Evaluation of HEVC decoders (SW, Hybrid and HW)

[loekf]

What kind of clock throttling do you expect during benchmarking?

I also have a 1050 Ti and during regular video playback all clocks fluctuate, as expected.

Throttling as in a different clocks depending on how many resources are used for video decoding. However, I now realise the decoder is pushed for maximum fps, so max resources, so max frequency.

Guess if you use VLC or ffmpeg you see more dynamics.

I will do the 10-bit stuff tonight.

[loekf]

My 10-bits results (MSI GTX 1050 TI 4GB):

1. UHD_PQ_Lovely_Swiss.ts

Decode: 159-179-211
Playback: 171-190-217

2. 2.Crowd_run_2160p50_Main10_100Mbps.mkv

Decode: 175-179-181
Playback: 173-176-177

3. jellyfish-140-mbps-4k-uhd-hevc-10bit.mkv

decode 130-146-165
playback 131-147-165

4. jellyfish-400-mbps-4k-uhd-hevc-10bit.mkv

Decode: 60-74-96
Playback: 60-75-96

5. Arri65_from_raw.mp4 (not 8K, but more like 6560 x 3100 pixels ?!)

Decode: 63-68-72
Playback: 64-68-72

Note: all playback benchmarks have output downscaled to 720p

Also looked at clocks during video playback using VLC. VLC 2.2.4 has some bug with DXVA2 (crashed). 3.0 beta worked fine and indeed during SD or 1080p x264 files (BD rip) the video clock "hoovers" around 540 - 570 Mhz. CPU load is not more than 4%. GPU and memory clock stay low as well (didn't write down the numbers).

[CruNcher]

1645 MHZ is a reasonable clock for a modern video decoder.

During decoding benchmarking (not playback benchmarking) the clocks are going up and down ?

It should be max all the time during benchmarking (decode or playback)

What is the stable clock during playback ?

Are you kidding for a Asic that's crazy never ever that's the VP8 Asic Frequency.

400-600 MHz sounds reasonable but never ever 1600, though in Nvidias Case we have to think indeed more about a MultiCore Structure.

And the overall Performance seems to depend on the core Count or rather a Virtualized Splitted core count.

http://i1.sendpic.org/i/bP/bPW89uDcWkoLwoDorkuSi10Zg0V.png


2) AMD RX 470 has obviously a different philosophy, covering 60fps only at "normal" bitrate clips up to 80Mbps which should cover a very large amount of real world clips out there.

I see that different i see it in the amount of Streams the ASIC is capable decoding at once the Performance can be either used to get a higher bitrate or higher resolution split.

That Nvidia can decode such high bitrates is only because their target was 8K so you either decode 2x4K(UHD) at reasonable bitrate or 1x8k at lower bitrate/FPS (half) then what is possible at 4K(UHD).

The same you could decode mostly hundreds of SD at once depending on your complexity ;)

You should more think of it like 0-100% Utilization if you dont go over the 100% Utilization you can do whatever you want, it doesn't matter what you playback or how much bitrate the highest complexity possible is the overall Performance ;)

I can easily playback 4X UHD AVC 30Mbit 30 fps Streams at once on GM204 and im not reaching the 100% ;)

4x30 = 120 Mbits @ 30 FPS

So in theory GM204 could playback 8K AVC at 60 mbits 30 fps stable ;)

http://oi66.tinypic.com/w9te09.jpg

Almost exactly the same result as if i Decode 2x the same complexity @ 60 FPS ;)

2x30 = 60 Mbits @ 60 FPS

http://oi64.tinypic.com/16krlo7.jpg

3x30 = 120 Mbits @ 60 FPS brakes apart and over utilizes the ASIC

http://oi64.tinypic.com/140cay1.jpg

[NikosD]

My 10-bits results (MSI GTX 1050 TI 4GB):


Thank you very much for your time and effort.

Since the 1st bonus clip is closer to 7K, I added a 2nd one in my 10bit HEVC list which now has 6 clips.

The 2nd bonus clip is 8K but huge (~2GB)

You can download it and test it from here:
https://mega.nz/#!8hcmwDoK!UwfMhPQvtiJw9SkZpijJijxyC4k45iqrNTwMVHjUZ6Y

My 10-bits results (MSI GTX 1050 TI
Also looked at clocks during video playback using VLC. VLC 2.2.4 has some bug with DXVA2 (crashed). 3.0 beta worked fine and indeed during SD or 1080p x264 files (BD rip) the video clock "hoovers" around 540 - 570 Mhz. CPU load is not more than 4%. GPU and memory clock stay low as well (didn't write down the numbers).

Can you use MPC-HC or MPC-BE or PotPlayer and try again ?

I would definitely like to know your GPU Video clock along with GPU core/ memory clocks but using more familiar (and better) video players during playback.

Using GPU-Z you could see the Video Engine utilisation and you could try some heavy (bitrate), high framerate clips like Netflixx 4K HEVC 8bit@60fps-700Mbps in order to see the clocks rising.

As a last favor, if you have more time and you are in a good mood, I would like to add your Kabylake's results.

I saw that it's using HD 630 iGPU at 1.0GHz.

Thanks in advance!

[NikosD]

Are you kidding for a Asic that's crazy never ever that's the VP8 Asic Frequency.

400-600 MHz sounds reasonable but never ever 1600, though in Nvidias Case we have to think indeed more about a MultiCore Structure.


I'm never kidding actually talking about HW evaluation/ benchmarks.

When Intel introduced for the first time in 2011 a highly integrated full fixed-function ASIC (QuickSync) for video decoding/encoding inside Sandybridge's (2nd generation Core) iGPU, the reason that destroyed the competition of ATI/Nvidia was mainly (among other things) it's clock ~ 1150 MHz, when ATI/Nvidia had half (or less) clock.

I was able and I'm still able to overclock this ASIC inside my Core i5-2400 at around 2000MHz even today.

Do you believe that Nvidia Pascal with those crazy GPU clocks ~2000MHz, is holding the video engine clock to a max benchmark mode (decode or playback) at ~400-600 MHz ?

Is it even possible to hit Kabylake's (7th generation Core) video decoding performance when the iGPU HD 630 of Kaby has a video clock of 1150MHz, with your Pascal's estimation of half clock (~400-600)?

Now I think you are kidding me and that was a good joke, I have to admit it.

[CruNcher]

Still you are wrong even with Decoding HEVC the Main Core has not that of a high Frequency as you suggest here or rather believe.

[loekf]

Thank you very much for your time and effort.

Since the 1st bonus clip is closer to 7K, I added a 2nd one in my 10bit HEVC list which now has 6 clips.

The 2nd bonus clip is 8K but huge (~2GB)

You can download it and test it from here:
https://mega.nz/#!8hcmwDoK!UwfMhPQvtiJw9SkZpijJijxyC4k45iqrNTwMVHjUZ6Y


Checked the 8K stream:

Decode: 53-56-58
Playback: 54-55-56

(using Nvidia's latest drivers, 378.78)

Also had a look at clocks again:

for a BD-rip, 1080p H.264, 9.2 Mbps video + 1.5Mbps DTS (bitstreaming) you see:

645 MHz for the GPU clock
570 Mhz for the GPU video clock
202.5 Mhz for the GPU memory clock

(clocks are stable, doesn't matter whether you have a quiet scene or a scene full of action and CGI)

suppose you take a 4K, HEVC file, 140 Mbps:

1835 MHz for the GPU clock
1645 Mhz for the video clock
1752 MHz for the memory clock

(using MPC-HE, x64, clocks traced using HWInfo)

Kabylake still pending.

[NikosD]

Checked the 8K stream:

Decode: 53-56-58
Playback: 54-55-56

(using Nvidia's latest drivers, 378.78)


Thanks, I will update my post.
It seems that we managed to find a file that Pascal 1050 Ti can't decode in real time, although it's close.

Also had a look at clocks again:

suppose you take a 4K, HEVC file, 140 Mbps:

1835 MHz for the GPU clock
1645 Mhz for the video clock
1752 MHz for the memory clock

It makes sense, that file needs maximum clocks, so I consider the 1645 MHz as benchmark mode clock (for playback and decode mode)

Next time see the video decoding utilisation.
Nvidia is the only company providing this info and is useful because you can see the margins of video decoding performance.
Use GPU-Z for that.

Kabylake still pending.

That would be extremely useful in order to compare it with Pascal.

[CruNcher]

Samsungs UHD Highlight Demo of

https://www.youtube.com/watch?v=pnuj1svkH8w

is slowly leaking through it seems ;)

[loekf]

Next time see the video decoding utilisation.
Nvidia is the only company providing this info and is useful because you can see the margins of video decoding performance.
Use GPU-Z for that.


Actually DXVA checker also reports the GPU (and CPU) load. It's usually always > 90% and breathing a little bit (a few %).

I guess the video decoder (nvdec or purevideo) utilization scales with clock speed. Not sure how it is calculated. The Nvidia SDK docs show little information, e.g.:

http://developer.download.nvidia.com/compute/cuda/5_5/rel/nvml/nvml.5.319.43.pdf

In other words, the video clock scales up or down based on what you "throw" at it, so the driver and GPU HW will always scale the clocks such that the load fits and the clocks are as low as possible (engineering hat on: make them such that you can process @ 90% + 10% margin).

So.. ideally the load should be around 90%.

[NikosD]

Actually DXVA checker also reports the GPU (and CPU) load. It's usually always > 90% and breathing a little bit (a few %).

So.. ideally the load should be around 90%.

What I'm talking about has absolutely nothing to do with GPU and CPU load.

Those two should be <5% or even ~0% during decode benchmark.

But for playback benchmark mode, yes the GPU utilisation should be close to 100%

DXVA checker for Nvidia cards should show a 3rd utilisation of Video decoder.

That is what I ask.

You can play the HEVC 10bit 140Mbps clips from inside DXVA Checker using normal playback (not benchmark) and check that 3rd parameter.

Otherwise just use GPU-Z sensors tab that report video utilisation.

[mixus]

Note :
4K 120 fps = 8K 30 fps
4K 240 fps = 8K 60 fps


8K - 30 fps = OK
4K - 180 fps = OK
8K - 60 fps = FAIL

If You Want 8K 60 Fps Wait Nvidia Volta or AMD Navi


Result
http://s6.dosya.tc/server8/mznlzn/Result.rar.html

[NikosD]

Note :
4K 120 fps = 8K 30 fps
4K 240 fps = 8K 60 fps


8K - 30 fps = OK
4K - 180 fps = OK
8K - 60 fps = FAIL

If You Want 8K 60 Fps Wait Nvidia Volta or AMD Navi


Result
http://s6.dosya.tc/server8/mznlzn/Result.rar.html

Hey,

I can see you have a Pascal GTX 1060.

What is your clocks (GPU/Memory/Video clock) during benchmarks and is there any difference in performance between your card and 1050 Ti that I have already posted ?

If performance difference is big, maybe you could post your results in my samples collection in order to put them in that comparison table.

Also, are there any links for those clips that you have provided screenshots only ?

Now, as a last point, I don't know about Volta but AMD's Navi is very difficult to include a 8K60fps decoder, simply because AMD is not interested in any non-consumer resolution, video codec, bitrate etc.

Polaris UVD has included HEVC decoding/ encoding after years of the HEVC release just because BluRay Ultra HD is close and has started to move a little.

We have to wait enough for AMD to give us 8K60fps decoder.

[mixus]

Hey,

I can see you have a Pascal GTX 1060.

What is your clocks (GPU/Memory/Video clock) during benchmarks and is there any difference in performance between your card and 1050 Ti that I have already posted ?

If performance difference is big, maybe you could post your results in my samples collection in order to put them in that comparison table.

Also, are there any links for those clips that you have provided screenshots only ?

Now, as a last point, I don't know about Volta but AMD's Navi is very difficult to include a 8K60fps decoder, simply because AMD is not interested in any non-consumer resolution, video codec, bitrate etc.

Polaris UVD has included HEVC decoding/ encoding after years of the HEVC release just because BluRay Ultra HD is close and has started to move a little.

We have to wait enough for AMD to give us 8K60fps decoder.

Yeah man, I have GTX 1060.

I think AMD and Nvidia will add to the next architecture.

I prepared video for you. please download and look.

http://s6.dosya.tc/server8/285lne/GTX_1060.rar.html

[NikosD]

Yeah man, I have GTX 1060.

I think AMD and Nvidia will add to the next architecture.

I prepared video for you. please download and look.

http://s6.dosya.tc/server8/285lne/GTX_1060.rar.html

Nice video, thank you for your time and effort.

I'm afraid you didn't answer a few questions.

Are those videos in public or private ?

Are there any links for 8K videos ?

I saw that you have also a HD 530 iGPU and you are familiar with DXVA Checker.

Why don't you test my samples collection with both your GPUs ? (1060 and HD 530)
If you have time and you are in a mood of course.

Now, the GPU core clock of your Pascal 1060 is huge - 2GHz - but GPU-Z doesn't provide the GPU Video clock.

You have to use HWiNFO in order to find out the GPU Video Clock during benchmarking.

One last thing.

I didn't see LAV video in your codecs.
Don't you use it ?

[NikosD]

Actually DXVA checker also reports the GPU (and CPU) load. It's usually always > 90% and breathing a little bit (a few %).


Now that I saw this video http://s6.dosya.tc/server8/285lne/GTX_1060.rar.html the monitoring parameter I was talking about regarding DXVA Checker is called VE usage (Video Engine usage) and it's right next to GPU/VE usage ;)

[aufkrawall]

The shown usage might be as mislading as the shown GPU usage as long as the GPU isn't running with full clocks.
Means if it only runs with ~1GHz but shows 50% VE/VPU usage, the latter value might be much lower for the same task if the GPU runs with a higher frequency.
So it's not a good indicator for the actual performance.

Btw: Dunno if this has already been mentioned here, but I've been told that all Pascal GPUs except of GP104 are able to decode VP9 10/12 bit.

[sneaker_ger]

Told by whom? GTX 1060 is GP106 and in tests it didn't show VP9 10 bit. (Haven't seen any Titan X or GTX 1080 Ti tests). And I don't remember anyone saying VP9 12 bit is supported on any Pascal card.

[aufkrawall]

Told by somebody close to Nvidia:
https://www.forum-3dcenter.org/vbulletin/showpost.php?p=11298988&postcount=36
You can question how well communication between Nvidia and their internal testers is, but unless you have a better bet...

[huhn]

just take this source:
https://developer.nvidia.com/nvidia-video-codec-sdk

some pascal card can do 12 bit VP9.

[sneaker_ger]

I see. Interesting.

[nevcairiel]

Yet 1060 (GP106) cards do not support it as of the most recent drivers, while we've heard that a 1050 does. So it would seem either some implementation isn't quite there yet (which would be odd if it shows up and works on a 1050), or the information isn't fully accurate.

[mixus]

Nice video, thank you for your time and effort.

I'm afraid you didn't answer a few questions.

Are those videos in public or private ?

Are there any links for 8K videos ?

I saw that you have also a HD 530 iGPU and you are familiar with DXVA Checker.

Why don't you test my samples collection with both your GPUs ? (1060 and HD 530)
If you have time and you are in a mood of course.

Now, the GPU core clock of your Pascal 1060 is huge - 2GHz - but GPU-Z doesn't provide the GPU Video clock.

You have to use HWiNFO in order to find out the GPU Video Clock during benchmarking.

One last thing.

I didn't see LAV video in your codecs.
Don't you use it ?


8K Video from www.youtube.com/user/martinheck1

8K 29.970 fps - Youtube MP4 66 Mbps.
8K 60 fps - Youtube 30 fps VP9 15 Mbps. (Converted 60 fps with http://www.spirton.com/convert-videos-to-60fps/) + Hybrid "www.selur.de"
4K 60 fps - http://demo-uhd3d.com/fiche.php?cat=uhd&id=152 (Converted 60 fps to 180 fps)

Sources Converted to HEVC Main10. (used Hybrid "www.selur.de")


I Don't know Both usage. Sorry ( HD 530 + Nvidia )

I checked with HWiNFO64. GPU Video Clock 1700 MHz. http://sendpic.org/view/1/i/dUPrpylf2BURgwk8zZwvJKui7IZ.png

I don't love MPC-HC. MPC-BE is better than MPC-HC.

Note: My upload speed 0.65 mbps. i don't make upload :(

I Know lav filters :)

[GTPVHD]

Unless someone with a GP102/GP104/GP106 running the latest 378.78 or later drivers & DXVAChecker posts a screenshot showing VP9_VLD_10bit_Profile2 like the link below.

http://forum.doom9.org/showthread.php?p=1787451#post1787451

I would find it hard to believe those GPUs suddenly picked up VP9 Profile2 support.

[NikosD]

Thanks for the info.


I Don't know Both usage. Sorry ( HD 530 + Nvidia )


What do you mean ?
You just plug a cable to HD 530 output and your monitor and run the tests selecting HD 530 from DXVA Checker and then Pascal 1060 :)


I checked with HWiNFO64. GPU Video Clock 1700 MHz. http://sendpic.org/view/1/i/dUPrpylf2BURgwk8zZwvJKui7IZ.png


Thanks.

[huhn]

Unless someone with a GP102/GP104/GP106 running the latest 378.78 or later drivers & DXVAChecker posts a screenshot showing VP9_VLD_10bit_Profile2 like the link below.

http://forum.doom9.org/showthread.php?p=1787451#post1787451

I would find it hard to believe those GPUs suddenly picked up VP9 Profile2 support.

only newer cards like the GP107 have this decoder this is no secret. maybe the 1080 ti will get it but it is still a GP102 so i don't think so.

* HEVC/VP9 10/12 bit decoding SW support coming in Video Codec SDK 8.0
** Max resolution support is limited to selected Pascal chips
*** VP8 decode support is limited to selected Pascal chips
**** VP9 10/12 bit decode support is limited to select Pascal chips

[loekf]

only newer cards like the GP107 have this decoder this is no secret. maybe the 1080 ti will get it but it is still a GP102 so i don't think so.

AFAIK, the release notes for the latest driver show:

http://us.download.nvidia.com/Windows/378.78/378.78-win10-win8-win7-desktop-release-notes.pdf

New Features
* Added support for the following products:
GeForce GTX 1080 Ti
GeForce GTX 1050 notebook
GeForce GTX 1050 Ti notebook
* Video SDK 8.0
High-bit-depth (10/12-bit) decoding (VP9/HEVC)
OpenGL input surface support for encoder
Weighted Prediction
H.264 ME-only mode enhancements

All Pascal based cards have the same video decoder, so GP102 - GP107. Difference is in clock speeds, #shaders, ROPs and memory size & speed.

I have never seen files/streams in VP9 format, I know Youtube serves them if you use Chrome/Edge/FF in some cases.

Found this one:

https://www.vcodex.com/news/a-nice-example-of-4k-video-encoded-using-vp9/

It's 4K, but ~13 Mbps, not very challenging... should be ...

[nevcairiel]

All Pascal based cards have the same video decoder, so GP102 - GP107.


They do not. Not sure how you draw that conclusion from the release notes, it doesn't mention anything about hardware support of those features either way.
The only semi-official word we received was that GP104 does not have VP9 10-bit support, but it was also claimed that GP106 does, which is either still disabled in current drivers, or wrong information - in which case only the GP107 in the 1050 has VP9 10/12-bit support.

[loekf]

They do not. Not sure how you draw that conclusion from the release notes, it doesn't mention anything about hardware support of those features either way.
The only semi-official word we received was that GP104 does not have VP9 10-bit support, but it was also claimed that GP106 does, which is either still disabled in current drivers, or wrong information - in which case only the GP107 in the 1050 has VP9 10/12-bit support.

Yeah, read the stuff on either videohelp or hardocp. Still puzzling why a GP107 would do 10-bits VP9 and a more expensive sibling like GP104 or GP106 would not.

Maybe it's tapeout related. GP107 got introduced after GP104 and GP106, so perhaps GP107 has video IP with a higher revision number (bug fixes).

EDIT: when I go this Chinese site DXVA checker reports VP9 profile 0 decode for GTX 1060 (= GP106):

https://www.mobile01.com/topicdetail.php?f=298&t=4866089

https://attach.mobile01.com/attach/201607/mobile01-3988607dd710f26cad9bbd49cf1b5bf0.jpg

And GTX1070, via https://forums.geforce.com/default/topic/989353/geforce-1000-series/cannot-get-hevc-x265-to-work-under-any-circumstances-could-i-have-a-hardware-defect-/

http://i.imgur.com/dTweKQE.jpg

and GTX1080:

http://www.notebookcheck.net/fileadmin/Notebooks/Sonstiges/Grafikkarten/NVidia/GTX1000/dxvachecker.png

So:

GP106 (GTX 1060), GP107 (GTX1050): support VP9
GP104 (GTX1070 and GTX1080): should be, think just rumours. Makes no sense if GTX1070 can do VP9, but GTX1080 can not.
GP102 (Titan etc, GTX1080 TI): got introduced later, should be

[wanezhiling]

when I go this Chinese site DXVA checker reports VP9 profile 0 decode for GTX 1060 (= GP106):
That's 8-bit VP9, not 10-bit.


Look at GP107 here: http://forum.doom9.org/showthread.php?p=1787451#post1787451

[CruNcher]

Yeah, read the stuff on either videohelp or hardocp. Still puzzling why a GP107 would do 10-bits VP9 and a more expensive sibling like GP104 or GP106 would not.

Maybe it's tapeout related. GP107 got introduced after GP104 and GP106, so perhaps GP107 has video IP with a higher revision number (bug fixes).

EDIT: when I go this Chinese site DXVA checker reports VP9 profile 0 decode for GTX 1060 (= GP106):

https://www.mobile01.com/topicdetail.php?f=298&t=4866089

https://attach.mobile01.com/attach/201607/mobile01-3988607dd710f26cad9bbd49cf1b5bf0.jpg

And GTX1070, via https://forums.geforce.com/default/topic/989353/geforce-1000-series/cannot-get-hevc-x265-to-work-under-any-circumstances-could-i-have-a-hardware-defect-/

http://i.imgur.com/dTweKQE.jpg

and GTX1080:

http://www.notebookcheck.net/fileadmin/Notebooks/Sonstiges/Grafikkarten/NVidia/GTX1000/dxvachecker.png

So:

GP106 (GTX 1060), GP107 (GTX1050): support VP9
GP104 (GTX1070 and GTX1080): should be, think just rumours. Makes no sense if GTX1070 can do VP9, but GTX1080 can not.
GP102 (Titan etc, GTX1080 TI): got introduced later, should be

Market, Target, Design choices anyways Nvidias was allways very clever where and when they support something at the right time to market ;)

Also what endsup in the PC had no priority yet because of the Security Architecture missing, thus why exposing of features was always first in Tegra where the Security architecture is fully actively in Place.

See the new Andorid Shield Maxwell shrink i bet with you it has the full GP107 VPx features ;)

In general Nvidias strategy is always logical based on the overall Target Systems capabilities most of the time.

So the more powerfull VPX in the low end GPU consumer hardware tries to compensate for a weaker overall System ;)

And then they make a decission sometime if it makes sense for the target to bring some features on the higher end via the Shader capabilities (invest into additional Software development and R&D) (first then these come mainly for the Professional GPUs, with a lot of Cuda Power).

So Nvidia might still expose Hybrid 10 bit support especially on Titan X, 1080 TI and GP102, not sure though how much sense it really would make, so close before Volta with the complete overhaul.

Also VP9 has not that support internaly they fully working on AV1.

[NikosD]

Yeah, read the stuff on either videohelp or hardocp. Still puzzling why a GP107 would do 10-bits VP9 and a more expensive sibling like GP104 or GP106 would not.

Maybe it's tapeout related. GP107 got introduced after GP104 and GP106, so perhaps GP107 has video IP with a higher revision number (bug fixes).

EDIT: when I go this Chinese site DXVA checker reports VP9 profile 0 decode for GTX 1060 (= GP106):

GP106 (GTX 1060), GP107 (GTX1050): support VP9
GP104 (GTX1070 and GTX1080): should be, think just rumours. Makes no sense if GTX1070 can do VP9, but GTX1080 can not.
GP102 (Titan etc, GTX1080 TI): got introduced later, should be

I really like guys who like to search for sources, googling in various sites, read Nvidia's (AMD's, Intel's) docs etc.

I'm one of them actually.

But I have to tell you sincerely, that the greatest source of almost every publicly available info regarding those issues is here and sometimes months before anyone else write about it ;)

You must read carefully and read more of this thread in order to get replies and solve mysteries regarding SW/HW video decoding.

And obviously I don't mean only myself, I just happened to start this thread.

[loekf]

You must read carefully and read more of this thread in order to get replies and solve mysteries regarding SW/HW video decoding.

And obviously I don't mean only myself, I just happened to start this thread.

Yeah, saw that too. Was too lazy to read through all 90+ pages. Appearently it was already known that Nvidia is doing strange things with their video IP (put a more advanced decoder in your later introduced low-end product). As someone also mentioned, high-end GPUs are targetting gamers, who are mostly interested in graphics' fps, not HEVC fps...

[mixus]

Thanks for the info.



What do you mean ?
You just plug a cable to HD 530 output and your monitor and run the tests selecting HD 530 from DXVA Checker and then Pascal 1060 :)



Thanks.

Hi bro.

HD 530 is very bad gpu. Because cannot play 4K main10 HEVC 30p or 60p.

[mixus]

I Installed Lastest Driver. 378.78

http://i.hizliresim.com/4PbBap.png

Note : VLC 3 now playable HEVC 4K and 8K.

http://i.hizliresim.com/m31YXP.png

http://i.hizliresim.com/j8Vkqj.png

[NikosD]

Hi bro.

HD 530 is very bad gpu. Because cannot play 4K main10 HEVC 30p or 60p.
It doesn't have 10bit HEVC decoder.

But you could test 8bit HEVC clips only.

[mixus]

It doesn't have 10bit HEVC decoder.

But you could test 8bit HEVC clips only.

For you :


INTEL HD 530 - HEVC 8 Bit

HEVC 4K MAIN 30 FPS 24 mbps --- Clip Name : LG_4K_OLED_Wonders
Decoder: Microsoft H265 Video Decoder MFT
Decoder Device: HEVC_VLD_Main
Frames: 7110
Time: 26,657 sec
FPS: 266,724 [235-301] fps
CPU Usage: 1 [0-3] %
GPU Usage: -

HEVC 4K MAIN 60 FPS 41 mbps --- Clip Name : Samsung_SUHD_Sensational_Picture
Decoder: Microsoft H265 Video Decoder MFT
Decoder Device: HEVC_VLD_Main
Frames: 7110
Time: 27,857 sec
FPS: 255,228 [131-290] fps
CPU Usage: 1 [0-5] %
GPU Usage: -

HEVC 8K MAIN 30 FPS 140 mbps --- Clip Name : NEW ZEALAND ASCENDING "Youtube"
Decoder: Microsoft H265 Video Decoder MFT
Decoder Device: HEVC_VLD_Main
Frames: 3620
Time: 54,951 sec
FPS: 65,877 [59-77] fps
CPU Usage: 1 [0-3] %
GPU Usage: -


Note:
MPC-HC + LAV Filters 0.69 (Copy-Back and Native)
Not played 8K Main 30p

Extra:
http://i.hizliresim.com/695AzP.png

[NikosD]

For you :

Extra:
http://i.hizliresim.com/695AzP.png

Oh man.

You really like doing things your own way!

Thanks for all this info, but we have here some specifics clips that I told you to test, using LAV video actually.

There is a collection of 8bit HEVC and 10bit HEVC clips in order for all of us to test the same clips.

Nevermind, thanks.

You made me smile :D

[mixus]

Oh man.

You really like doing things your own way!

Thanks for all this info, but we have here some specifics clips that I told you to test, using LAV video actually.

There is a collection of 8bit HEVC and 10bit HEVC clips in order for all of us to test the same clips.

Nevermind, thanks.

You made me smile :D

:D


I5 6600K 4.4 GHz - VP9 8K Profile 0 15 mbps
http://i.hizliresim.com/1LMzY1.png

I5 6600K 4.4 GHz - H264 8K 66 mbps
http://i.hizliresim.com/Ly19pJ.png

INTEL HD 530 - 4K 120 fps HEVC Main 8 Bit 223 Mbps
http://i.hizliresim.com/r3ZPQB.png

INTEL HD 530 - 4K 120 fps HEVC Main 8 Bit 223 Mbps
http://i.hizliresim.com/7qMRkm.png

INTEL HD 530 - 4K 120 fps HEVC Main 8 Bit 223 Mbps
http://i.hizliresim.com/P0PXqv.png


Final

HD 530 Optimized HEVC 4K Main 60 Fps and AVC 4K 60 Fps.

Not Ready Main10.

[mixus]

GTX 1060 6 GB - 4K 120 fps HEVC Main 8 Bit 223 Mbps
http://beta.hizliresim.com/ldvnvb

[P.J]

EVGA GeForce GTX 960 SuperSC ACX 2.0+ 2GB (GPU Video Clock: 1202.3 MHz)
Win 10 x64 1607, Forceware 378.78, LAV x64 0.69.0.14

1) LG_4K_View-the-Feeling (2160p@29.970fps ~25Mbps)

Decode: 137/145/163
Playback: 141/145/165

2) Samsung 4K@30fps-80Mbps file -0235

Decode: 119/120/122
Playback: 121/123/126

3) Samsung 4K@30fps-80Mbps file -0230

Decode: 114/115/117
Playback: 116/116/121

4) Crowd_run 4K@50fps -300Mbps

Decode: 83/93/97
Playback: 83/94/98

5) 1st Bonus clip:
Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit

Decode: 48/59/70
Playback: 47/59/71

1.UHD_PQ_Lovely_Swiss-51Mbps@60fps

Decode: 112/124/153
Playback: 111/122/155

2.Crowd_run_2160p50_Main10_100Mbps

Decode: 108/109/113
Playback: 108/112/115

3.Jellyfish-140-mbps-4k-uhd-hevc-10bit

Decode: 94/103/114
Playback: 92/102/113

4.Jellyfish-400-mbps-4k-uhd-hevc-10bit

Decode: 38/55/71
Playback: 44/55/72

[NikosD]

Perfect !

Thank you.

[Kadano]

Hi all,

I recently bought an Intel G4560 because of its amazing performance per price ratio. However, the one potential problem I see with it is the integrated graphics HD 610, which seems to be half of a HD 630 used in all Kaby Lake CPUs that are sold for 85€ or more.

Is it known whether the HD 610 is also chopped down in encoding / decoding ability or is the performance only reduced for 3D / gaming applications?
If it is not known, I'd like to help by running tests on the G4560.

[NikosD]

l
Is it known whether the HD 610 is also chopped down in encoding / decoding ability or is the performance only reduced for 3D / gaming applications?
If it is not known, I'd like to help by running tests on the G4560.

The 3D power of Intel's IGPU plays a significant role in HW encoding, because Intel uses hybrid encoding and not pure fixed function (actually it has a choice of pure fixed function but it's crippled)

But in HW pure decoding only the clock really matters *

In playback benchmark mode and realtime playback, 3D power plays a role in scaling playback and also using EVR CP or even heavier renderers, like MadVR.

Anyway, I'm personally more than interested in testing your IGPU in all video samples using both decode and playback mode.

If you do it, post your results here or send me a PM.

* There are cases that the higher GPU models have more HW units than the lower but I think it only matters for encoding not decoding (they are encoders actually)

[Kadano]

Anyway, I'm personally more than interested in testing your IGPU in all video samples using both decode and playback mode.

If you do it, post your results here or send me a PM.
How do I run decode / playback benchmark with LAV Filters? Do I need to set some flag in PotPlayer / mpc-hc? I've looked through your posts but I can't find any information on how to prepare the testing settings.

[Kadano]

How do I set up the testing environment? I've looked through your post, but wasn't able to find any details.

In DXVA Checker > "DirectShow/MediaFoundation Decoder", do I set the Target to "DS + MF (32bit)" or something else? Also, after "Re-Check (Connect to Renderer)" on LAV Video Decoder, the text turns red. Does this mean there is something wrong?

Also, do I use 32bit or 64bit for playback?

[NikosD]

OK here we go:

Use LAV Video x64 one of the latest versions along with DXVAChecker v3.15 and you select "DS + MF 64bit" for all benchmark modes (decode & playback)

For playback your selection should be 1280 x 720.

The LAV video turning red means success.
It successfully connected to the renderer (EVR) using DXVA2.

Run each of the 12 files at least 4 times for each mode and write down the average of min/avg/max, usually excluding the first run because it loads the file.

Be extra careful to disable everything from the Intel's control panel, because Intel by default enables a few options for "better" video quality.

Be sure to disable everything.

[CruNcher]

This is really interesting

http://images.anandtech.com/doci/11229/cropped_evrcp_575px.png
http://images.anandtech.com/doci/11229/cropped_kodi_575px.png


Though i see absolutely no information about Anandtechs Testing Methodology here on the EVR-CP side it is MPC-HC though Fullscreen Ex or Fullscreen Windowed on Win 10

for UHD it is definitely a master difference at those 12W @ peak between glitch free playback and not on Kodi with a peak GPU load of ~5% at ~7W

[wanezhiling]

No doubt about it, vanilla EVR is always the best&only choice for weak platforms.

[CruNcher]

yeah also comes close on my own cherry trail results but there the overhead also came from the gui rendering and timers of MPC-BE, especially the seekbar was critical even with EVR.

So i wonder if Anandtechs result is with any gui rendering or without

also in case of MadVR im not sure didn't test it on low power but with the options all on lowest and DXVA all the time it might be ok but in general then still the CPU overhead is higher anyways then even EVR-CP.