing close to 20-30% on my system. I was recording 4K 60FPS at 100Mbps and the video was being stored on my NAS over a 1Gbit network, so not even locally on a fast SSD or anything. Maybe there's something wrong software-wise on your end, or are you perhaps saving high-bitrate video on the same drive you're running the game from or on a slow drive in general? No idea what Share does if the drive isn't fast enough to write, but if the game can't load assets quickly enough because the drive is busy I can definitely see hitching with GPU load drops popping up.
     

  5. shaqobe

    shaqobe Guest

    Messages:
    5
    Likes Received:
    0
    GPU:
    nvidia
    Hmmm. Thanks very much for checking. The directory for my ShadowPlay recordings and temporary files are on a different HDD than the SSD my games are installed on.

    I've done some research and there's a lot of conflicting information out there from SLI users on ShadowPlay's performance impact. Very strange that only a very specific category--DX11 games in exclusive fullscreen--are affected for me. If it's indeed something in the software environment, I will start process of elimination disabling things one-by-one, or maybe reformatting and installing only Nvidia driver+GFE, and see if that fixes it.

    Just curious, do you happen to have MSI Afterburner/RTSS running? I wonder if that has anything to do with it.

    Also, would you mind telling me your driver and GFE versions (if not the latest)?
     
    Last edited: Jul 20, 2017
  6. fantaskarsef

    fantaskarsef Ancient Guru

    Messages:
    15,636
    Likes Received:
    9,512
    GPU:
    4090@H2O
    Well the time I had 980 SLI I was not experiencing serious performance drops, besides a bit of a millisecond "freeze" when I recorded the shadowplay buffer to record afterwards. But certainly I never noticed a 30% performance hit when using it to record all the time. Used it for Overwatch mostly, if that's important.
     
  7. Bloodred217

    Bloodred217 Master Guru

    Messages:
    356
    Likes Received:
    1
    GPU:
    2x GTX 1080 G1 Gaming 8GB
    I do have MSI AB and RTSS running, though I don't have the OSD enabled. I'm running 384.80 and GFE 3.7.0.81, I update pretty recently, should still be latest I imagine.

    Do you get performance drops with both Instant Replay and actual recording disabled? If you do it's probably not due to the capture/recording itself, maybe something doesn't play nice with the Share OSD, like some other program hooking into your games? Could be some specific issue with how different hooks interact in DX11, but maybe they're fine in other APIs.

    In terms of diagnosing software issues, I'd just make a backup image of the OS partition, then do a fresh install and see how it behaves. You can then restore from the image, the whole process would probably take like an hour.
     
  8. shaqobe

    shaqobe Guest

    Messages:
    5
    Likes Received:
    0
    GPU:
    nvidia
    Yeah I get the performance drop just from having Share switched on; Instant Replay is disabled and not manually recording. On hindsight, I was a little off on the performance impact I stated previously. My GPU usage drop is up to 30%, but the FPS drop is 10-20% depending on application. Basically, the performance drop you saw in TW:WH from having Instant Replay and/or manually recording on? That's what I get as long as Share is enabled. It's behaving as if Instant Replay or manual recording is enabled at all times, even when I disable them.

    I closed AB and RTSS as well as Steam and Origin overlays, anything that hooks AFAIK, and it made no difference. Share overlay and overlay icons are also disabled. Tried disabling Aero in Windows 7 as well, nada. I messaged a couple of W10 SLI users who have the same problem I do so it's not the OS version. And yeah it's only in DX11 fullscreen, no issues in DX11 windowed/borderless or in games using OGL 1-4 and DX9. I don't have any DX10 games to test.

    I'm using SLI in x8/x8 and HWiNFO 'GPU bus usage' (I'm assuming this is PCIe) hasn't crossed 25%, so I ruled out insufficient PCIe BW as the culprit. My screen is only 1080p. I don't think it's a CPU/RAM limitation either as Share doesn't impact my CPU utilizations at all.

    I restored a clean system image from right after I installed Windows (before drivers and software) and applied the latest GeForce driver and GFE, but it made no difference. Tried 368.69 and GFE v2.11.4.0 (that last version before GFE v3 renamed ShadowPlay to Share) as well.

    So I am totally stumped. I feel like I've tried everything but see no rhyme or reason why this should be happening and am still no closer to arriving at a solution.
     
  9. Bloodred217

    Bloodred217 Master Guru

    Messages:
    356
    Likes Received:
    1
    GPU:
    2x GTX 1080 G1 Gaming 8GB
    This is the only thing that stands out for me, I'm running x16/x16. What resolution are you playing at, 4K? I've tested x8/x8 and x16/x16 on pretty much identical systems (different mobo) and I've seen significant difference in SLI performance in certain games. It only generally becomes something worth considering at 4K (or higher). I upgraded from a typical Z87 x8/x8 board to one with x16/x16 (with a PLX PCIe bridge) and I've seen pretty significant differences in certain games, for instance I gained like 10FPS/15% in Witcher 3, DOOM wasn't even really scaling at all and was also stuttering/micro-freezing occasionally with maxed shadows.

    Are you getting 25% bus usage with Share disabled too, or is it lower? Also in what game are you getting that? It's actually pretty high compared to what I usually see, though I'm on x16/x16, but it might be limiting SLI scaling/performance despite the number looking low. For instance in TW3 AA is bandwidth intensive - I get 2-5% bus usage with AA off, 12-15% with it on and SLI scaling is measurably worse. On x8/x8 with AA it's even worse than x16/x16, so it definitely hurts performance despite that number not being close to 100%. The thing is, I can't see why just an overlay would increase PCIe load to any significant margin, doesn't make sense to me.

    This kind of makes sense I think, I'm only speculating but for the bus to show 100% usage, it would have to be under 100% load 100% of the time, but would it be used 100% of the time? Probably not, a GPU probably spends some time getting data from the other one (PCIe bus at 100% load), but once the data transfer is complete it uses it to actually render the next frame (PCIe bus low load). If this were to happen you could very well see "low" bus load, but the GPUs are still actually spending time just waiting on bus transfers. Basically the bus could have relatively short bursts of max load, then longer periods of low load. That could definitely affect performance while the bus load reading we see would just show some average of both the max load peaks and the long inactive periods. If the max load peaks took less time, the GPUs might perform better because they're waiting less. This is all hypothetical since I don't know how SLI actually transfers data (or what exactly goes over PCIe and what goes over the bridge), but I hope you understand what I'm going for and why PCIe bandwidth can be important.
     
  10. shaqobe

    shaqobe Guest

    Messages:
    5
    Likes Received:
    0
    GPU:
    nvidia
    Like I said, my screen resolution is only 1080p.

    From what I've observed, bus usage is directly related to frame rate. Higher FPS = higher % bus usage.

    I've also observed that the relative performance impact of Share is greater at higher frame rates, and lower at lower frame rates. However, if anything bus usage is lower with Share enabled (due to the lower FPS).

    I ran some tests in BF4, Metro 2033 Redux, and Unigine Heaven:

    BF4 Test Range / 100% Res Scale / Share Off: 134.3 FPS, 92%/99% GPU usage, 29%/24% bus usage
    BF4 Test Range / 100% Res Scale / Share On: 113.1 FPS, 82%/91% GPU usage, 23%/23% bus usage

    BF4 Test Range / 200% Res Scale / Share Off: 51.8 FPS, 98%/100% GPU usage, 13%/9% bus usage
    BF4 Test Range / 200% Res Scale / Share On: 48.1 FPS, 93%/96% GPU usage, 11%/11% bus usage

    Metro 2033 Redux Main Menu / SSAA 0.5 / Share Off: 113.0 FPS, 96%/98% GPU usage, 18%/18% bus usage
    Metro 2033 Redux Main Menu / SSAA 0.5 / Share On: 79.2 FPS, 68%/75% GPU usage, 13%/12% bus usage

    Metro 2033 Redux Main Menu / SSAA Off / Share Off: 67.1 FPS, 98%/99% GPU usage, 11%/11% bus usage
    Metro 2033 Redux Main Menu / SSAA Off / Share On: 50.8 FPS, 75%/79% GPU usage, 8%/8% bus usage

    Unigine Heaven / 720p / Share Off: 29.6 FPS, 99%/100% GPU usage, 3%/3% bus usage
    Unigine Heaven / 720p / Share On: 29.4 FPS, 97%/97% GPU usage, 4%/4% bus usage.

    Unigine Heaven / 1080p / Share Off: 18.0 FPS, 99%/99% GPU usage, 4%/4% bus usage
    Unigine Heaven / 1080p / Share On: 17.5 FPS, 98%/97% GPU usage, 4%/2% bus usage

    On a side note about PCIe BW limitations in SLI, I found this DudeRandom84 YouTube channel (can't post link yet) where he tests really high-end cards like 1080 Tis and 1080s in SLI at 4K and 8K and he gets full GPU utilization even though he has only a 7700K and Z170 Deluxe non-PLX motherboard.
     
    Last edited: Jul 22, 2017

  11. Bloodred217

    Bloodred217 Master Guru

    Messages:
    356
    Likes Received:
    1
    GPU:
    2x GTX 1080 G1 Gaming 8GB
    Makes sense to have higher bus usage with higher FPS, at the very least more frames are being submitted so there has to be more CPU-GPU traffic. I don't think the OSD alone really has anything to do with PCIe bandwidth though and 1080p isn't high enough for SLI to have significant bandwidth issues AFAIK. Not sure what could be the culprit, those are some pretty huge drops you're getting.

    I tried a higher FPS game too, namely Overwatch in training mode. I got ~174FPS without Share and ~172FPS with the OSD enabled, so almost no difference, probably within the margin of error really.

    It doesn't happen in every game and GPU usage isn't really a good way to assess that it is happening in the first place. You'd need to measure SLI scaling (vs. single GPU) using different PCIe configurations, just seeing good GPU usage doesn't necessarily mean it's as good as it gets. You can also disable "well-known" bandwidth-limited graphical settings to see if they bring a large boost to SLI scaling, such as temporal AA.
     

Share This Page