Log in

View Full Version : MadVr best setting for Geforce 2070 Super


empleat
6th June 2020, 22:24
I read Madvr guide and i read to use lanczos for high end pc. However i am getting on i5 9600kf - not even maxed and geforce 2070 super only 10 ms average rendering time. I am using svp pro yet, for frame interpolation to 144fps. My gpu is not even maxed - 86% load. I am using artifact removal: banding high high, reduce compression artifacts 6 high. And trade performance for quality makes 0 difference. I need to get at least to 6.9 ms rendering time, but there are to many options and i have simply no idea what is best for my setup, there are to many hw configurations, didn't see even benchmarks anywhere. I read cuvid decoder is good, but it doesn't work anymore. I am using MPC HC 1.9.4 32bit. And i am using current DXVA 2 copy back, because it in SVP Pro guide, but other decoders work too. Tested couple there wasn't much of a difference.

Can you please advice me what to use for max quality for my setup, while keeping performance enough for 144fps ty!

Ha i found random youtube video:https://www.youtube.com/watch?v=4hyQD9-7G8c And copied setting for scaling algorithms and removed reduce compression artifacts, this was taking 6ms of rendering time. Anyway while i am asking already, if you could advice me best possible setting, i want my gpu to work hard, while i have new computer so i can enjoy it :)

EDIT: it is strange, for top quality video 10k bit rate, i have only 4-5ms rendering time, but for 3.7k bitrate, i have 11-12 ms rendering time - it is avs codec, so it seems mad vr setting doesn't matter. But more what source is playing.

Asmodian
12th June 2020, 08:46
Are your sources different resolutions? What resolution is your display? It sounds like your files are not the same codec? As you say, there are a lot of options so we need more information. :)
It would be strange to have the higher bitrate version be faster unless the slower source was not using hardware decoding for some reason, check LAV Video while playing the videos.

Debanding high should only be used on heavily banded sources, high/high is a bad default setting. I leave it off unless I notice banding in the source but low/low or low/high are also reasonable. Same with reduce compression artifacts, these algorithms do blur detail away, no need to run them all the time. Reduce compression artifacts is also pretty expensive, I would drop if if you want interpolated 144 fps. Even if you have the GPU performance I would not run it as high as 6 by default, unless most of your souses do have bad compression artifacts? :(

If you have extra GPU power using it on Jinc or NGU med chroma scaling (if your source is native luma resolution) or better luma scaling if not. If your source is not native resolution (1440p display?) then spend it on luma scaling instead, keep chroma on Bicubic60 until you find a good luma/image setting and then set chroma upscaling. Bicubic60 is already a pretty decent chroma scaling option and it is really fast. Bicubic 150 is great for fast downscaling too.
IMO:
Luma/Image: NGU > Jinc > Lanczos/Spline > Cubic 60/75
Chroma: NGU > super-xbr 100 > Jinc > Spline > Bicubic60. I don't really like lanczos for chroma and I think spline is debatable too (they are very similar), I do like bicubic for chroma if I need the speed so I cannot use Jinc or better.

CUVID is not better anymore. Personally I use DXVA2 copy back too, it has been the most reliable for me. D3D11 Native is the fastest but it only works if you don't need the CPU to access the video.

Have you tried smooth motion? It is nothing like interpolation but it does work really well at 144 Hz (do not combine with interpolation). ;)

empleat
12th June 2020, 20:56
I am on 1920x1080. Sources are usually 720, or 1080. Yea that source which had 12 ms rendering time was AVC. Otherwise majority is mpeg i think. I change gamma and than there was pixelation, so that's reason i use it. Because 255 colors is too dark, but limited range is too bright and ugly. But i don't know why higher quality video played fine, if this could cause problem in this case. Gonna test that, next time i will be playing video with high rendering time.

I am using NGU sharp for chroma.
I don't have luma under scalling algorithms.
For image downscalling i am ussing: SSIM 2d - strenght:100%, all 3 box checked, middle box relaxed, bottom box strength 150%
For image upscalling: ngu sharp: algorithm quality: 1st box very high, 2nd mad vr decide, 3rd very high
next box both mad vr decide
last box: 1st Jinc AR, 2st box SSIM 1D 100 AR Linear Light

Otherwise i know MadVr have profiles and you can switch that with shortcuts, but finding each source what codec it uses etc. and i don't know optimal setting for that anyway. It is too laborious. I want to use only 1 profile for everything. Both cpu and gpu not even maxed, so i should probably look for different setting. Gonna test that Bicubic too next time something has high rendering time. And hw acceleration is on in MPC HC - dxva copy back and in svp pro as well.

Also i used smooth motion previously, it can't match SVP Pro. SVP is much better! And deinterlacing is off too, both in MadVr and in Mpc HC.

I think i tried direct11 as well and there was no difference. I tried pretty much every decoder there was.

EDIT: seems like image upscalling was problem, when i switched to lanczos it went down 1.8ms from 10. It is pity it looked great on NGU sharp with settings above. Still weird, from that video, it was optimal setting for 4k nvidia 2070... Maybe it doesn't work just for some sources, as i was having only 5ms~ for 10k bitrate mpeg sources. Btw jinc working too and doesn't look much worse. After more tweaking, i found it was maybe, because under quadripling, it was let MadVr to decide, but every other setting i kept same. And i have 3ms now, when i set only if needed, or scalling factors is 2.0 bigger.