Evaluation of HEVC decoders (SW, Hybrid and HW)

[Zetti]

http://bluesky23.yukishigure.com/en/DXVAChecker.html
Version 4.0.2 (03/04/2018)
Added an index to the GPU engine name when there were more than one engine with the same name
Minor fix

[Zetti]

http://bluesky23.yukishigure.com/en/DXVAChecker.html
Version 4.1.0 (03/25/2018)
Now show "VP9_Profile0_NoDXVA_AMD" on Decoder Device when your AMD Radeon GPU supports VP9 hybrid decode acceleration
Fixed error when checking device under specific conditions
Fixed error when checking DirectShow/MediaFoundation decoder on Windows Vista
Changed GPU selector to GPU menu
Added "Submit Device Information to Database" menu item to GPU menu
Minor changes and fixes

[NikosD]

https://techreport.com/r.x/2017_11_29_Here_s_a_first_look_at_the_battery_life_of_HP_s_Ryzen_powered_Envy_x3/dxva.png
Ryzen 5 2500U

Unfortunately, latest ultra fast (for gaming) iGPU RX Vega M GH of Hades Canyon doesn't support native VP9 (not even 8bit/ Profile 1).

It's only hybrid VP9 acceleration like Polaris and discrete VEGA.

https://s18.postimg.org/x79xhhat5/Radeon-_DC.png

[huhn]

they have the intel iGPU not a huge deal.

[NikosD]

they have the intel iGPU not a huge deal.

Unfortunately they don't:
https://www.anandtech.com/show/12572/the-intel-hades-canyon-nuc8i7hvk-review-kaby-lakeg-benchmarked/7

[huhn]

Unfortunately they don't:
https://www.anandtech.com/show/12572/the-intel-hades-canyon-nuc8i7hvk-review-kaby-lakeg-benchmarked/7

d3d11 copyback?

i'm not here to say this "system" is a nice thing or something but you can access the intel decoder if you want to.

and there is no question if the intel iGPU is present.

and how they test "play back "accuracy" anadtech seriously you even get a dropped frame every 24 mins...

[NikosD]

d3d11 copyback?

i'm not here to say this "system" is a nice thing or something but you can access the intel decoder if you want to.

and there is no question if the intel iGPU is present.

and how they test "play back "accuracy" anadtech seriously you even get a dropped frame every 24 mins...

No, you are wrong here.

It seems that Intel's system is using outputs only from RX Vega M, so no Intel iGPU can be used directly although it's present.

And from the article I read this:
madVR 0.92.12 was evaluated with MPC-HC 1.7.15 (unofficial release) with its integrated LAV Filters 0.71. The video decoder was set to Direct 3D mode, with automatic selection of the GPU for decoding operations.

It seems obvious to me that Intel wants RX Vega M to be used for video decoding and rendering, but the initial statements of Intel said that Intel's iGPU is present exactly for that reason, to be used for decoding and rendering should be executed on RX Vega M.

But then why did they leave Intel's iGPU active if they wanted RX Vega M only to be used even for decoding ?

It's just another weird decision from a big company.

What a common situation.

[NikosD]

d3d11 copyback?

i'm not here to say this "system" is a nice thing or something but you can access the intel decoder if you want to.

and there is no question if the intel iGPU is present.

and how they test "play back "accuracy" anadtech seriously you even get a dropped frame every 24 mins...


You could be right actually.

From Tom's Hardware I get a different picture for Hades Canyon (Kabylake-G)

All Kaby Lake-Gs support dual-channel DDR4-2400 (no ECC) memory.

You'll notice that they also sport HD Graphics 630 and AMD's discrete GPU, too.

Intel doesn't use its integrated engine for either NUC's display outputs.

Instead, that hardware remains available for OpenCL-based compute workloads and QuickSync-accelerated software.

Some Kaby Lake-G products will use HD Graphics 630 for driving up to three 4K monitors. In those cases, the processor will power gate (turn off) the Vega GPU and its HBM2 during light graphics workloads

Anandtech and ganesh should try that copy-back mode (D3D9 and especially D3D11) you mentioned just for testing.

Maybe it works, after all.

We have to wait for a better review than Anandtech's or a forum member from here to test it.

[Morku]

For LAV Video I can decide which Hardware Device is to use.
But what's about Edge/Firefox when I go for youtube and not to use Vega? Is there a setting for it?

I am also very interested in the new NUC models as HTPC. But disadvantages in Hardware Acceleration would be a completly no go for it.

[Warner306]

For LAV Video I can decide which Hardware Device is to use.
But what's about Edge/Firefox when I go for youtube and not to use Vega? Is there a setting for it?

I am also very interested in the new NUC models as HTPC. But disadvantages in Hardware Acceleration would be a completly no go for it.

I wouldn't use a NUC if I wanted to use madVR. The new NUCs will hardware decode, though, via the Intel iGPU.

[huhn]

DXVA will not work it needs a connected display.
if an intel iGPU is present (and it is) there is absolutely no reason why D3D11 should not work.

if lavfilter can use d3d11 copyback every program on windows can do the same in theory. it's more of the question if they make the work of adding that for these NUCs or not. checking 2 GPUs for the hardware acceleration possibilities is clearly doable.

[NikosD]

DXVA will not work it needs a connected display.

IIRC, the fake display trick on Windows 7 or better works for DXVA2 copy-back of LAV filters besides QuickSync of course.

But I'm not sure.

[huhn]

a GPU doesn't know the difference between a fake display and a real one so yes this would work but it is far from practical and why should someone do that when d3d11 is an option.

[NikosD]

D3D11 is not working on Windows 7, it needs a newer version.

Window 7 is sill a 50% OS

Also, I think D3D11 is not as popular, stable and with a predictive behavior like DXVA2.

[Morku]

I wouldn't use a NUC if I wanted to use madVR. The new NUCs will hardware decode, though, via the Intel iGPU.

Why not? Decoding with Intel GPU and scalling processing with Vega GPU would sound like a nice solution if possible and I am not overlook something important.

@NikosD
People who still use Windows 7 will know why.
For multimedia reason I don't see a point to stay for Windows 7.
Where did you get the information that D3D11 is not stable and predictive like DXVA2?
And if we are still talking about the newer NUC, Kabylake/Coffeelake don't support Windows 7 anyway.

[NikosD]

Regarding D3D11, just from my personal (little) experience using MS and LAV decoders.

Windows 7 is not official supported, but you can install it on Kabylake.

[huhn]

D§D11 is rock stable i'm not even sure if there is a single know crash that can happen under normal usage.

the lavfilter d3d11 implementation is using some parts from DX11 that is not present on windows 7.

if you want to use new stuff you have to use a new OS too it's simple as that and people with new hardware will use far far more than 50 % windows 10 i even go as far and say that a modern 2018 PC with windows 7 will be a total excepetion.

there are a lot of reasons not to use a system like that for madVR like zotac has PCs in that size with a 1060 in it for far less but do as you please...

[NikosD]

Don't take my posts personally and I almost never speak for myself in this thread.

My comments are general, trying always to include most cases, even the corner ones.

I install new Windows OS on all (almost) of my machines, even before the official release.

I hope it's more clear now.

[Zetti]

http://bluesky23.yukishigure.com/en/DXVAChecker.html
Version 4.1.1 (04/01/2018)
Fixed wrong buffer size in presenter
Minor changes

[el Filou]

The only letdown here is that Intel will not support HDR on Gemini Lake NUCs. You will need to get at least an i3 NUC in order to play HDR videos in Windows 10.Well that is a shame. :(

[clsid]

Encoding. Real-time is implicit for decoding.

[NikosD]

Cannonlake is dead more or less, especially 2+2. Intels next meaningful GPU progress will come with Icelake.

I'm really confused after the latest bunch of new Coffeelake CPUs and 300 series chipsets, which are actually Cannonlake's PCH.

Which will be Intel's next CPU for desktop and when ?

Is it a Cannonlake for desktop (CNL-S) at 10nm around Q3 2018 ?

Is it an Icelake for desktop using 10nm+ process on 2019 ?

Maybe a new Coffeelake at 14nm+++(+) ?

Very frustrated situation.

[el Filou]

Maybe they'll pull a Broadwell again and release just 2 SKUs for desktop Cannon Lake ?

[aufkrawall]

Also, I think D3D11 is not as popular, stable and with a predictive behavior like DXVA2.
My experience with various GPUs of Intel, AMD and Nvidia shows the opposite, D3D11VA without copyback is simply superior over DXVA2 in practice.
In mpv, you can use zero copy decoding, which is even faster than DXVA2 + EVR on Intel, while having uncompromised quality at the same time regarding internal processing precision (no banding). And unlike DXVA2, it doesn't manipulate colors in general.

[Yups]

I'm really confused after the latest bunch of new Coffeelake CPUs and 300 series chipsets, which are actually Cannonlake's PCH.

Which will be Intel's next CPU for desktop and when ?

Is it a Cannonlake for desktop (CNL-S) at 10nm around Q3 2018 ?

Is it an Icelake for desktop using 10nm+ process on 2019 ?

Maybe a new Coffeelake at 14nm+++(+) ?

Very frustrated situation.



There is an 8C Coffeelake-R coming this year. The next major upgrade will be Icelake-S next year, probably sometime in H2 2019.

[NikosD]

Thanks.

Then I'll probably buy next month my new system.

I'll wait for a Core i5 8400 vs Ryzen 2600 comparison using a B360 vs B350 chipset.

I lean toward Core i5 8400 due to iGPU since dGPUs are still too expensive and Ryzen doesn't have iGPU.

I could add a dGPU later, when Chinese Ethereum ASICs do their job for GPU prices.

There is an 8C Coffeelake-R coming this year.

But a Core i5 9400 6C/12T vs Core i5 8400 could be even more interesting, depending on its price though.

[Morku]

Unfortunately, latest ultra fast (for gaming) iGPU RX Vega M GH of Hades Canyon doesn't support native VP9 (not even 8bit/ Profile 1).

It's only hybrid VP9 acceleration like Polaris and discrete VEGA.


they have the intel iGPU not a huge deal.

Unfortunately they don't:
https://www.anandtech.com/show/12572/the-intel-hades-canyon-nuc8i7hvk-review-kaby-lakeg-benchmarked/7

No, you are wrong here.

It seems that Intel's system is using outputs only from RX Vega M, so no Intel iGPU can be used directly although it's present.

And from the article I read this:


It seems obvious to me that Intel wants RX Vega M to be used for video decoding and rendering, but the initial statements of Intel said that Intel's iGPU is present exactly for that reason, to be used for decoding and rendering should be executed on RX Vega M.

But then why did they leave Intel's iGPU active if they wanted RX Vega M only to be used even for decoding ?

It's just another weird decision from a big company.

What a common situation.

How about the new advanced graphic setting in Windows 1803?
https://pureinfotech.com/set-gpu-app-windows-10/

Shouldn't be possible to set the option for the Browser to use Intel GPU? Anandtech posted a picture, that DXVA checker do have access to Intel 630 GPU with VP9.
But how to set Intel GPU for Power saving and Vegas for High performance?
Would be UHD Blu-rays possible than?

[nsnhd]

Movie & TV has improved, plays more formats, more smoothly.

[Zetti]

http://bluesky23.yukishigure.com/en/DXVAChecker.html
Version 4.1.2 (05/19/2018)
Fixed app crash issue on Radeon GPU under specific conditions
Fixed app error issue when using D3D11 video acceleration during video palyback and benchmark under specific conditions
Minor changes and fixes

[NikosD]

The unbelievable truth regarding 10nm process of Intel:
In the end we have a chip being built on a troubled 10nm process.
In six months Intel can make 15% of a day’s production on those production lines.
The resultant chips are abjectly broken, they are 2+1 but the +1 doesn’t work which means they are selling a CPU with half the die turned off, an expensive proposition given the cost of the process and the shatteringly low yields.
Even with the GPU turned off, the resultant CPU uses twice the power of the 14nm devices to run slower than those with a GPU.

OEMs won’t touch these 10nm parts either voluntarily or with ‘not bribes’ so Intel has to twist arms and force them to make laptops and “Make it ‘look real“.
Why?
To put out a data point that they can build ‘truth’ and ‘alternative facts’ around when it comes time to talk to analysts.
The 10nm Cannon Lake parts aren’t real and never will be viable, financially or technically speaking.
Feel free to believe the PR messaging but you can’t say you don’t know what is really going on now.

https://semiaccurate.com/2018/05/29/is-intels-upcoming-10nm-launch-real-or-a-pr-stunt/

[mariner]

Request for VP9 8k10bits test sample

Greettings, many thanks to Nikos and others for doing all the hard work adn sharing their findings here.

The 8k streams found on youtube are mostly of 8bit variant. Would anyoen be kind enough to share a link for VP9 8k10bit clip?

Many thanks and best regards,

[sneaker_ger]

At least youtube-dl doesn't show any 8K HDR streams, always 4K and lower in HDR, 8K in SDR/8 bit. So I guess Youtube simply doesn't yet offer 8K HDR?

[littleD]

https://www.youtube.com/watch?v=1La4QzGeaaQ

here you have but i didnt download this video and didnt check actual properties. firefox addon download helper offer download but with additional software.

[sneaker_ger]

No, same as the other videos. No 8K HDR (unless there is a secret way I don't know about).
youtube-dl -F 1La4QzGeaaQ
[youtube] 1La4QzGeaaQ: Downloading webpage
[youtube] 1La4QzGeaaQ: Downloading video info webpage
[youtube] 1La4QzGeaaQ: Extracting video information
[info] Available formats for 1La4QzGeaaQ:
format code extension resolution note
249 webm audio only DASH audio 58k , opus @ 50k, 2.01MiB
250 webm audio only DASH audio 78k , opus @ 70k, 2.68MiB
171 webm audio only DASH audio 126k , vorbis@128k, 4.67MiB
140 m4a audio only DASH audio 128k , m4a_dash container, mp4a.40.2@128k, 5.12MiB
251 webm audio only DASH audio 150k , opus @160k, 5.29MiB
278 webm 256x144 144p 108k , webm container, vp9, 30fps, video only, 3.68MiB
160 mp4 256x144 144p 112k , avc1.4d400c, 30fps, video only, 2.55MiB
242 webm 426x240 240p 240k , vp9, 30fps, video only, 7.85MiB
133 mp4 426x240 240p 247k , avc1.4d4015, 30fps, video only, 5.48MiB
330 webm 256x144 144p60 HDR 248k , vp9.2, 60fps, video only, 8.73MiB
243 webm 640x360 360p 444k , vp9, 30fps, video only, 14.57MiB
331 webm 426x240 240p60 HDR 508k , vp9.2, 60fps, video only, 18.86MiB
134 mp4 640x360 360p 739k , avc1.4d401e, 30fps, video only, 16.87MiB
244 webm 854x480 480p 830k , vp9, 30fps, video only, 26.89MiB
332 webm 640x360 360p60 HDR 1093k , vp9.2, 60fps, video only, 40.84MiB
135 mp4 854x480 480p 1359k , avc1.4d401f, 30fps, video only, 36.82MiB
247 webm 1280x720 720p 1728k , vp9, 30fps, video only, 54.49MiB
333 webm 854x480 480p60 HDR 2017k , vp9.2, 60fps, video only, 76.49MiB
302 webm 1280x720 720p60 2664k , vp9, 60fps, video only, 91.83MiB
136 mp4 1280x720 720p 2724k , avc1.4d401f, 30fps, video only, 72.69MiB
248 webm 1920x1080 1080p 2748k , vp9, 30fps, video only, 96.11MiB
298 mp4 1280x720 720p60 4175k , avc1.4d4020, 60fps, video only, 117.08MiB
303 webm 1920x1080 1080p60 4499k , vp9, 60fps, video only, 157.80MiB
334 webm 1280x720 720p60 HDR 4560k , vp9.2, 60fps, video only, 175.81MiB
137 mp4 1920x1080 1080p 5047k , avc1.640028, 30fps, video only, 131.74MiB
299 mp4 1920x1080 1080p60 7030k , avc1.64002a, 60fps, video only, 203.39MiB
335 webm 1920x1080 1080p60 HDR 7046k , vp9.2, 60fps, video only, 269.70MiB
271 webm 2560x1440 1440p 9018k , vp9, 30fps, video only, 304.80MiB
308 webm 2560x1440 1440p60 13575k , vp9, 60fps, video only, 471.00MiB
336 webm 2560x1440 1440p60 HDR 16513k , vp9.2, 60fps, video only, 639.35MiB
313 webm 3840x2160 2160p 18432k , vp9, 30fps, video only, 649.89MiB
315 webm 3840x2160 2160p60 26946k , vp9, 60fps, video only, 1015.76MiB
337 webm 3840x2160 2160p60 HDR 29405k , vp9.2, 60fps, video only, 1.11GiB
272 webm 7680x4320 4320p60 31841k , vp9, 60fps, video only, 834.77MiB
17 3gp 176x144 small , mp4v.20.3, mp4a.40.2@ 24k
36 3gp 320x180 small , mp4v.20.3, mp4a.40.2
43 webm 640x360 medium , vp8.0, vorbis@128k
18 mp4 640x360 medium , avc1.42001E, mp4a.40.2@ 96k
22 mp4 1280x720 hd720 , avc1.64001F, mp4a.40.2@192k (best)

Gotta find a sample elsewhere (or encode yourself).

[mariner]

Thank fro the kind reply, sneaker and littedD.
The stream appears to be 8bit like many others, no. 272.
I guess sneaker is right about youtube-dl not picking up 8K HDR streams. Would anyone know if it's a youtube-dl limitation?
Many thanks and best regards,

[nevcairiel]

I'm not convinced YouTube actually offers those streams. youtube-dl just gets the info from the server, and if its not there...

[NikosD]

[U]The 8k streams found on youtube are mostly of 8bit variant. Would anyoen be kind enough to share a link for VP9 8k10bit clip?

Many thanks and best regards,

I'm at the beach right now :)

I'll try a Google search later but probably sneaker is right and you have to encode it yourself, it's a little bit difficult request.

I'll do my best.

[mariner]

I'm at the beach right now :)

I'll try a Google search later but probably sneaker is right and you have to encode it yourself, it's a little bit difficult request.

I'll do my best.

Both you and nev deserve that.
:)

[mariner]

Intel Graphics HD 620

Greetings.

Thinking of getting an old entry level Kabylake with mobile i5 7200U/DDR3 running Win7(hacked?). Would like some advice on the hardware decoding capabilities of the HD 620 graphics processor.

1. What's the concensus of this forum on HD620's hardware decoding of 10bit HEVC and VP9, 8K & 4K? Full or hybird? Any bitrate limitation?
2. Is Win7 up to the task? Or is Win10 required? ie D3D11 etc.
3. If DXVA Checker shows HEVC_VLD_Main10, how does one tell if it is full or partial hardware decoding?
4. Does i5 7200u have the power to do softwaer decoding of these 10bit video? i7 7500U?

Many thanks and best regards,

[huhn]

1. they are hardware decoder and they have a bitrate limitation but that'S unimportant for real usages because it is insanely high.

2. lavfilter D3D11 needs windows 8 or newer.

3. you don't know it from this information.

4. no... not even close

[NikosD]

[U]
3. If DXVA Checker shows HEVC_VLD_Main10, how does one tell if it is full or partial hardware decoding

I agree with the previous post of huhn, but I want to add a few things.

For HD 620, what you see in DXVA Checker is full hardware decoding for all codecs/resolutions.

But generally speaking you can't say for sure.

For me, the best way to find out what is really going on, is by monitoring CPU and mainly GPU loading while decoding in DXVA native mode (LAV) and pure decoding DXVA Checker benchmark (not playback mode)

Full hardware decoding gives almost 0% usage for both CPU and GPU loading.

If you see a two digit figure for GPU loading like 20% or more, then it's probably hybrid decoding.

For the rest of your questions, huhn has already replied.

BTW, my first research regarding VP9 8K 10bit gave me nothing.

Almost all of the Google results gave me a YouTube announcement regarding 8K VP9 10bit HDR support, which seems that it doesn't exist.

[mariner]

Thanks for the kind reply huhn and Nikos.

1. So the HD620 should have no issues with the likes of 400mbps Jellyfish and MBC UHD Music Core GFriend - Rough (4in1). Any known test clip that it cannot handle?
2. A question for those with experience running Win7 on Kabylake. Is the good old DXVA2 up to the task?

Many thanks and best regards,

[dredik]

1. On a laptop with i3-7100u with hd620 (2*4Gb dual channel RAM) MBC UHD Music Core GFriend - Rough (4in1) was lagging pretty much. Cant say about hd620 in other CPUs. But I must say that GTX1050Ti 4Gb paired with core i5-7400 (desktop) cant handle this clip either (I could not make it play smooth without hccups :(). 400mbps Jellyfish had no problems on i3-7100u - played great, without frame drops.
2. KabyLake CPUs is not supported by Windows 7. You can install Windows7 on KabyLake system but there is no official HD620 windows7 driver.

[huhn]

why do you even care about these dumb test files?

no commercial file will ever be encoded like this UHD BD are capped at a something like 125 mbit sorry i don't know the exact number because it doesn't matter.

8k may not even use HEVC as a codec but a codec that is not yet supported like AV1.

i mean your GPU can't output 8K 60 hz and no other GPU too and DXVA downscaling is pathetic.

[hajj_3]

UHD bluray max bitrate is 128Mbps. The maximum of kaby is ridiculously high, something like 400Mbps so all consumer h265 videos can be hardware decoded.

[littleD]

Intel Graphics HD 620

Greetings.

Thinking of getting an old entry level Kabylake with mobile i5 7200U/DDR3 running Win7(hacked?). Would like some advice on the hardware decoding capabilities of the HD 620 graphics processor.

1. What's the concensus of this forum on HD620's hardware decoding of 10bit HEVC and VP9, 8K & 4K? Full or hybird? Any bitrate limitation?
2. Is Win7 up to the task? Or is Win10 required? ie D3D11 etc.
3. If DXVA Checker shows HEVC_VLD_Main10, how does one tell if it is full or partial hardware decoding?
4. Does i5 7200u have the power to do softwaer decoding of these 10bit video? i7 7500U?

Many thanks and best regards,I have 620. It plays 8k files nicely i mean hevc and vp9 and rest is 4k only including h264. All files play fluently with dxva, if you wanna use d3d11 forget about high resolutio and 10 bit cause it stutters.
dxva checker cannot show You if chip uses partial decoding. There were some partial decoding profiles long time ago, but they were partial by it means.
4k 50 fps 10bit is slow in software for 7200u, i have no sample with 10 bit 25 fps.
Intel team always drop old windows systems early, i see no drivers for hd620 and windows 7...

[mariner]

Thanks for the kind reply dredik, littleD and otheres.

If HD620 cannot cope with the MBC clip despite being identified by DXVA Checker as 8K HEVC 10bits capable, would anyone know if the UHD 605 (Gemini Lake) or UHD 620 (8th Gen.) can handle this?

Thanks and best regards,

[NikosD]

Does anyone know the difference between Titan V and Pascal's latest video decoder like 1050/1050 Ti/1080 Ti ?

According to DXVA Checker they have similar features but according to Nvidia they belong to different VPDAU sets.

Maybe the have a difference in speed.

Is it Titan V any faster in video decoding than 1080 Ti for example ?

[huhn]

do you really care about a 10k cards hardware decoder?

for end user it should be more interesting if nvidia is releasing new gaming GPU at gamescom.

these quadro cards are most likely 12 nm i only found the die size which is nearly as big as volta with a lower transistor count so unlikely to be 7 mn.

the next big thing for hardware decoder should be AV1 and it's a little bit early for it but i guess doable if they really really wanted to do it but windows has to add it too the 960 had a VP9 decoder but it took quite some time before it was usable.
so i don't except anything new from nvidia this round.

[Cary Knoop]

Is it Titan V any faster in video decoding than 1080 Ti for example ?
Video decoding is not very efficient for CUDA cores. NVIDIA and other GPUs have dedicated encoding and decoding modules. The GPU is more important for image processing.

For image processing, the Titan V is definitely faster but it depends how much of a difference you will actually see!

Most video processing applications use single precision functions as there is no need for double precision functions. The special thing about the Titan V is fast double precision functions (good for artificial neurons) and you pay for that functionality. You can buy three 1080Ti's for one Titan V!

So unless your CUDA software takes advantage of double precision functions for image processing (not likely!) a lot of the gain from the Titan V will be unused.