Evaluation of HEVC decoders (SW, Hybrid and HW)

[Asmodian]

Adaptive? Optimal Power? When I get back I home I will test too.

[NikosD]

...much worse than pascal..

only 300 MHz?? Regarding features, the driver still misses H.264 8K.
Regarding speed, VP9 8bit definitely can't much the 38 FHD simultaneously decoded streams advertised for Turing Tesla T4 :)
The decoded dropped frames are too many, but the decoding clock is too low as well.

Maybe someone could download the 8K60fps VP9 clip and try that in benchmark (full speed) mode.

The issue could be just a wrong driver's setting from nVidia.

[nevcairiel]

That youtube video runs perfect on my 2080 Ti, but my GPU never clocks that low because I have 3 screens connected (one of which is 144Hz) - which keeps idle clocks higher. Video Engine load 75%-80%~
Only one or two drops at the start of playback, nothing otherwise.

Its this one, btw: https://www.youtube.com/watch?v=1La4QzGeaaQ (format 272, 8K VP9)

LAV DXVA2 (Native) Decode Performance
Avg 75.6 FPS (didn't DXVA checker use to show min/max fps?)

[NikosD]

Its this one, btw: https://www.youtube.com/watch?v=1La4QzGeaaQ (format 272, 8K VP9)

LAV DXVA2 (Native) Decode Performance
Avg 75.6 FPS (didn't DXVA checker use to show min/max fps?)I have reported the issue of missing min/max fps to DXVA Checker's developer a month ago and he is going to put it again in the next version.
I have a beta in my hands, showing min/max fps.

For VP9 8K60fps clip the average of 76fps is good but not far away of the real-time decoding.

What is the performance of a Pascal (any) card on the same clip ?

[wanezhiling]

1050ti 68fps
uhd620 64fps

[NikosD]

1050ti 68fps
uhd620 64fpsA difference of around ~10% could be just the clock, so the VP9 HW decoder of Turing and Pascal is probably the same.
Thanks.

[Zetti]

Microsoft Launches Free AV1 Video Codec for Windows 10

https://news.softpedia.com/news/microsoft-launches-free-av1-video-codec-for-windows-10-523704.shtml

[NikosD]

To be honest, the main problem of Turing cards and especially 2080 Ti, is not the video decoder performance or drivers.

It's the meaningless selling point of RTX technology, completely useless right now and unfortunately even when games come with RTX support, they are going to run at 60fps only for resolutions of 1080p and below.

What's the point of giving 1500€ to run games at 1080p ?

The value for money of Turing cards is the worst ever for any card I remember of.

And above all these, the failure rate of 2080 Ti cards is unprecedented.

They just die after a few working weeks and a lot of times even the replacement just dies too!

This is simply tragedy...

[Manni]

To be honest, the main problem of Turing cards and especially 2080 Ti, is not the video decoder performance or drivers.

It's the meaningless selling point of RTX technology, completely useless right now and unfortunately even when games come with RTX support, they are going to run at 60fps only for resolutions of 1080p and below.

What's the point of giving 1500€ to run games at 1080p ?

The value for money of Turing cards is the worst ever for any card I remember of.

And above all these, the failure rate of 2080 Ti cards is unprecedented.

They just die after a few working weeks and a lot of times even the replacement just dies too!

This is simply tragedy...

I'm glad I passed on the 2080ti due to the lack of HDMI 2.1 and the 14nm process. Waiting for the next gen happily with my 1080ti in the meantime.

Do you have a link to the info re failure rates? I have no interest in the current gen but I do plan to get the next gen when they offer a HDMI 2.1 model with 7nm process. I agree the price is a rip-off, but given the quality you get with MadVR and the fact that video processors costing three or four times as much don't give that quality, I can still justify putting that kind of money in a GPU. I sure wouldn't if it was only for gaming.

I'm hoping that Madshi will, at some point, exploit the new Tensor Cores in the RTX models. When that happens, it will become even more justified, at least for my use. I certainly won't buy a new GPU without Turing to replace my 1080ti, if only just for that probable MadVR capability to make the most of the new technology in the short/mid term.

EDIT: no worries, found a link: https://www.techradar.com/uk/news/nvidias-flagship-rtx-2080-ti-graphics-cards-are-failing-more-than-they-should
EDIT2: and another one with actual numbers: https://www.digitaltrends.com/computing/nvidia-rtx-2080-ti-founders-issues/
Doesn't seem as bad as you say, especially if you avoid the FE, which I would never buy anyway (I usually buy EVGA for nVida)...

[Zetti]

Is GTX 1100 series still not on the way??

I hope because GTX 1000 series is 2 years old.

[huhn]

the 2000 series is out

[Zetti]

Yes RTX 2000 series is out but the price is too high.

[nevcairiel]

Yes RTX 2000 series is out but the price is too high.

Why would you think t here would be a 11 series if they went straight to 20? :)

[huhn]

that doesn'T change that the new cards are out if you like the new prices or not.

[Zetti]

Why would you think t here would be a 11 series if they went straight to 20? :)
I have seen news articles about both GTX 1100 Series and RTX 2000 Series will be released.

RTX 2000 Series is released, but i still waiting on GTX 1100 Series.

[NikosD]

Microsoft Launches Free AV1 Video Codec for Windows 10

https://news.softpedia.com/news/microsoft-launches-free-av1-video-codec-for-windows-10-523704.shtmlPlease take a look here for a quick benchmark:
http://forum.doom9.org/showthread.php?p=1857315

[nevcairiel]

RTX 2000 Series is released, but i still waiting on GTX 1100 Series.

There is no 11 series. Before the release, a lot of people were speculating if it would be 11 or 20. 20 won. There were never any hints that both would exist.

[Zetti]

Ok thanks for info.

[aufkrawall]

libdav1d is a nice improvement for AV1 decoding, already works fine for me for the YouTube testvideos with lower CPU load than libaom or the Microsoft decoder.

[NikosD]

One of the worst client humiliation and disrespect coming from nVidia of course...Who else ? (besides Intel)
https://www.techspot.com/news/77445-nvidia-addresses-failing-geforce-rtx-2080-ti-cards.html

[aufkrawall]

I want my own personal ranting thread too. :)

[NikosD]

Oh, nonsense.

I was just replying to Manni using words of TechSpot's article.
Nothing more and nothing less.

[vanden]

Hello,

Is there a recent processor test (Ryzen & Threadripper (Zen, Zen +), Coffee Lake & Coffee Lake Refresh) in Software (CPU) mode ?

And sorry (if it's a few pages before) I did not have the courage to read the 123 pages ...

[NikosD]

Hello,



Is there a recent processor test (Ryzen & Threadripper (Zen, Zen +), Coffee Lake & Coffee Lake Refresh) in Software (CPU) mode ?



And sorry (if it's a few pages before) I did not have the courage to read the 123 pages ... No, there isn't.
For mainly two reasons:

1) I don't own such hardware

2) Most of the people use hardware acceleration for decoding.

BTW, speaking of HW acceleration, Intel's latest Gen11 iGPU combined with Sunny Cove architecture (actual implementation will probably be called Ice Lake) on 2019 will include parallel decoders for multi-stream concurrent decoding or faster one stream decoding.

From anandtech.com
Intel also stated that by using parallel decoders it can either support concurrent video streams or they can be combined to support a single large stream, and this scalable design will allow future hardware to push the peak resolutions up to 8K and beyond.

[NikosD]

There is a huge lie regarding Gen11 iGPU or at least misinformation.

It's not the first Teraflop iGPU by Intel and probably not the fastest one.

Skylake, back to 2015, had a GT4e version (Iris Pro 580) that leveraged 72 EUs and Max 1.15 TFlops performance using also 128MB of embedded memory.

I doubt that 2019 iGPU Gen11 will be faster than that.

[Zetti]

New rumors for GTX 1100 series.

https://wccftech.com/nvidia-geforce-gtx-1160-and-geforce-rtx-2060-launch-specs-rumor/

https://www.tomshardware.com/news/nvidia-geforce-gtx-11-series-1160-ti-rumor,38283.html

[Yups]

There is a huge lie regarding Gen11 iGPU or at least misinformation.

It's not the first Teraflop iGPU by Intel and probably not the fastest one.

Skylake, back to 2015, had a GT4e version (Iris Pro 580) that leveraged 72 EUs and Max 1.15 TFlops performance using also 128MB of embedded memory.

I doubt that 2019 iGPU Gen11 will be faster than that.


The first mainstream "for everyone" 1 Tflop iGPU from Intel for budget notebooks upwards. Skylake GT4e was basically non existent and its performance lacklustre (only 50% faster with 3x the EU count). Even Intels GT3e is more or less dead. I also believe that 64 Gen11 EUs will be easily faster/more efficient than 72 Gen9 EUs in real world circumstances at the same power envelope. Gen11 is so much newer with a better memory compression, bigger cache and lots of other improvements.

[Zetti]

http://bluesky23.yukishigure.com/en/DsfMftViewer.html
Version 1.8.0 (01/06/2019)
Added support to check dll information about the following MFT in Windows10 Version 1809
AV1 Video Extension
HEIF Image Extension
VP9 Video Extension
Webp Image Extension
Added some subtype names
Minor changes and fixes

[NikosD]

Nice.

But still waiting for his new DXVA Checker with min/avg/max fps and some other fixes.

It's been almost 9 months from the last version.

[Arm3nian]

How do you test LAV D3D11 decode performance with DXVA Checker? The option is always grayed out.

[NikosD]

How do you test LAV D3D11 decode performance with DXVA Checker? The option is always grayed out. IIRC you need at least Win 8 OS

[Arm3nian]

I'm on Windows 10, and the latest version of everything else. D3D11 works in LAV, even though all the decoding options show inactive for some reason. It's there in the play and playback performance menus, but not in the decode performance.

[nevcairiel]

You cannot test native D3D11 mode in LAV with anything but madVR itself.

[NikosD]

I'm on Windows 10, and the latest version of everything else. D3D11 works in LAV, even though all the decoding options show inactive for some reason. It's there in the play and playback performance menus, but not in the decode performance. Indeed. I can't see the option.
Is there somewhere in this thread a D3D11 native performance post for LAV ?

[huhn]

as new said it can only be used with madVR which has quite some over head by doing thing properly compared to other renderer so it can't be measured comparable.

[nevcairiel]

The performance should be rather quite similar to DXVA2 Native anyway. Its practically the same code, just with different D3D objects. The big advantage is more freedom in accessing the data from a video renderer, without loss in quality.

[aufkrawall]

D3D11VA even runs on the slowest Intel IGPs with mpv's zero copy decoding (while 10 bit video being free of banding at the same time, as "dumb-mode" allows dithering).
I can play 4k 60fps VP9 10 bit (btw: thanks again @nev for implementing it in ffmpeg!) HDR video + tonemapping on Gemini Lake (until it throttles).

[wanezhiling]

https://lists.freedesktop.org/archives/amd-gfx/2018-May/022291.html
https://lists.freedesktop.org/archives/amd-gfx/2018-May/022282.html
Vega20, includes VCE 4.1 and two UVD 7.2 instances.

[NikosD]

https://videocardz.com/77895/the-new-features-of-nvidia-turing-architecture



Main 4:4:4 12 profile hardware decoding? Quite an upgrade from Pascal's Main12 profile 4:2:0 hardware decoding.Latest nVidia SDK 9.0 is out and it seems that decoding features are the same like Volta, adding HEVC 4:4:4 up to 8K and 12bits.

The HEVC/VP9 support up to 8K and 12bits is impressive, but the info regarding H.264 8bit support of 8K resolution is not accurate, it's still 4K.

Regarding my previous comment of Turing Tesla decoding performance, it seems that only professional cards (Quadro, Tesla) will be that fast.

Turing's decoder performance regarding consumer/gaming cards should be on par of the very rare Volta's decoder which is a little (?) faster than Pascal's second generation decoder (1050 Ti etc)

Maybe a few decoding benchmarks of Turing vs Pascal (new) would be useful.

[NikosD]

Intel's upcoming Gen11 Icelake iGPU besides 64 EUs vs 24 EUs of Gen9 has also doubled Media Decoding/Encoding performance by adding a second MFX unit.

Intel has also added 4:4:4 support for both HEVC and VP9 and VP9 encoding.

It still misses 12bit that nVidia already supports and it doesn't mention anything of AV1 support.

[sneaker_ger]

Yes, it was said that Gen11 wouldn't have AV1.
There is not any AV1 acceleration yet but it's "coming soon" past Gen11 and Intel is "fully committed" to AV1.
https://www.phoronix.com/scan.php?page=news_item&px=Intel-Gen11-Graphics-Details

[nevcairiel]

Hardware development cycles are likely too long to even be able to make Gen11 if they wanted to. I wouldn't expect AV1 hardware before Mid/End 2020.

[Zetti]

http://bluesky23.yukishigure.com/en/DXVAChecker.html
Version 4.2.0 (03/30/2019)
Discontinued support for AMD VP9 hybird decoder
Fixed driver setting issue on specific intel iGPU
Updated DSF/MFT Viewer
Minor Changes, fixes and improvements

[NikosD]

http://bluesky23.yukishigure.com/en/DXVAChecker.html Min/Max fps is back, but this was fixed months ago.

I see no differences from 4.1.2 justifying the delay, so why it took so long to be released is a mystery.

[Elo]

Please excuse me if the following does not belong in this thread. I find it very difficult to understand the Video processing in a software player (and for me in particular the J River Media Center MC24/25) for UHD material to such a level that I can dimension a 4k system.The unclear points is mainly how the processing tasks are controled and split between different HW components in the system. Can for instance the decding be done by the GPU on the CPU and the rendering be done on a external GPU?. I am about to upgrade my system to a true 4k system which involves changing out nearly all components in the chain (except power amplifiers and loudspeakers) and today there is several possible system layouts to consider. Currently I am using a high end “surround” processor/amplifier which acts as a video switch and sound decoder taking on all video sources , and pass them on to monitor and loudspeakers. The new smart TVs offers another possible layout namely using the smart TV as the videoswitch (Samsung QLEDS) and using the ARC (audio return channel) for the sound. These TVs have a input box with all connections that is connected to the TV through a thin cable containing fibers and power supply. Further these TVs have built in apps for the most popular streaming sources (much like the Apple TV). Sources not covered by dedicated Apps can be played through Chromcast Ultra or similar.

It seems to me that the 9th Gen Intel Coffe Lake on a motherboard that supports HDMI 2 through a DP 1.4 to HDMI 2 conversion chip would have the nessesary funcionality to support MC’s video playback system based on madVDR rendering. But does it have the nessesary computational power? I would like someone with bettere expertise than me to comment on this?

If we look at external GPUs Nvidia 1660 upwards they are gigantic in size and power consumption. I am not at all interested in gaming an only need a system capable of playing all videomaterial thrown at it. It seems to be to be an overkill to use 2060 GPU alongside Intels 7th generations CPU onwards.

From the 8th gen if Intels CPUs (Kerby Lake) with onchip GPU there is HW support for H.264, H.265, HDR etc. (4k):


With a mother board with Kerby Lake CPU and a “weak” GPU like Nvidia 1050 (exists in a passively cooled version), can the decoding (4k) be done in hardware on the CPU (rather its GPU) and the rendering on the external GPU and thus utlizing the HW capabilities to its maximum? (Controled by MC24 and madvr with monitor is connected throug the HDMI connection at the external GPU)
Further using a 9th generation Intel CPU on a motherboard with HDMI 2. Is this strong enough to perform high level rendering and decoding of H.265 coded video? If yes this opens up for the last generation Intel NUC to offer a bvery small and effective HTPC solution ONLY for media playback (no gaming).
Can these tasks be placed and controlde through the set up interface of madVR?


I would appreciate it a lot if someone can help me understand these issues!

.../Elo

[huhn]

1. you could that that but i doubt this is going to be any faster.
2. it can easily decode it but the processing power is pretty weak an AMD APU should be a better and cheaper choice.
3. no madVR is a renderer not a decoder/splitter this is job of the decoder/splitter.

you know there a re threads for madVR in forum?

[Elo]

1. you could that that but i doubt this is going to be any faster.
2. it can easily decode it but the processing power is pretty weak an AMD APU should be a better and cheaper choice.
3. no madVR is a renderer not a decoder/splitter this is job of the decoder/splitter.

you know there a re threads for madVR in forum?

Thank you ... Yes I know and will adress the issue there aswell..
I am suprised that you say the AMD APU ia a better choice? Can you direct me to benchmarks that expands on this?

[huhn]

wrong thread :-).
there is a reason the fastest "IGPU" intel has is an AMD card.
https://www.notebookcheck.net/Iris-Plus-Graphics-655-vs-Vega-11-vs-Vega-6_8828_8470_8466.247598.0.html

and an apu with an vega 11 cost a fraction.
the hardware decoder on amd are not the best i have no clue if they now have VP 9 profile 2 and if you care about this used but still niche codec.

[Zetti]

http://bluesky23.yukishigure.com/en/DXVAChecker.html
Version 4.2.1 (04/20/2019)
Fixed Media Foundation decoder check prosess crash issue when GPU does not support D3D11 video acceleration
Fixed Device Information submit issue since April 18
Fixed "Show in Groups" setting issue on DSF/MFT Viewer
Minor fix

[NikosD]

Just after a little more than one month and a half of its release, I decided to buy after many, many years a new gaming card.

I'm talking of course about nVidia GTX 1660, which I bought in a budget (230 €) dual fan implementation, but not low-cost quality.

The 1660 card has a reference boost clock of 1785 MHz and a TDP of 120W by nVidia, mine has a 1830 MHz boost and a TDP of 130W but apparently it's out of the box overclocked to 1965 MHz.

Running with a GPU Clock of 1965 MHz means that the Video clock is 1815 MHz, which was the running clock during testing.

I used DXVA Checker v4.2.1 with LAV filters v0.74.1.1 in DXVA2 native mode on Win 10 x64 (17763.475) and my system has a Core i3 4170 with 16GB (2x8GB) Dual Channel DDR3-1600MHz RAM.

Turing video decoder is especially optimized for HEVC and even more specialized for 8K resolution.

Check my results here:
https://forum.doom9.org/showthread.php?p=1799099#post1799099

Now, regarding older codecs used by Blu-ray like H.264, VC-1, MPEG2 I did a little test and find out that Intel with Haswell and onwards is still the king:


MPEG2 (1080p clip 38Mbps / 30fps)


Haswell@1.3GHz 1470 fps

Turing@1.8GHz 720 fps


VC-1 (Devil May Cry 1080p clip 40Mbps / 60fps)


Haswell@1.3GHz 1067 fps

Turing@1.8GHz 753 fps


H.264 (Ducks Take Off 2160p clip 370Mbps / 50fps)


Haswell@1.3GHz 144fps

Turing@1.8GHz 102 fps

Now, regarding more modern codecs than these used by Blu-ray - like VP9 - the situation is a little different, meaning that the performance is close to Pascal but only a little faster.

The 8K VP9 performance is not so optimized like 8K HEVC (8bit/10bit) for Turing.
Actually it should be the exact same video decoding engine of Pascal, running a little faster due to faster Turing clocks.

Using 8K VP9 8bit clip from YouTube https://www.youtube.com/watch?v=1La4QzGeaaQ as a reference, we get:


Turing@1.8GHz 74/77 fps/78

Pascal@1.65GHz 68 fps

Kabylake UHD 620@1.15GHz 64 fps

In real-time playback of the clip above, using my Core i3 4170 - Win 10 x64 - nVidia 1660 system and Chrome v74 or Edge, I see no dropped frames decoding the VP9 8K 60fps stream (full screen or window) while Video Engine load of GTX 1660 is steady at ~77%