Evaluation of HEVC decoders (SW, Hybrid and HW)

[edwdevel]

http://jpegshare.net/d2/3a/d23a630cf160163c5057946f6df5056c.jpg.html

[edwdevel]

Here is an MPC-BE screenshot of sony camp:

http://jpegshare.net/37/88/3788aa4cd0ee38811ba7290c84f9e884.jpg.html

[NikosD]

NikosD, I did test the arri 7K clip, as it could be downloaded quickly, but the others are very large and the servers refused fast downloading. The "girls" video takes a long time to benchmark and this cannot be done at work :)

Post a link to your 4k VP9 videos and I'll try to view them.

This is a link with all of my samples listed in a server with a max download of 1MB/s
ftp://helpedia.com/pub/multimedia/testvideos/HEVC_HW_decoders_clips/

Take a look here:
ftp://helpedia.com/pub/multimedia/testvideos/HEVC_HW_decoders_clips/8K_Ex_sample.mp4

It's a 8K sample with only 133MB footprint on your disk.

I would like also you to test this 4K clip with huge 700Mbps bandwidth.
It's only 787MB, take a look here:
ftp://helpedia.com/pub/multimedia/testvideos/HEVC_HW_decoders_clips/Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit.mkv

OK, here is a link of YouTube with 4K60fps in VP9 codec:
https://www.youtube.com/watch?v=ppy_gjBp71E

You have to manually select 4K60 fps from the "Settings"->"Quality" -> 2160p60 (4K)

Now, regarding your tests and your DXVA Checker screenshot, it seems that you have the full-featured video decoder of Pascal 1050/1050 Ti chips regarding features but the decoding speed of 1030 is between AMD RX 470 and Nvidia 960.

Something like that: (regarding decoding speed)
RX 470 < GT 1030 < GTX 960

Your 2GB VRAM is a no go for Netflix 4K streaming, which requires >3GB VRAM

Also, as others have already mentioned here, using madVR with 4K content requires >2GB VRAM, especially for the clips you mentioned to test like "Exodus" and "Life of Pi" clips which are HDR.

Did you find the "GPU Video Clock" ?

[v0lt]

Here is an MPC-BE screenshot of sony camp:

http://jpegshare.net/37/88/3788aa4cd0ee38811ba7290c84f9e884.jpg.html
Thank you.
The GT 1030 is enough to play 4k HEVC HDR 60fps video. :)

[huhn]

why should there be a problem at FHD?

[mrcorbo]

Seems to work in Edge on my GTX 1060. Tested up to 8k VP9 w/ 10% CPU usage and no dropped frames. Chrome can't even do this, actually, at least not on my system. Same video pegs my CPU and drops frames like crazy.

Wanted to update this to say that YouTube HDR clips are also playing in Edge in HDR if you turn on HDR mode in Windows (and you have an HDR TV/monitor, obviously).

In "Stats for Nerds", Codec is reported as VP9.2 and color is reported as smpte2084(PQ)/bt2020.

This as of Windows 10 Insider Build 16232. Not sure if it works on any earlier version.

Notably, on my 1060 this seems to be using hybrid acceleration and at 60fps drops frames like crazy. It can mange 24fps (at least) smoothly, though.

[NikosD]

Notably, on my 1060 this seems to be using hybrid acceleration and at 60fps drops frames like crazy. It can mange 24fps (at least) smoothly, though.

Only 1030/1050/1050 Ti cards have fixed-function HW decoder for 10bit VP9, used by YouTube for HDR clips.

And Kabylake from Intel of course.

I think 1060 doesn't even have a hybrid 10bit VP9 decoder.
It must be CPU only.

[jmonier]

Only 1030/1050/1050 Ti cards have fixed-function HW decoder for 10bit VP9, used by YouTube for HDR clips.

And Kabylake from Intel of course.

I think 1060 doesn't even have a hybrid 10bit VP9 decoder.
It must be CPU only.

The 1060 DOES have a HW decoder.

[NikosD]

The 1060 DOES have a HW decoder.
We are talking about VP9 10bit.

Are you following the conversation ?

[v0lt]

@jmonier
GTX 1050 (https://forum.doom9.org/showthread.php?p=1802947#post1802947) vs GTX 1060 (https://forum.doom9.org/showthread.php?p=1802971#post1802971)
The GTX 1060 does not have a VP9_VLD_10bit_Profile2, but it does exist in the GTX 1050. Need or do not need this feature, this is another question.

[NikosD]

Latest Nvidia drivers are out - 384.76 - and it seems that Nvidia kept its promise, eventually.

They added support of DX12 on Fermi cards, although it's just support not optimized drivers for DX12.

https://s4.postimg.org/xzc6qs72l/GTX-580-_Ge_Force-384.76-_DX11-vs-_DX12.jpg

[NikosD]

OK, it seems that 384.76/384.80 drivers of Nvidia, enabled one more undocumented feature - besides DX12 support on Fermi - and that is Netflix UHD support on Pascal cards.

If someone has a 1030/ 1050 which have <3GB VRAM could try a Netflix UHD account and tell us if it works.

You need Netflix Windows Store (UWP platform) app or Edge browser, Windows 10 and HDCP 2.2 4K monitor.

http://techreport.com/news/32195/nvidia-384-76-drivers-quietly-enable-uhd-netflix-on-pascal

https://www.techpowerup.com/234882/nvidia-geforce-384-series-driver-removes-need-for-new-cpus-for-4k-netflix

[NikosD]

It is the first time I see Video Decoding usage of Intel's iGPU outside of Intel's own tools (GPA monitor)

https://s9.postimg.org/cg72m9bin/small_gpu_monitor.jpg

[P.J]

It is the first time I see Video Decoding usage of Intel's iGPU outside of Intel's own tools (GPA monitor)

I can't find anything in GPA monitor :confused:

[NikosD]

Do you have GPA monitor installed ?

It's a special tool from Intel, it's not in the drivers.

You have to read the manual.

[P.J]

Do you have GPA monitor installed ?

It's a special tool from Intel, it's not in the drivers.

You have to read the manual.

Yea but it doesn't show video engine load, only fps.

[aufkrawall]

I tried out using the Skylake IGP for decoding with rendering and signal output via dGPU, and it worked like a charme on Linux with mpv and vaapi-copy (mpv btw. worked great in general with the IGP on Linux).
On Windows, we are still limited to the old QuickSync decoder in LAV Filters which doesn't support HEVC 10 bit and VP9, right? So Kabylake users with dGPU are cut off their best video functionality...

[NikosD]

I think that DXVA copy-back can do that using recent versions of LAV filters, but nevcairiel should reply for sure.

[nevcairiel]

You can do that if the GPU is active. It needs a connected screen for that.

[aufkrawall]

I think that DXVA copy-back can do that using recent versions of LAV filters, but nevcairiel should reply for sure.
Only D3D11VA-CB could do that, but nev stated it'd be really ugly for him to implement.
mpv has D3D11VA-CB support, but it's slow and it seems to be lacking support for using different devices.

You can do that if the GPU is active. It needs a connected screen for that.
It wouldn't be enough to simply add a second cable that connects the display with the IGP and dGPU at the same time, right? If I'm not mistaken, there needs to be an actual signal send to the display?

[huhn]

even a fake display is enough so there is no need to send a signal.

far from perfect but better than nothing.

[aufkrawall]

It always annoyed me to the death that the mouse cursor could leave the real display and got lost on the fake one.

[pankov]

aufkrawall,
I think it was since the Creators Update that if you position your fake display in the bottom right corner of a real one so they have only one corner touching (one vertex to be precise) the mouse won't be able to pass to the fake one.

---------------
| |
| R |
| |
---------------
---------------
| |
| F |
| |
---------------

[huhn]

this works but this doesn't change that it is not practical for the AVG user.

[edwdevel]

It's a 8K sample with only 133MB footprint on your disk.

I would like also you to test this 4K clip with huge 700Mbps bandwidth.

OK, here is a link of YouTube with 4K60fps in VP9 codec:
https://www.youtube.com/watch?v=ppy_gjBp71E

Your 2GB VRAM is a no go for Netflix 4K streaming, which requires >3GB VRAM

Also, as others have already mentioned here, using madVR with 4K content requires >2GB VRAM, especially for the clips you mentioned to test like "Exodus" and "Life of Pi" clips which are HDR.

Did you find the "GPU Video Clock" ?

I've edited my 29th June 2017, 18:52 forum entry with the "tunnelflag" 700Mbs clip and the "8Kexsamp" 8K sample.

I've also done some benchmarking on the vp9 youtube clips, and have done more investigation on different platforms using the vp9 codec on youtube. To get to the point with the gtx 1030, I am working with my old 2core3.0Ghz wolfdale and Windows7, and it seems that the GPU is not being used for decoding under Chrome, as the cpu is maxed out at 1080p vp9, although there is some GPU activity (23%), and there are many dropped frames at 1080p. Furthermore, it seems that the coding efficiency is not too good for vp9, as even 1080p seems to require around 50Mbs to run without frame drops. Here at work with the 1030 and around 50Mbs max internet (if that) I must drop down to 720p before I have no dropped frames (this is with your provided clip).

Contrast the above with my home system, which is a sandybridge 4core3.4Ghz system with an Nvidea GT 970. My internet speed at home is 80Mbs+, and 1080p is flawless although the cpu is running high (50%). Even 4K can probably be decoded by the cpu, but the bitrate required is <100Mbs for the vp9 clip, so I run out of internet bandwidth.

The bottom line is that it is youtube and chrome that dictate the efficiency of the download. Google should immediately convert the download of all of it's content to HEVC, and jettison vp9. Then my poor old PC at work with its good-enough gt 1030 can view all of these clips at a reasonable internet bit-rate. :)

I think it's clear that 2G is enough to decode *and* downsize basically any 4k clip. If netflix needs more than this it has its' own reasons, but certainly technically the 1030 can do the job.

I've also done some madvr investigation, and believe 2G is still adequate as far as memory is concerned, but madvr is a completely different animal as far as renderers go and basically it acts like a 3D game that needs as much GPU power as you can afford. Still, I can run both the move clips (exodus and lifeofpi) using madvr (via MPC-BE) with the GT 1030 perfectly, but clearly the 1030 is overloaded at 60fps with, for instance, the "Sony Camp" clip. I'll comment more on what I've discovered in response to a different reply to my June29 comment.

[el Filou]

I've also done some benchmarking on the vp9 youtube clips, and have done more investigation on different platforms using the vp9 codec on youtube. To get to the point with the gtx 1030, I am working with my old 2core3.0Ghz wolfdale and Windows7, and it seems that the GPU is not being used for decoding under Chrome, as the cpu is maxed out at 1080p vp9, although there is some GPU activity (23%), and there are many dropped frames at 1080p. Furthermore, it seems that the coding efficiency is not too good for vp9, as even 1080p seems to require around 50Mbs to run without frame drops. Here at work with the 1030 and around 50Mbs max internet (if that) I must drop down to 720p before I have no dropped frames (this is with your provided clip).
It looks like Chrome wants/needs (maybe a dev can clear that out?) at least Windows 10 1607 or later to enable VP9 acceleration: https://codereview.chromium.org/2182263002/

On my own 2-core Wolfdale under Windows 10, YouTube VP9 is perfectly smooth in 1080 with absolutely no frame drops whatsoever with a 30 Mbps connection. 2160 is a bit too much for my connection, but if I buffer for long enough it will also play without any frame drops.

[huhn]

2G for madVR should be totally save for 1080p output.

madVR uses an advanced HDR-SDR conversation which is maybe to much for a 1030 even at 1080p but a 1030 should be able to do 60 FPS with the correct settings.

and why should google change from VP9 too HEVC your card should be able to hardware decode it in chrome and they need to pay for HEVC unlike VP9.

[aufkrawall]

aufkrawall,
I think it was since the Creators Update that if you position your fake display in the bottom right corner of a real one so they have only one corner touching (one vertex to be precise) the mouse won't be able to pass to the fake one.

---------------
| |
| R |
| |
---------------
---------------
| |
| F |
| |
---------------
Thx, good to know. Let's just hope this is by design and not actually a bug.

[NikosD]

To get to the point with the gtx 1030, I am working with my old 2core3.0Ghz wolfdale and Windows7, and it seems that the GPU is not being used for decoding under Chrome, as the cpu is maxed out at 1080p vp9, although there is some GPU activity (23%), and there are many dropped frames at 1080p.

Your OS is too old, at least for Microsoft and probably Google, you need to upgrade to Win 10 Creators Update in order to HW accelerate VP9 8bit/10 bit.

Using Windows 10 you could try even Edge - the Win 10 only browser of Microsoft - which probably can HW accelerate VP9 too.

Furthermore, it seems that the coding efficiency is not too good for vp9, as even 1080p seems to require around 50Mbs to run without frame drops. Here at work with the 1030 and around 50Mbs max internet (if that) I must drop down to 720p before I have no dropped frames (this is with your provided clip).

Contrast the above with my home system, which is a sandybridge 4core3.4Ghz system with an Nvidea GT 970. My internet speed at home is 80Mbs+, and 1080p is flawless although the cpu is running high (50%). Even 4K can probably be decoded by the cpu, but the bitrate required is <100Mbs for the vp9 clip, so I run out of internet bandwidth.

The bottom line is that it is youtube and chrome that dictate the efficiency of the download. Google should immediately convert the download of all of it's content to HEVC, and jettison vp9. Then my poor old PC at work with its good-enough gt 1030 can view all of these clips at a reasonable internet bit-rate. :)


You are totally wrong here.

I have a VDSL2 (FTTC) true 50Mbps line and I can easily keep up with the 4K60fps stream of the clip provided above by me.

For 1080p the required bandwidth is below <20Mbps.

Probably there is something wrong with your line - maybe it's not real 80Mbps.

I can't see something wrong with Google servers or VP9, it must be your side.


I think it's clear that 2G is enough to decode *and* downsize basically any 4k clip. If netflix needs more than this it has its' own reasons, but certainly technically the 1030 can do the job.


I don't own a 4K monitor, it's only 1080p, but from what I have read here the problem is when you try to decode and display 4K clips on 4K native resolution without downsizing - this is where you need more than 2G VRAM - and probably the same goes for Netflix.

[NikosD]

I've edited my 29th June 2017, 18:52 forum entry with the "tunnelflag" 700Mbs clip and the "8Kexsamp" 8K sample.


Your two new results, place 1030 above 960 and closer to 1050 and even a little above for "tunnel" clip.

Did you change something (driver, setting) ?

Maybe you could run again some of your previous tests with your new configuration or it is something coincidental.

[edwdevel]

Your two new results, place 1030 above 960 and closer to 1050 and even a little above for "tunnel" clip.

Did you change something (driver, setting) ?

Maybe you could run again some of your previous tests with your new configuration or it is something coincidental.

Hmm. It looks like the latest Nvidia drivers are far better than the previous ones- I've edited again my my 29th June 2017, 18:52 forum entry with new benchmarks for the 384.76 drivers.

I think that is the only change to the June29 setup.

The new benchmarks seem 50% faster. I think everyone with a Pascal benchmark should re-do them with 384.76 to ensure there is not something strange in my setup :)

[edwdevel]

It looks like Chrome wants/needs (maybe a dev can clear that out?) at least Windows 10 1607 or later to enable VP9 acceleration: https://codereview.chromium.org/2182263002/

On my own 2-core Wolfdale under Windows 10, YouTube VP9 is perfectly smooth in 1080 with absolutely no frame drops whatsoever with a 30 Mbps connection. 2160 is a bit too much for my connection, but if I buffer for long enough it will also play without any frame drops.

Yes, I'm sure you're right about Chrome needing Windows10 for hardware VP9 decoding. I wonder if it has something to do with the DirectX version? At any rate, if Google wanted, it means they could make it work in Windows7 as well, it's just a matter of desire.

As far as how youtube does VP9 streaming is concerned, it is difficult to know just what the actual bitrate of the clip is from the "stats for nerds" that youtube provides. Apparently youtube just buffers the desired video at the maximum bit rate the user can support and then dribbles it out as needed. But obviously, if the average bit rate at the desired resolution (say 1080p or 4k) cannot be sustained, then the video will drop frames no matter how fast the cpu or gpu are. Furthermore, as I have discovered and is also obvious, different clips have different required bit rates. We should keep our discussion to the clip NikosD provided- https://www.youtube.com/watch?v=ppy_gjBp71E, which happens to be a very difficult 4k video requiring a high bit rate. Other VP9 4K videos obviously have a much lower bit rate requirement.

To compare actual bit-rate requirements for different videos, we need VP9 4K 10 bit downloadable content to actually calculate the bit-rates. I haven't seen any such videos to test, but I haven't also looked very hard :)



r

[NikosD]

Yes, I'm sure you're right about Chrome needing Windows10 for hardware VP9 decoding.

As far as how youtube does VP9 streaming is concerned, it is difficult to know just what the actual bitrate of the clip is from the "stats for nerds" that youtube provides. Apparently youtube just buffers the desired video at the maximum bit rate the user can support and then dribbles it out as needed. But obviously, if the average bit rate at the desired resolution (say 1080p or 4k) cannot be sustained, then the video will drop frames no matter how fast the cpu or gpu are.


Did you see my previous reply regarding those issues ?

[edwdevel]

2G for madVR should be totally save for 1080p output.

madVR uses an advanced HDR-SDR conversation which is maybe to much for a 1030 even at 1080p but a 1030 should be able to do 60 FPS with the correct settings.

and why should google change from VP9 too HEVC your card should be able to hardware decode it in chrome and they need to pay for HEVC unlike VP9.

For 1080p madvr, the 2G memory of the GT 1030 is perfectly adequate (at least for the clips I have tested). It is 4k 60fps HEVC that is questionable.

We need HEVC for the simple reason that it is more efficient, so that *I can view the same video at the same quality* for half the bit-rate, which is critical because for most people internet bandwidth is finite. I know Google doesn't care about this, but all of its' users do. :)

[sneaker_ger]

To compare actual bit-rate requirements for different videos, we need VP9 4K 10 bit downloadable content to actually calculate the bit-rates. I haven't seen any such videos to test, but I haven't also looked very hard :)
Look no further:
https://www.youtube.com/watch?v=tO01J-M3g0U (download using youtube-dl)

[huhn]

We need HEVC for the simple reason that it is more efficient, so that *I can view the same video at the same quality* for half the bit-rate, which is critical because for most people internet bandwidth is finite. I know Google doesn't care about this, but all of its' users do. :)

HEVC isn't even remotely close to 2x compression efficiency compared to VP9.

[edwdevel]

Look no further:
https://www.youtube.com/watch?v=tO01J-M3g0U (download using youtube-dl)

Have you actually downloaded this file? I'm having trouble getting it: I keep getting the H264 720*1280 version. Please download it yourself, ensure that it is in fact the 4K VP9 version, play it using MPC-BE or DXVA checker, and then let me know how you did it :).

[nevcairiel]

We need HEVC for the simple reason that it is more efficient, so that *I can view the same video at the same quality* for half the bit-rate, which is critical because for most people internet bandwidth is finite. I know Google doesn't care about this, but all of its' users do. :)

HEVC and VP9 are in a similar efficiency range, you won't get any significant savings from using HEVC over VP9 - if any at all.

[sneaker_ger]

Have you actually downloaded this file?
It is not one file. Youtube provides 41 different streams for this video. You can get a list via
youtube-dl -F https://www.youtube.com/watch?v=tO01J-M3g0U

Then download e.g stream #337 (4kp60 VP9 10 bit HDR):
youtube-dl -f 337 https://www.youtube.com/watch?v=tO01J-M3g0U

https://rg3.github.io/youtube-dl/

[stax76]

I'm also using youtube-dl, great tool, here is a great Firefox extension which can download it too:

https://addons.mozilla.org/de/firefox/addon/youtube-video-and-audio-dow/?src=api

[NikosD]

Hmm. It looks like the latest Nvidia drivers are far better than the previous ones- I've edited again my my 29th June 2017, 18:52 forum entry with new benchmarks for the 384.76 drivers.

I think that is the only change to the June29 setup.

The new benchmarks seem 50% faster. I think everyone with a Pascal benchmark should re-do them with 384.76 to ensure there is not something strange in my setup :)

Now that the driver is tuned enough and your results look close to 1050/1050 Ti, I suggest you to download everything from here:
ftp://helpedia.com/pub/multimedia/testvideos/HEVC_HW_decoders_clips/

and test them with latest LAV x64 filters and DXVA Checker in both decode and playback mode of 1280x720 scaling in order to put your results here in the table:
https://forum.doom9.org/showthread.php?p=1799099#post1799099

[clsid]

What is the current state of decoding on AMD GPUs with recent drivers? Does HEVC work properly now? Have the interlaced video problems been fixed?

[NikosD]

You can take a look at my links regarding performance of HW decoding on both H.264 & H.265.

It is more than fast enough for anything meaningful e.g Blu-ray UHD (not corner cases like 8K)

HEVC never had a problem using Polaris card, although it was a little bit on the edge regarding performance.

But not anymore, new drivers solved that "problem".

Now, regarding H.264 interlaced content, my last test using latest drivers for AMD, Nvidia, Intel and latest LAV x64 filters gave me a little surprise.

AMD has now exactly the same behavior like Nvidia and Intel.

So, H.264 interlaced is definitely fixed too.

[clsid]

Good news. Thanks. Since which driver version has the performance increased? 17.4 or later?

[NikosD]

Yes, around there 17.4.x

[edwdevel]

HEVC and VP9 are in a similar efficiency range, you won't get any significant savings from using HEVC over VP9 - if any at all.

Please understand I am not interested at all in proving HEVC is better than VP9 or vice versa. I was under the impression, from basic scientific consensus, that in encoding efficiency HEVC produces about half the bit rate of VP9 for videos at the 4k level. Just google "HEVC vs VP9 efficiency" for many papers that show this. Here is one:

http://iphome.hhi.de/marpe/download/Performance_HEVC_VP9_X264_PCS_2013_preprint.pdf

But, as I've been shown a 4k video that is encoded under both HEVC and VP9, I will undertake to look at both and report back :)

[huhn]

it's from 2013 and the used encoder plays a major if not the most important role for compression.

what so ever in term of compression potential VP9 and HEVC are not far apart.

[nevcairiel]

Please understand I am not interested at all in proving HEVC is better than VP9 or vice versa. I was under the impression, from basic scientific consensus, that in encoding efficiency HEVC produces about half the bit rate of VP9 for videos at the 4k level. Just google "HEVC vs VP9 efficiency" for many papers that show this. Here is one:

http://iphome.hhi.de/marpe/download/Performance_HEVC_VP9_X264_PCS_2013_preprint.pdf

That paper uses PSNR to measure video. That might be "scientific", but its not any way to actually measure video quality accurately (which in itself is not something thats really "solved", but PSNR is one of the worse things to use). Early HEVC encoders achieved good PSNR by creating blury video, which was very obvious to any human viewer, but rated quite well in PSNR tests.

[littleD]

dxva output with newest intel drivers

[NikosD]

Well, it seems that I managed to come up to some conclusions regarding VP9 acceleration (hybrid or pure fixed-function) of the three well known browsers (Chrome, Firefox, Edge) using Youtube on different hardware.

First of all VP9 acceleration is supported by all three nowadays and it definitely works on recent hardware supporting fixed-function VP9 acceleration like Kabylake for Intel and 950/960 for Nvidia, along with all Nvidia Pascal cards of course.

Some browsers could have problems with 8K VP9 acceleration though, but that's a very new and extremely high resolution option of Youtube streams, added lately.

The other fact is that for Chrome and Chromium based browsers you need a very recent Win OS, which is Windows 10 RS1 (Anniversary Update with a release on August 2016) or later according to this:

https://groups.google.com/a/chromium.org/forum/#!topic/chromium-reviews/qSePjavyJsU

The first official stable Chrome version supporting VP9 acceleration is v54, released after Anniversary Update on November 2016.

Now, Edge as we know is a Windows 10-only browser so obviously you need a Windows 10 OS and Edge enabled VP9 acceleration in Windows Insider builds before Anniversary Update, but the first stable Edge version supporting VP9 acceleration is inside Windows 10 RS1 (Anniversary Update) according to this:

https://blogs.windows.com/msedgedev/2016/04/18/webm-vp9-and-opus-support-in-microsoft-edge/#aPHsQSA3aIvOpy38.97

I don't know when exactly Firefox enabled in stable releases the VP9 acceleration and what are its Win OS requirements, but it seems that using Windows 10 and latest versions like v55, it works.

VP9 hybrid acceleration was firstly supported by Intel iGPUs like Broadwell and Skylake and AMD added that functionality officially on December 2016 inside Crimson 16.12.1 drivers only for Chrome:

VP9 Decode Acceleration: 4K 60Hz GPU-Accelerated Video Streaming enabled on supported Google™ Chrome web browsers.

Requires supported Chrome™ web browser versions with Hardware Acceleration enabled.
Compatible with AMD Radeon™ GCN and Radeon RX 400 series enabled products on Windows® 7/8.1/10.


It says Windows 7/8.1 but as I told you that is wrong according to my sources.

Also, as I read, it's an OpenCL based solution not only for RX 400/RX 500 series but some previous generations, too.

The first driver didn't enable the VP9 hybrid acceleration by default and the second driver 16.12.2, fixed that according to AMD.

Fixed Issues

Chromium may fail to utilize hybrid decode for VP9 content.


The truth is that Chrome never enabled by default the VP9 hybrid acceleration for Polaris cards.

You have to force enable it using this switch:

chrome.exe --enable-accelerated-vpx-decode=2

The reason is probably low performance.

But for Intel and CPUs prior to Kabylake, things are even worse.

All of the VP9 hybrid acceleration capable CPUs, like Broadwell and Skylake, are not allowed to use VP9 hardware decoding of Chrome according to this patch:

https://groups.google.com/a/chromium.org/forum/#!topic/chromium-reviews/BghP-1eqyhg

I have no idea regarding Edge and Firefox support of Intel's hybrid VP9 acceleration.

For Polaris cards, the actual use of that somewhat crippled VP9 hybrid decoder, lasted only 4 months between the first driver support and Windows 10 Creators Update which broke VP9 acceleration.

So, for that reason, I have opened a thread in AMD forums for anyone interested to read and contribute to the common effort of an actual fix - if possible - of VP9 hybrid acceleration.

You can read and write your comments here:
https://community.amd.com/thread/218278

I have a Core i5 2400 right now with a RX 470/8GB card and a real 50Mbps VDSL2 (FTTC) line, testing 4K60 fps clips and I think AMD should have kept its promise of offering VP9 4K60fps support using Polaris cards.