Evaluation of HEVC decoders (SW, Hybrid and HW)

[CruNcher]

Originally Posted by v0lt

Arri65_from_raw.mp4 (HEVC 10-bit 6560x3100x23,976 263Mbit)

Geez i had to scratch together that Space somewhere with all the encoding/decoding tests running

Thx this is directly from ARRI RAW->SDR Grade to HEVC ?

And the Scene came from the Alexa 6.5K Film mode, which Lens was used ?

yeah a beast I/O peaks can be also very high

Though lot of noise, hmm encoder wasn't x265 but Reference 10 bit Encoder ?

The Alexa Mounted on the bottom of a Helicopter facing straight downwards ?

But that Horizontal Pan is so perfectly precise really a Helicopter with some Gyro Stabilization ?

[v0lt]

Originally Posted by CruNcher

Thx this is directly from ARRI RAW->SDR Grade to HEVC ?

And the Scene came from the Alexa 6.5K Film mode, which Lens was used ?

I do not know the details. Information from the official page:

ALEXA 65 Footage
All three ALEXA 65 clips were shot with project rate 23.98 in ARRIRAW Open Gate format. One clip is in sync speed the other two in 60 fps high speed. The used lenses are Prime 28, 50 and 80mm.

[NikosD]

Does anyone have the slightest idea why an 8 core RyZen 7 has 4.8 billion transistors which is the same as Broadwell-EX with 15 cores (!) ?

To give you an idea, biggest Broadwell-E processor, the 10 core Core i7-6950K, has "only" 3.2 billion transistors.

What is inside RyZen 7 occupying so many transistors ?

[huhn]

512kb level 2 cache per core intel is using 256 kb per core , maybe more IO most of the south bridge is on the CPU now and http://techreport.com/news/22100/amd...ansistor-count

not that it matter but we will know fast enough

[CruNcher]

Dont forget the Trustzone powered by ARM

http://www.amd.com/en-us/press-relea...2012jun13.aspx

The Hypervisor is a ARM Cortex Ax which will be tightly integrated most probably rather similar to how Intels ME works in combination with their AES 256 Intstructions

http://hackaday.com/2016/11/28/neutr...gement-engine/

In the End this will allow you to play the 4K content Bus Encrypted via DXVAs Protected Media Path and controlled by the Trustzone in Realtime for integrity.



Which gonna mean this time we have fought content piracy successful and brought it finally successful to the PC it was about time as soon as everyone switched to Windows 10 and coming OSes we can pull the switch and transfer all the content finally and MSE is integrated as well in the chain, also we can finally start to introduce all this new exciting monetarization ways via tighter OS Content Control a Dream comes true so many years of work finally become reality

Now that every Hardware is ready it will strengthen our tumbling Economies and loses over the coming years finally reduced and the Web will get cleaner as well for AVG users in the transition Process most Online Pirates and Real Criminals wont even realize what hit them, sometime it needs force to make avg users understand what's good for them

[el Filou]

Originally Posted by CruNcher

Which gonna mean this time we have fought content piracy successful and brought it finally successful to the PC it was about time as soon as everyone switched to Windows 10 and coming OSes we can pull the switch and transfer all the content finally and MSE is integrated as well in the chain, also we can finally start to introduce all this new exciting monetarization ways via tighter OS Content Control a Dream comes true so many years of work finally become reality

Now that every Hardware is ready it will strengthen our tumbling Economies and loses over the coming years finally reduced and the Web will get cleaner as well for AVG users in the transition Process most Online Pirates and Real Criminals wont even realize what hit them, sometime it needs force to make avg users understand what's good for them

It's hard to know when you're serious or not from your posts, but this time I have to hope you're joking.
Like every DRM technology before it, it will eventually be either cracked or bypassed and won't stop piracy, all while bringing additional costs to consumers, producers and every intermediate party.
Every DRM scheme is a disease and this is the HIV of DRM.

[CruNcher]

That's your opinion but those who implement it think different it's bringing order to the chaos in their view of it and it will have a impact on society in several ways good/bad, of course the economical impact in the future will be huge and so will be it's misuse therefore we have to fight which rules are rules Society can agree upon like we have found rules for Law and Order in Real Life but it needs to be incorporated directly in IT Systems and that is happening right now

Some respect the Law and Order system some don't and you get punished and jailed if you disrespect it and the Society it's Values it's Protecting being it Real Property or Intellectual or your Life itself sooner or later IT Systems will completely fall under this Judgement in Realtime you can't avoid it with the rising complexity and issues coming from things like IOT or Augmentation not if your Societies System is broken in itself and you need to drive it on the way it is.

Society needs Protection on this Layer DRM/DRS will ensure it like the real Law and Order System ensures what's your property and what's not and what happens if you go against the Rules you gonna get Punished how that Punishments will look and these Future Courtrooms is still to be decided.

The biggest Problem is that currently the decisions about it are purely driven by Corporate Economical interests and that needs to stop and the overall Process becoming more transparent.

[jkauff]

I would be in favor of an end to piracy if it was accompanied by an end to price-gouging. The popular movies and TV shows most people want to own have recouped their cost many, many times over by the time they're released on disc. Disc manufacturing is a trivial cost to major studios (not to mention the cost of streaming, which is basically nothing). Blu-ray movies should cost half what they do, and the gouging will get even worse with 4K.

[kypec]

Originally Posted by jkauff

I would be in favor of an end to piracy if it was accompanied by an end to price-gouging. ...

My words exactly. Content owners and providers should understand that restricting online access to only certain regions is wrong. I have started to use and pay monthly Netflix account in my country last year when it became available here in Slovakia but it still irritates me that there is almost never a TV show or movie that I want to watch. Most of the content is either available only to U.S. customers or not available at all, not even the movies that were released like 3+ years ago. I mean, what's the point of paying online streaming service when there is no interesting content to watch.

[NikosD]

It's been around 2.5 years since my first post in this thread, evaluating SW and Hybrid HEVC decoders of that time, since no pure fixed-function HW HEVC was available.

The samples that I had chosen back then, they were all low bitrate in order to be feasible for SW and hybrid decoders to decode them in reasonable framerate.

Now that all big three (AMD, Intel, Nvidia) provide pure HW HEVC decoders (8bit & 10bit) it's time to gather a new collection of samples that fit better to HW video decoding and post results for all three.

You can find a link to download each clip right below it or you can find them altogether in this host: (thanks to mariush)

ftp://helpedia.com/pub/multimedia/te...ecoders_clips/ (Max 2 downloads/user, Max 1MB/s)


All tests were done using DXVA Checker v3.x, LAV Video 0.6x in benchmark decode and benchmark playback mode (downscaled to 1280 x720) using DXVA native


8bit HEVC 4K/8K samples

Cards/iGPUs tested:

nVidia Turing (up to 8K)
Nvidia Pascal (up to 8K)
Nvidia GM206 (3rd gen Maxwell) (up to 4K)
Intel Kabylake (up to 8K)
Intel Skylake (up to 8K)
AMD Polaris (up to 4K)


1) LG_4K_View-the-Feeling (2160p@29.970fps ~25Mbps)
http://demo-uhd3d.com/fiche.php?cat=uhd&id=101


Decode:


GTX 1660 - 338/357/370

GTX 1060 - 217/223/227

GTX 1050 Ti 187/193/204

GTX 1050 172/180/194

Kabylake HD 620 156/170/195

Kabylake HD 610 138/165/194

GTX 960 137/145/163

RX 470 97/105/111



Playback:


GTX 1660 - 340/358/368

GTX 1060 - 217/223/227

GTX 1050 Ti 195/199/204

GTX 1050 183/190/201

Skylake HD 530 146/168/202

Kabylake HD 620 146/160/187

GTX 960 141/145/165

Kabylake HD 610 97/123/147

RX 470 101/107/113




2) Samsung 4K@30fps-80Mbps file -0235
http://imaging-resource.com/PRODS/sa...HD_Sample1.MP4



Decode:


GTX 1660 - 289/295/307

Kabylake HD 620 181/188/194

GTX 1060 - 183/185/189

Kabylake HD 610 134/173/183

GTX 1050 Ti 158/162/164

GTX 1050 149/155/162

GTX 960 119/120/122

RX 470 96/99/106


Playback:


GTX 1660 - 293/302/308

Skylake HD 530 174/190/201

GTX 1060 - 185/186/190

Kabylake HD 620 172/177/180

GTX 1050 Ti 156/161/165

GTX 1050 148/154/158

Kabylake HD 610 130/140/151

GTX 960 121/123/126

RX 470 95/98/102


3) Samsung 4K@30fps-80Mbps file -0230
http://www.imaging-resource.com/PROD...ProQuality.MP4


Decode:


GTX 1660 - 248/254/263

GTX 1060 - 175/176/179

Skylake HD 530 160/164/168

Kabylake HD 620 153/156/158

Kabylake HD 610 149/153/158

GTX 1050 Ti 148/153/157

GTX 1050 137/141/145

GTX 960 114/115/117

RX 470 94/96/99


Playback:


GTX 1660 - 257/258/262

GTX 1060 - 175/176/180

Skylake HD 530 147/155/160

GTX 1050 Ti 148/154/158

Kabylake HD 620 148/149/152

GTX 1050 138/143/148

Kabylake HD 610 107/122/128

GTX 960 116/116/121

RX 470 95/98/102


4) Crowd_run 4K@50fps -300Mbps
https://mega.nz/#!RVNy0BhB!DpsJlf5n_...vT0P12VikS3xko


Decode:


GTX 1660 - 176/180/180

GTX 1060 - 142/145/147

GTX 1050 Ti 113/125/128

Skylake HD 530 112/120/123

Kabylake HD 620 101/110/113

Kabylake HD 610 97/109/113

GTX 1050 92/104/109

GTX 960 83/93/97

RX 470 60/70/77


Playback:


GTX 1660 - 176/182/182

GTX 1060 - 139/145/147

GTX 1050 Ti 115/125/128

Skylake HD 530 102/114/119

Kabylake HD 620 96/106/109

GTX 1050 94/106/114

GTX 960 83/94/98

Kabylake HD 610 72/92/98

RX 470 60/72/79


5) 1st Bonus clip:
Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit
https://yadi.sk/i/swMdlJHMuB6yR


Decode:


GTX 1660 - 96/109/127

GTX 1060 - 81/91/101

GTX 1050 Ti 66/80/93

Kabylake HD 610 64/80/96

Kabylake HD 620 62/76/88

GTX 1050 49/64/73

GTX 960 48/59/70

RX 470 33/44/51


Playback:


GTX 1660 - 96/109/127

GTX 1060 - 81/91/101

GTX 1050 Ti 66/80/92

Kabylake HD 620 62/75/85

Kabylake HD 610 57/72/87

GTX 1050 49/64/74

GTX 960 47/59/71

RX 470 34/46/55


6) 2nd Bonus clip (8K):
8K_Ex_sample- 7680x4320@24fps-53Mbps
http://art11111.com/image/video/8K_Ex_sample.mp4


Decode:


GTX 1660 - 95/105/108

GTX 1060 - 55/59/62

GTX 1050 46/52/55

Kabylake HD 620 42/46/48

Kabylake HD 610 41/46/49



Playback:


GTX 1660 - 95/105/108

GTX 1060 - 55/59/62

GTX 1050 45/52/54

Kabylake HD 620 41/44/47

Kabylake HD 610 36/41/43



10bit HEVC 4K/8K samples

Cards/iGPUs tested:

nVidia Turing (up to 8K)
Nvidia Pascal (up to 8K)
Nvidia GM206 (3rd gen Maxwell) (up to 4K)
Intel Kabylake (up to 8K)
AMD Polaris (up to 4K)


1.UHD_PQ_Lovely_Swiss-51Mbps@60fps
http://demo-uhd3d.com/files/uhd4k/De...vely_Swiss.zip

Decode:


GTX 1660 - 310/340/365

GTX 1060 - 215/222/242

GTX 1050 171/179/197

GTX 1050 Ti 159/179/211

GTX 960 112/124/153

Kabylake HD 610 74/115/152

RX 470 73/87/108



Playback:


GTX 1660 - 305/335/360

GTX 1060 - 213/221/237

GTX 1050 185/191/205

GTX 1050 Ti 171/190/217

GTX 960 111/122/155

RX 470 80/94/116

Kabylake HD 610 52/76/87



2.Crowd_run_2160p50_Main10_100Mbps
https://www.sendspace.com/file/giy7br


Decode:


GTX 1660 - 292/298/298

GTX 1060 - 205/208/209

GTX 1050 Ti 175/179/181

GTX 1050 161/164/165

Kabylake HD 620 155/157/158

Kabylake HD 610 112/118/121

GTX 960 108/109/113

RX 470 86/89/98 (CPU usage ~20% on quad core)


Playback:


GTX 1660 - 290/294/294

GTX 1060 - 204/207/208

GTX 1050 Ti 173/176/177

GTX 1050 162/166/169

Kabylake HD 620 137/139/141

GTX 960 108/112/115

RX 470 94/96/98

Kabylake HD 610 68/75/79



3.Jellyfish-140-mbps-4k-uhd-hevc-10bit
http://jell.yfish.us/media/jellyfish...hevc-10bit.mkv


Decode:


GTX 1660 - 177/208/256

GTX 1060 - 151/169/191

GTX 1050 Ti 130/146/165

Kabylake HD 620 117/133/151

GTX 1050 112/128/147

Kabylake HD 610 94/110/124

GTX 960 94/103/114

RX 470 67/80/111


Playback:


GTX 1660 - 178/210/255

GTX 1060 - 151/169/191

GTX 1050 Ti 131/147/165

GTX 1050 114/131/150

Kabylake HD 620 111/123/139

GTX 960 92/102/113

RX 470 74/85/96

Kabylake HD 610 65/77/83



4.Jellyfish-400-mbps-4k-uhd-hevc-10bit
http://jell.yfish.us/media/jellyfish...hevc-10bit.mkv


Decode:


GTX 1660 - 86/105/130

GTX 1060 - 67/84/109

GTX 1050 Ti 60/74/96

GTX 1050 53//83

Kabylake HD 620 54/67/89

Kabylake HD 610 53/65/83

GTX 960 38/55/71

RX 470 32/40/52


Playback:


GTX 1660 - 85/104/129

GTX 1060 - 67/84/109

GTX 1050 Ti 60/75/96

Kabylake HD 620 53/66/85

GTX 1050 42/55/74

GTX 960 44/55/72

Kabylake HD 610 44/55/68

RX 470 32/41/53



5) 1st Bonus clip (7K):
5.Arri65_from_raw-6560x3100-263Mbps@24fps
http://www.dmitry.com/tests/ARRI/Arri65_from_raw.mp4


Decode:


GTX 1660 - 103/103/106

GTX 1060 - 74/76/77

GTX 1050 Ti 63/68/72

Kabylake HD 620 54/56/58

GTX 1050 52/55/59

Kabylake HD 610 40/47/50



Playback:


GTX 1660 - 100/103/107

GTX 1060 - 74/76/77

GTX 1050 Ti 64/68/72

GTX 1050 52/55/59

Kabylake HD 620 48/51/53

Kabylake HD 610 30/35/37



6) 2nd Bonus clip (8K):
MBC UHD Music Core GFriend - Rough (4in1) 8k@60fps-80Mbps
https://mega.nz/#!8hcmwDoK!UwfMhPQvt...iqrNTwMVHjUZ6Y


Decode:


GTX 1660 - 98/103/107

GTX 1060 - 59/60/63

GTX 1050 Ti 53/56/58

GTX 1050 52/54/56

Kabylake HD 620 43/48/56

Kabylake HD 610 31/35/41


Playback:


GTX 1660 - 96/101/104

GTX 1060 - 59/60/63

GTX 1050 Ti 54/55/56

GTX 1050 52/54/55

Kabylake HD 620 38/41/47

Kabylake HD 610 24/25/29


Comments:

1) It is now clear from the benchmark results, that Intel has changed its strategy to bring the fastest HW video decoder and targets the most efficient.

When Sandybridge back on 2011 introduced for the first time QuickSync fixed-function decoder/encoder, it was the fastest ASIC of that time with a clock around 1300 MHz, when AMD and Nvidia had clocks around ~500 MHz.

But Intel in every generation manages to increase performance and at the same time lowers the clock, so Kabylake's iGPU has exactly the same clock as Skylake (1150 MHz) which is lower than Sandy, 6 years ago.

On the other hand, Nvidia with Pascal cards managed to reach ~1700 MHz which is almost 50% faster clock than Kabylake and reached in raw terms the decoding speed of Intel, but not so efficiently in order to achieve that clock.

Turing cards go even faster at ~1800 MHz for Video Clock and for HEVC seem to be in a different category especially for 8K clips which seem to be a lot more optimized than Pascal.

But 8K VP9 is almost the same for Turing and Pascal, like 4K H.264 too.


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.


3) Turing GTX 1660 has a video clock of 1800 - 1815 MHz, while GTX 1050 Ti runs at 1645 MHz "GPU Video Clock" and GTX 1060 at 1708.5 MHz.
GTX 1050 runs at 1582 - 1607 MHz
The GPU Video Clock difference is 4% for Pascal cards but the real-world video decoding performance is ~15% faster for 1060 in most of the clips (besides the huge bitrate clips in which they are a little closer)


4) Nvidia GTX 960 has a "GPU Video Clock" of 1202.3 MHz, so a 1050 Ti is ~37% faster clock for clock.

It seems that the HEVC 8bit decoder follows exactly the clock difference, so the decoder should be actually the same.

But regarding HEVC 10bit decoder, the first clips show an almost double difference than 8bit decoder favoring, of course, the Pascal chip.

The huge bitrate 10bit clip (400Mbps) has exactly 37% difference though, like the clock difference.


5) Skylake HD 530 and Kabylake HD 630, both run at the same speed of 1150 MHz and Kabylake HD 620/610 run at 1050 MHz.
The results of pure decoding speed of Kabylake and Skylake look similar on 8bit clips, they could have the exact same HEVC 8bit decoder (only difference in MHz)
The same goes to 610/620 which could/ should have the same speed on both 8bit & 10bit HEVC clips and the only difference should be the playback scaled down to 720p speed due to a lot more units of 620.
Different drivers and different memory configuration (single channel/ dual channel) could affect performance on iGPUs, too.

[loekf]

Time permitting, will try to put my newly installed GTX 1050TI through the test.

I was looking for these numbers. What happened to AMD's Polaris UVD ? Looks like it has much lower processing power and has a "all-is-ok-if-I-output-60fps" philosophy.

[NikosD]

I will add tomorrow 10bit HEVC clips too, so just wait for another day

AMD was always conservative to media performance.

They don't care for Media Servers and multiple stream decoding, so they are just to the point for one stream performance.

[loekf]

Originally Posted by NikosD

I will add tomorrow 10bit HEVC clips too, so just wait for another day

AMD was always conservative to media performance.

They don't care for Media Servers and multiple stream decoding, so they are just to the point for one stream performance.

BTW, this link is not correct:

http://imaging-resource.com/PRODS/sa...HD_Sample1.MP4

[NikosD]

Originally Posted by loekf

BTW, this link is not correct:

http://imaging-resource.com/PRODS/sa...HD_Sample1.MP4

When I click on this link, it immediately starts downloading the file.

[loekf]

Originally Posted by NikosD

When I click on this link, it immediately starts downloading the file.

Weird, here at work it works. At home I got a 404 error at that Samsung site (or missing page).

Anyways, lemme download LAV tonight and do some testing.

[loekf]

Okay here we go:

Tools: DXVA checker rev. 3.15, LAV Filters 0.69.0

System: Intel Core i5 7400 (Kabylake), Nvidia GTX1050 TI
Decode benchmarks settings: 64-bit
Playback benchmarks settings: 64-bit, downscaled to 1280 x 720

For all results: 5 runs, usually first run is worst. Took the last run.

1. LG_4K_View-the-Feeling.mp4

Decode: 187 - 193 - 204
Playback: 195 - 199 - 204

2. yvid_SAM_0235-UHD_Sample1.mp4

Decode: 158 - 162 - 164
Playback: 156 - 161 - 165

3. yvid_SAM_0230-ProQuality.mp4

Decode: 148 - 153 - 157
Playback: 148 - 154 - 158

4. crowd_run_2160p50_300Mbps.mkv

Decode: 113 - 125 - 128
Playback: 115 - 125 - 128

5. Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit.mkv

Decode: 66 - 80 - 93
Playback: 68 - 80 - 92

BTW, your link to test case 3 is really wrong. Issue is the wrong file name. Should be:

http://www.imaging-resource.com/PROD...ProQuality.MP4

[NikosD]

Originally Posted by loekf

Okay here we go:

Tools: DXVA checker rev. 3.15, LAV Filters 0.69.0

System: Intel Core i5 7400 (Kabylake), Nvidia GTX1050 TI
Decode benchmarks settings: 64-bit
Playback benchmarks settings: 64-bit, downscaled to 1280 x 720

For all results: 5 runs, usually first run is worst. Took the last run.

Your results are perfect, containing all min/avg/max values.

We usually abandon the first run and the best run (if it's too high) and take the average of the rest.

Because maybe your last run has too high or different than average results.

Also, what GPU did you test ?

Kaby or 1050 Ti ? I think 1050 Ti, but at what clocks ? (core/ memory)

Can you run the exact same tests on both GPUs ? (Kaby and 1050 Ti)

I will upload in a moment the 10bit samples, so you can run the test on them too.

Thanks!

Update:
Your results do not much with GTX 1060.
It has a good difference.
Clocks maybe ?

Originally Posted by loekf

BTW, your link to test case 3 is really wrong. Issue is the wrong file name. Should be:

http://www.imaging-resource.com/PROD...ProQuality.MP4

Right.
I updated the link.
Thanks!

[loekf]

Originally Posted by NikosD

Also, what GPU did you test ?

Only the GTX 1050 TI inside my HTPC. The iGPU of the Kabylake CPU is disabled, guess automatically via the BIOS. To test the Kabylake iGPU guess I would have to take the GPU out and install proper Intel drivers.

The 1050 TI is a MSI with 1455 MHz clock (7GHz memory). So compared to a 1060 that's roughly 15% lower. The 1060 usually also has a 8 GHz vs 7 GHz clock. Looks like the difference between the two is smaller than you would expect based on clock speeds.

I will try to make time for another round of testing.

[NikosD]

Originally Posted by loekf

The 1050 TI is a MSI with 1455 MHz clock (7GHz memory). So compared to a 1060 that's roughly 15% lower. The 1060 usually also has a 8 GHz vs 7 GHz clock. Looks like the difference between the two is smaller than you would expect based on clock speeds.

I will try to make time for another round of testing.

Use a monitoring tool like GPU-Z to see the GPU clocks when you run the tests but write down the results without GPU-Z running.

I posted the new five HEVC 10bit clips.

[NikosD]

Originally Posted by loekf

I will try to make time for another round of testing.

I have updated the previous post with your 8 bit results.

Waiting for new feedback regarding the 10 bit clips.