madVR - high quality video renderer (GPU assisted)

[*MidnightWatcher*]

The question for me now is this -- Should I pick up a cheap used RTX 2080 or spend more and buy an RTX 3070? What settings would an RTX 2080 not be able to do that the 3070 would?

[Alexkral]

What settings would an RTX 2080 not be able to do that the 3070 would?

The first thing I always think of when I see these kinds of questions is frame interpolation. In addition to be partly done on the GPU, it is not the same to process with madVR at 24 fps than at 60 or even more.

[chros]

The question for me now is this -- Should I pick up a cheap used RTX 2080 or spend more and buy an RTX 3070? What settings would an RTX 2080 not be able to do that the 3070 would?
Which GPU do you have now?
What do you need it for? Gaming? then yes. Do you need the new inventions, HDMI 2.1/AV1? yes, again.
Would you only use it for madvr? then good question. Do you need RTX for it at all?! A used Pascal 1080/Ti would be cheaper :)
There's no point "future proofing" unless you need those features right now.

The first thing I always think of when I see these kinds of questions is frame interpolation. In addition to be partly done on the GPU, it is not the same to process with madVR at 24 fps than at 60 or even more.
Yes, but personally I don't need this.

[SamuriHL]

If you're upscaling lower resolutions at 60fps then faster speeds are warranted. How much so depends on far too many factors for anyone to realistically answer. I would also not buy an RTX card banking on the fact that madshi is going to be building new tensor core algorithms. He absolutely IS building new tensor core algorithms. However, we have no guarantees at all that they will ever see the light of day in madvr on a PC. They may, but, that's not a selling point of a new card. So right now, all that matters is raw speed. I personally use copyback D3D because I want menus when I stick a disc in the machine. The faster memory speeds on the new 30xx series may or may not improve things on that front. It remains to be seen if the 2xfp32 will have a real impact on what can be done in madvr, as well. These are all unknowns until we get the cards in our hands. But ultimately what card is right for you depends on your needs/wants/desires and what content you will be viewing and on what type of screen. I'm running a 2070 right now which does 99% of what I want. If I wasn't also into gaming and looking to avoid picking up a new console, I might not even bother chasing after the use case I'm looking at solving with a new card and just stick with the 2070. But I weighed the options for myself, and I see the 30xx cards as being a very useful upgrade overall. YMMV

[huhn]

No, thats hardly the same thing, because SMT doesn't actually mean it can run twice the instructions. Its an apples to oranges comparison entirely.
The CUDA core layout was reportedly changed so that it actually has hardware to run two FP32 computations, instead of just one.

then doubling the number of available cores and saying a "core" can do FP32x2 is easily misleading. because it really looks like it can do FP32+FP32 or FP32+INT32 where turing was always able to do FP32+INT32 (INT 32 where independent cores) this is a great update if true.

they have double the CUDA core per SU or they can do FP32x2 i have a hard time believing they can do both at the same time.

[*MidnightWatcher*]

Which GPU do you have now?
What do you need it for? Gaming? then yes. Do you need the new inventions, HDMI 2.1/AV1? yes, again.
Would you only use it for madvr? then good question. Do you need RTX for it at all?! A used Pascal 1080/Ti would be cheaper :)
There's no point "future proofing" unless you need those features right now.
No gaming, using madVR only. My current GPU is a 1060 6GB and it does about 75% of what I want. If I upgrade I'd go for an RTX model for the tensor cores, just in case madshi ever (fingers crossed) released an update that took advantage of it somehow.

[Stereodude]

I have a 1080 non-Ti and I can barely play back 2160p60 to a UHD TV without dropping frames using madVR. I had to dial things way back to like bicubic chroma upscaling and most other stuff turned off. I wouldn't consider a Pascal card if that's what you want to do.

Edit: I should add that I couldn't find settings to let me play back HDR 2160p60 without dropped frames if it doesn't have black bars. SDR is okay.

[jkauff]

I have a 1080 non-Ti and I can barely play back 2160p60 to a UHD TV without dropping frames using madVR. I had to dial things way back to like bicubic chroma upscaling and most other stuff turned off. I wouldn't consider a Pascal card if that's what you want to do.
Hmm. I have a 1080 in my HTPC attached to my 4K HDR television, running madVR with custom optimized resolutions. I can easily upscale my Blu-ray movies with NGU High.

Is there any reason you're using 2160p60? Doesn't your TV support lower resolutions?

[VBB]

He must be talking about 60fps content, like Gemini Man, and also using tone mapping at the same time. I don't use tone mapping, but my GTX 960 can easily play such content.

[Asmodian]

they have double the CUDA core per SU or they can do FP32x2 i have a hard time believing they can do both at the same time.

I agree, I think the number of CUDA cores reported is actually the number of FP32 units, not cores that can each do two FP32 operations. So these new "cores" are weaker than the old cores because they cannot do INT32 at the same time, more of an expansion of the old INT32 units into FP32 units than really doubling everything in what they used to call a CUDA core. Of course, marketing is trying to make it sound good and in the process of explaining what they did technically they end up saying two things that are mutually exclusive.

In this case I would probably do the same thing. If you double the FP32 units you have to double the reported core count, everyone knows each core is a FP32 unit. But then architecturally the way you got double the FP32 units did not double everything traditionally included in a CUDA core. With no good way to define GPU "cores" in the first place, and consumers not getting excited by architectural details, Nvidia had no option but to be confusing... and they picked the confusion that made them look better.

But this is not like SMT either, there really is twice the hardware we traditionally cared about in a GPU core. It is not a doubling of the front end to better utilize idle hardware or similar.

[Crimson Wolf]

This from Q/A on reddit


[Tony Tamasi] One of the key design goals for the Ampere 30-series SM was to achieve twice the throughput for FP32 operations compared to the Turing SM. To accomplish this goal, the Ampere SM includes new datapath designs for FP32 and INT32 operations. One datapath in each partition consists of 16 FP32 CUDA Cores capable of executing 16 FP32 operations per clock. Another datapath consists of both 16 FP32 CUDA Cores and 16 INT32 Cores. As a result of this new design, each Ampere SM partition is capable of executing either 32 FP32 operations per clock, or 16 FP32 and 16 INT32 operations per clock. All four SM partitions combined can execute 128 FP32 operations per clock, which is double the FP32 rate of the Turing SM, or 64 FP32 and 64 INT32 operations per clock.

Doubling the processing speed for FP32 improves performance for a number of common graphics and compute operations and algorithms. Modern shader workloads typically have a mixture of FP32 arithmetic instructions such as FFMA, floating point additions (FADD), or floating point multiplications (FMUL), combined with simpler instructions such as integer adds for addressing and fetching data, floating point compare, or min/max for processing results, etc. Performance gains will vary at the shader and application level depending on the mix of instructions. Ray tracing denoising shaders are good examples that might benefit greatly from doubling FP32 throughput.

Doubling math throughput required doubling the data paths supporting it, which is why the Ampere SM also doubled the shared memory and L1 cache performance for the SM. (128 bytes/clock per Ampere SM versus 64 bytes/clock in Turing). Total L1 bandwidth for GeForce RTX 3080 is 219 GB/sec versus 116 GB/sec for GeForce RTX 2080 Super.

Like prior NVIDIA GPUs, Ampere is composed of Graphics Processing Clusters (GPCs), Texture Processing Clusters (TPCs), Streaming Multiprocessors (SMs), Raster Operators (ROPS), and memory controllers.

The GPC is the dominant high-level hardware block with all of the key graphics processing units residing inside the GPC. Each GPC includes a dedicated Raster Engine, and now also includes two ROP partitions (each partition containing eight ROP units), which is a new feature for NVIDIA Ampere Architecture GA10x GPUs. More details on the NVIDIA Ampere architecture can be found in NVIDIA’s Ampere Architecture White Paper, which will be published in the coming days.


What's also interesting is on the encoder/decoder side, and to the person who asked what 30xx series can do thht 20xx can't? AV1 decode is one:


[Jason Paul] Hi Carmen813, for RTX 30 Series, we decided to focus improvements on the video decode side of things and added AV1 decode support. On the encode side, RTX 30 Series has the same great encoder as our RTX 20 Series GPU. We have also recently updated our NVIDIA Encoder SDK. In the coming months, livestream applications will be updating to this new version of the SDK, unlocking new performance options for streamers.


So, it looks like nvidia will be exposing some new options to the encoder side with the new SDK, even to users of the 20xx series.

[Stereodude]

Hmm. I have a 1080 in my HTPC attached to my 4K HDR television, running madVR with custom optimized resolutions. I can easily upscale my Blu-ray movies with NGU High.

Is there any reason you're using 2160p60? Doesn't your TV support lower resolutions?
I can upscale p24 Blu-rays using NGU high too. I'm talking about 2160p60 content.

He must be talking about 60fps content, like Gemini Man, and also using tone mapping at the same time. I don't use tone mapping, but my GTX 960 can easily play such content.
Yes, p60 content. No, I'm not using tone mapping. Straight HDR pass through. The render times have to be in the 13ms range or it will drop frames. That's really hard to hit for 2160p60 HDR content.

[VBB]

That's what I meant when I said my 960 plays that just fine. Just lower the chroma setting enough. Your 1080 shouldn't break a sweat.

[chros]

I'm talking about 2160p60 content. No, I'm not using tone mapping. Straight HDR pass through. The render times have to be in the 13ms range or it will drop frames. That's really hard to hit for 2160p60 HDR content.
Do you use copyback? And with which display?

That's what I meant when I said my 960 plays that just fine. Just lower the chroma setting enough. Your 1080 shouldn't break a sweat.
Yep, my 1060 6GB with d3d11-native and RGB 8 bit output can do (4k UHD Gemini man):
- madvr tonemapping hdr output @700 nits: Bicubic60 chroma, and both compromise HDR option are set
- hdr passthrough: NGU Sharp Low (!) chroma + enhance details 2.0 (image enhancement)

But this title is DV, so the problem is mute since we have to use an external device for DV :)

No gaming, using madVR only. My current GPU is a 1060 6GB and it does about 75% of what I want. If I upgrade I'd go for an RTX model for the tensor cores, just in case madshi ever (fingers crossed) released an update that took advantage of it somehow.
I have the same and it does 100% for me :) But in this case just wait for 3060 / 3050.

[chros]

they have double the CUDA core per SU or they can do FP32x2 i have a hard time believing they can do both at the same time.

I agree, I think the number of CUDA cores reported is actually the number of FP32 units, not cores that can each do two FP32 operations. So these new "cores" are weaker than the old cores because they cannot do INT32 at the same time, more of an expansion of the old INT32 units into FP32 units than really doubling everything in what they used to call a CUDA core.
...
But this is not like SMT either, there really is twice the hardware we traditionally cared about in a GPU core. It is not a doubling of the front end to better utilize idle hardware or similar.



[Tony Tamasi] One of the key design goals for the Ampere 30-series SM was to achieve twice the throughput for FP32 operations compared to the Turing SM. To accomplish this goal, the Ampere SM includes new datapath designs for FP32 and INT32 operations. One datapath in each partition consists of 16 FP32 CUDA Cores capable of executing 16 FP32 operations per clock. Another datapath consists of both 16 FP32 CUDA Cores and 16 INT32 Cores. As a result of this new design, each Ampere SM partition is capable of executing either 32 FP32 operations per clock, or 16 FP32 and 16 INT32 operations per clock. All four SM partitions combined can execute 128 FP32 operations per clock, which is double the FP32 rate of the Turing SM, or 64 FP32 and 64 INT32 operations per clock.
...
Doubling math throughput required doubling the data paths supporting it, which is why the Ampere SM also doubled the shared memory and L1 cache performance for the SM. (128 bytes/clock per Ampere SM versus 64 bytes/clock in Turing). Total L1 bandwidth for GeForce RTX 3080 is 219 GB/sec versus 116 GB/sec for GeForce RTX 2080 Super.

Like prior NVIDIA GPUs, Ampere is composed of Graphics Processing Clusters (GPCs), Texture Processing Clusters (TPCs), Streaming Multiprocessors (SMs), Raster Operators (ROPS), and memory controllers.

The GPC is the dominant high-level hardware block with all of the key graphics processing units residing inside the GPC. Each GPC includes a dedicated Raster Engine, and now also includes two ROP partitions (each partition containing eight ROP units), which is a new feature for NVIDIA Ampere Architecture GA10x GPUs.


Thanks guys, it's getting a bit clearer :)

[ryrynz]

Thanks guys, it's getting a bit clearer :)So much quote, you make mobile users cry [emoji14] You can edit down quotes, I recommend it.

[chros]

Those quotes *are* edited :)

[Stereodude]

Do you use copyback? And with which display?
Yes, using copyback and Sony 75Z9D. I get about 17ms render times using Bicubic 75 and have made just about every performance tradeoff/compromise madVR offers. Windows 8.1 and the 1080 is clocking itself up to it's limits.

[Greenhorn]

Thanks guys, it's getting a bit clearer :)

Just to provide a bit of additional context, NVIDIA SMs before Turing (or was it Pascal?) were already capable of 32 FP32 instructions per clock. Turing just traded this capability for the ability to execute 16xFP32 and 16xInt32 simultaneously.

Ampere is just restoring the old functionality (because I guess games don't really use INT32 a lot) while keeping the concurrent INT32 execution. (See page 34 of the Ampere whitepaper. It doesn't seem to mention if it's 32xFP32 and 32xINT32 or 32xFP32 and 16xINT32.)

To try and bring this back to MADVR, does it use ints for stuff or is it all floating-point?

[mclingo]

Hi, is this a bug or a lack of my understanding of the current bandwith on HDMI 2.0, my 4k 10 bit 23,976 HDR movies play fine but I have one or two 25hz 4k HDR videos which wont play in HDR (hdr triggers but picture is washed out), if I force them to play at 23hz they are ok.

I thought we were good up to 10bit 30hz for 4k HDR on HDMI 2.0?

tried two recent AMD drivers.

[SamuriHL]

To try and bring this back to MADVR, does it use ints for stuff or is it all floating-point?

Pretty sure all madvr math internally is FP. It's those rounding errors using INT that cause problems and why madvr is a better solution for rendering.

[chros]

Yes, using copyback and Sony 75Z9D. I get about 17ms render times using Bicubic 75 and have made just about every performance tradeoff/compromise madVR offers. Windows 8.1 and the 1080 is clocking itself up to it's limits.
That's why then, that's how faster (https://forum.doom9.org/showthread.php?t=176642) d3d11-native is: a 1060 becomes way faster than a 1080.
Can't you use it on Win8.1? (I forgot it) If you can:
- you don't need black bar detection with a TV!
- do you need deinterlacing, bd-menus? If not, then forget it.

If you can't: install Win10, it's way cheaper than buying another card :)

[Stereodude]

That's why then, that's how faster (https://forum.doom9.org/showthread.php?t=176642) d3d11-native is: a 1060 is way faster than a 1080.
Can't you use it on Win8.1? (I forgot it) If you can:
- you don't need balck bar detection!
- do you need deinterlacing, bd-menus? If not, then forget it.

If you can't: install Win10, it's way cheaper than buying another card :)
d3d11-native is a decoding setting for LAV filters? Or what specific option are you talking about?

I'll have to look at it later when I get home.

Win10 is free on this system since I did the upgrade back when M$ was giving it away free so the HW would be recognized in the future, then rolled back to Win 8.1.

[VBB]

Yes, you would change that in LAV. Make sure that "Hardware Device to use" is set to "Automatic (Native)".

[NoTechi]

Hi, is this a bug or a lack of my understanding of the current bandwith on HDMI 2.0, my 4k 10 bit 23,976 HDR movies play fine but I have one or two 25hz 4k HDR videos which wont play in HDR (hdr triggers but picture is washed out), if I force them to play at 23hz they are ok.

I thought we were good up to 10bit 30hz for 4k HDR on HDMI 2.0?

tried two recent AMD drivers.

Did you try the last official driver 20.4.2? I had some problems with the latest "beta" drivers when it comes to watch live tv (1080p at 50Hz) with dvbviewer but switching back to 20.4.2 with a clean install and choosing "standard" while installing I have no issues at all. It is not 4k HDR but might be worth a try.

NoTechi

[chros]

d3d11-native is a decoding setting for LAV filters?

Yes, you would change that in LAV. Make sure that "Hardware Device to use" is set to "Automatic (Native)".

As VBB said, take a look at the Lav screenshot (https://www.avsforum.com/threads/guide-building-a-4k-htpc-for-madvr.2364113/page-142#post-58165252) here. I can update that post during the weekend with my current settings if you like.

[chros]

Just to provide a bit of additional context, NVIDIA SMs before Turing (or was it Pascal?) were already capable of 32 FP32 instructions per clock. Turing just traded this capability for the ability to execute 16xFP32 and 16xInt32 simultaneously.

Ampere is just restoring the old functionality (because I guess games don't really use INT32 a lot) while keeping the concurrent INT32 execution. (See page 34 of the Ampere whitepaper. It doesn't seem to mention if it's 32xFP32 and 32xINT32 or 32xFP32 and 16xINT32.)
Interesting, now I'm getting confused again.
What do you think it means regarding to the currently stated cuda core counts?

To try and bring this back to MADVR, does it use ints for stuff or is it all floating-point?

Pretty sure all madvr math internally is FP. It's those rounding errors using INT that cause problems and why madvr is a better solution for rendering.
That's a valid question: do we have any data how the 2 generations (pascal vs turing) behaves with the same cuda core counts? E.g. 1080 vs 2070S?

[mclingo]

Did you try the last official driver 20.4.2? I had some problems with the latest "beta" drivers when it comes to watch live tv (1080p at 50Hz) with dvbviewer but switching back to 20.4.2 with a clean install and choosing "standard" while installing I have no issues at all. It is not 4k HDR but might be worth a try.

NoTechi

yeah tried few, all the same

[Yviena]

Is the HDR buttons in MadTPG broken changing the different nits values does nothing to the whitepatch when the TV/display is in HDR mode.

[Stereodude]

As VBB said, take a look at the Lav screenshot (https://www.avsforum.com/threads/guide-building-a-4k-htpc-for-madvr.2364113/page-142#post-58165252) here. I can update that post during the weekend with my current settings if you like.
That shaved about 5ms off the render time. It doesn't seem to like pausing and resuming though.

[oldpainlesskodi]

Hi, is this a bug or a lack of my understanding of the current bandwith on HDMI 2.0, my 4k 10 bit 23,976 HDR movies play fine but I have one or two 25hz 4k HDR videos which wont play in HDR (hdr triggers but picture is washed out), if I force them to play at 23hz they are ok.

I thought we were good up to 10bit 30hz for 4k HDR on HDMI 2.0?

tried two recent AMD drivers.

Can you upload a clip?

I can try on my setup.

BTW - no problem here with 60fps HDR, providing Madvr is set to 10bit.

[RXP]

Is the HDR buttons in MadTPG broken changing the different nits values does nothing to the whitepatch when the TV/display is in HDR mode.

Yep - reported to the bug tracker. Really frustrating because I can't do a 3D LUT with HDR mode on.

[chros]

That shaved about 5ms off the render time. It doesn't seem to like pausing and resuming though.
Good, take a look at the general settings (https://www.avsforum.com/attachments/088-capture_089_06082019_183624-png.2599674/) of madvr, it should work. If it doesn't then maybe you need to upgrade to Win10.

Interesting, now I'm getting confused again.
What do you think it means regarding to the currently stated cuda core counts?

That's a valid question: do we have any data how the 2 generations (pascal vs turing) behaves with the same cuda core counts? E.g. 1080 vs 2070S?
One more thing popped into my mind: does this new architectural change need applications (e.g. madvr) to be rewritten or will it work automatically (by the drivers)?
Because if it needs to be rewritten then we will be screwed, and only the half the claimed cuda cores will be used :)

[huhn]

programs like games or madVR most likely don't even know about this.
most stuff is in the hand of the GPU driver anyway and if it get's FP32 shader it will be able to use all of it's shaders for this task.

[chros]

That's good news if that's the case.

[CZ Eddie]

Where can I find a collection of MKV clips or something for various torture testing so I can eek out MadVR performance?
Like, a video for testing motion blur that is 480p, 720p, 1080p and 4k.
Or another video for testing background noise.
And other for testing detail.
Etc.

[jkauff]

Why the desire to push madVR to the max? IMO, you should be using your eyes instead of benchmarking.

Play a 1080p movie you're very familiar with. Set madVR to use NGU Sharp - high for luma doubling (for example), then play it again using NGU Sharp - very high. Can you see the difference? If not, why waste GPU resources on a level of algorithm quality you can't perceive? Use other movies from your collection to test 480p, 720p, etc.

Trust your eyes. You're just looking for a baseline anyway, since the source quality of each movie will dictate the minor adjustments you'll want to make at playback.

[ryrynz]

As above. Use the content you like to watch and test with that.

[mclingo]

Can you upload a clip?

I can try on my setup.

BTW - no problem here with 60fps HDR, providing Madvr is set to 10bit.


https://1drv.ms/u/s!AgvFafeelEBilI4RhSmODSr2EwjaMg?e=KWAhVE

here is the clip, 4k HDR @ 25fps, it plays and HDR is activated but is washed out.

However, no idea how you've managed to get HDR 60hz @ 4k as HDMI 2.0 definately doesnt have the bandwidth for this, you must running this at 8 bit.. but AMD cards cant do HDR at 8 bit, are you sure you have send output to HDR ticked in latest MADVR, sure you're not just tone mapping to SDR?

[oldpainlesskodi]

Plays fine here (at 50hz). AMD will switch into HDR @ 60hz 8bit, provided Madvr is set at 10bit or higher and not auto. Tested with the latest beta (113) works fine with passthrough. I have acutally gone back to madVR 92.09, as I want Nnedi3.

Just noticed you are not running FSE - I HAVE to use FSE for the TV to kick into HDR mode, otherwise, Madvr reports that it's using AMDHDR, BUT, the TV is in SDR mode, and looks washed out.

Don't know if that helps.

[huhn]

maybe you should try it at 50/25 HZ.

[DMU]

...you must running this at 8 bit.. but AMD cards cant do HDR at 8 bit, are you sure you have send output to HDR ticked in latest MADVR, sure you're not just tone mapping to SDR?

Set 8 bit output in driver settings and 10 bit in madVR settings.
Or 10bit@25Hz in driver settings and 10 bit in madVR settings.

[mclingo]

you cant get HDR in MADVR on AMD card without setting 10 bit in MADVR so thats already set, I already have the driver set to 8, i'll try setting 10 bit driver and then 50hz.


I thought 30hz was max for 2160p HDR 10 bit output though on HDMI 2.0?, how are you all getting more than that?

[mclingo]

none of that works, my desktop is set to 2160p 4:4:4 full RGB 8 bit, MADVR is set to 10 bit, FSE if on, is did work when I set my desktop to 10 bit YBCR 422, so I get those with this working are also running their desktops in 10 bit 422?

I should be able to get 2160p 10 bit HDR up to 30hz though surely with a fully 18gps video chain?

[huhn]

newest test build ?

[mclingo]

not sure is there anything newer than 113?

[huhn]

try again with the normal release version.

this version blocks 10 bit output if the GPU is 8 bit in case of AMD madVR shouldn't known but worth a try.

[mclingo]

no joy with that mate, its an odd one this, however there isnt much 25 fps HDR out there so i'm not too bothered to be honest, I can just drop to tone mapping to SDR when I come up against this or create a profile, its just really odd its behaving this way on my system where other peoples are ok... unless they are all running their driver at 422 which I dont like doing.

[CZ Eddie]

Would it be expecting too much to think that live OTA 1080i television can be upscaled to 4K using the same extreme settings you would for 1080i MP4/MKV?

[el Filou]

Container/transport shouldn't matter for rendering.