Evaluation of HEVC decoders (SW, Hybrid and HW)

[clsid]

Why is the GPU usage so high? Did they perhaps use PS2.0 bicubic resize? DXVA scaling would be interesting is MPC-HC ever gets it.

[pirlouy]

No doubt about it, vanilla EVR is always the best&only choice for weak platforms.
Maybe this has already been asked but why madVR which seems to be well coded can compete with EVR if you you use basic filters (bilinear) and you enable all "trade quality for performance" settings ?

[huhn]

madVR is nothing a modern iGPU can't handle.

adding DXVA scaling to EVR sounds pretty pointless to me. DXVA scaling is too limited it only works with a low number of pixel formats.

[NikosD]

Why is the GPU usage so high? Did they perhaps use PS2.0 bicubic resize? DXVA scaling would be interesting is MPC-HC ever gets it.
Kadano, a user with iGPU HD 610 just informed me that if you don't disable the default "Noise reduction" from the control panel you get 70fps.

If you disable it, you get 122 fps.

Anandtech could have easily forgotten the default settings on, something that I always suggest to disable.

[pirlouy]

madVR is nothing a modern iGPU can't handle.But why ? If you use dxva scaling and all "trade quality" settings, madVR should be at the same level, no ?

[CruNcher]

Kadano, a user with iGPU HD 610 just informed me that if you don't disable the default "Noise reduction" from the control panel you get 70fps.

If you disable it, you get 122 fps.

Anandtech could have easily forgotten the default settings on, something that I always suggest to disable.

Nope if you read also previous htpc targeted platform tests you will see they are aware of the Driver Settings Influence on the Shader Overhead and recommend to disable any post processing on the Driver side especially with MadVR.

But why ? If you use dxva scaling and all "trade quality" settings, madVR should be at the same level, no ?

You should think so but MadVR has also has by default no matter which settings you chose a higher CPU overhead and consumes more CPU cycles which can get critical as well @ 6W-12W overall CPU/GPU target ;)

[nevcairiel]

But why ? If you use dxva scaling and all "trade quality" settings, madVR should be at the same level, no ?

huhn used double negative, meaning that even iGPUs can handle madVR if you stick to modest settings.

[vlado08]

Intel NUC7i3BNH
CPU: i3-7100U 2.4Ghz
GPU: Intel HD 620
RAM: 16GB DDR4-2132MHz - dual channel
Windows 10 x64 1607 os build 14393.969
Graphics driver 21.20.16.4590
LAV Video x64 0.69.0
DXVA Checker 3.15.0


Power Options - High Performance
(GPU @ 1000Mhz during benchmark)
video enhancement - disable; monitor@1080/59p

HEVC 8bit:

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

Decode: 156/170/195 CPU 3[0-6]
Playback: 146/160/187 CPU 5[3-10]

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

Decode: 181/188/194 CPU 3[1-5]
Playback: 172/177/180 CPU 4[2-7]

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

Decode: 153/156/158 CPU 3[1-4]
Playback: 148/149/152 CPU 5[3-7]

4) Crowd_run 4K@50fps -300Mbps

Decode: 101/110/113 CPU 3[1-6]
Playback: 96/106/109 CPU 4[1-7]

5) 1st Bonus clip:
Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit

Decode: 62/76/88 CPU 2[1-7]
Playback: 62/75/85 CPU 4[2-8]

6) 2nd Bonus clip:
8k_ex_sample 8k@24fps - 53Mbps

Decode: 42/46/48 CPU 2[0-5]
Playback: 41/44/47 CPU 3[2-5]

HEVC 10bit:

2.Crowd_run_2160p50_Main10_100Mbps

Decode: 155/157/158 CPU 4[3-6]
Playback: 137/139/141 CPU 6[5-6]

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

Decode: 117/133/151 CPU 4[2-11]
Playback: 111/123/139 CPU 6[4-11]

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

Decode: 54/67/89 CPU 4[1-9]
Playback: 53/66/85 CPU 5[2-13]

5.1st Bonus clip:
Arri65_from_raw 6560x3100@23.976fps 263Mbps

Decode: 54/56/58 CPU 3[1-7]
Playback: 48/51/53 CPU 6[4-10]

6.2nd Bonus clip:
MBC UHD Music Core GFriend - Rough (4in1) 8k
7680x4320@59.94fps 82.4Mbps

Decode: 43/48/56 CPU 3[0-8]
Playback: 38/41/47 CPU 4[0-10]

[vlado08]

Intel NUC7i3BNH
CPU: i3-7100U 2.4Ghz
GPU: Intel HD 620
RAM: 16GB DDR4-2132MHz - dual channel
Windows 10 x64 1607 os build 14393.969
Graphics driver 21.20.16.4590
LAV Video x64 0.69.0
DXVA Checker 3.15.0


Power Options - High Performance
(GPU @ 1000Mhz during benchmark)
video enhancement - disable; monitor@1080/59p

HEVC 8bit:

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

Decode: 156/170/195 CPU 3[0-6]
Playback: 146/160/187 CPU 5[3-10]

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

Decode: 181/188/194 CPU 3[1-5]
Playback: 172/177/180 CPU 4[2-7]

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

Decode: 153/156/158 CPU 3[1-4]
Playback: 148/149/152 CPU 5[3-7]

4) Crowd_run 4K@50fps -300Mbps

Decode: 101/110/113 CPU 3[1-6]
Playback: 96/106/109 CPU 4[1-7]

5) 1st Bonus clip:
Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit

Decode: 62/76/88 CPU 2[1-7]
Playback: 62/75/85 CPU 4[2-8]

6) 2nd Bonus clip:
8k_ex_sample 8k@24fps - 53Mbps

Decode: 42/46/48 CPU 2[0-5]
Playback: 41/44/47 CPU 3[2-5]

HEVC 10bit:

2.Crowd_run_2160p50_Main10_100Mbps

Decode: 155/157/158 CPU 4[3-6]
Playback: 137/139/141 CPU 6[5-6]

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

Decode: 117/133/151 CPU 4[2-11]
Playback: 111/123/139 CPU 6[4-11]

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

Decode: 54/67/89 CPU 4[1-9]
Playback: 53/66/85 CPU 5[2-13]

5.1st Bonus clip:
Arri65_from_raw 6560x3100@23.976fps 263Mbps

Decode: 54/56/58 CPU 3[1-7]
Playback: 48/51/53 CPU 6[4-10]

6.2nd Bonus clip:
MBC UHD Music Core GFriend - Rough (4in1) 8k
7680x4320@59.94fps 82.4Mbps

Decode: 43/48/56 CPU 3[0-8]
Playback: 38/41/47 CPU 4[0-10]

[Kadano]

I tested Intel Pentium G4560 / HD 610. No dedicated GPU attached. 1x 8GB DDR3 RAM (1600MHz) used.

8bit HEVC

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

Decode: 165 fps [138-194]
Playback [1280x720]: 123 fps [97-147]

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

Decode: 173 fps [134-183]
Playback [1280x720]: 140 fps [130-151]

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

Decode: 153 fps [149-158]
Playback [1280x720]: 122 fps [107-128]

4) Crowd_run 4K@50fps -300Mbps

Decode: 109 fps [97-113]
Playback [1280x720]: 92 fps [72-98]

5) 1st Bonus clip:
Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit

Decode: 80 fps [64-96]
Playback [1280x720]: 72 fps [57-87]

6) 2nd Bonus clip (8K):
8K_Ex_sample- 7680x4320@24fps-53Mbps

Decode: 46 fps [41-49]
Playback [1280x720]: 41 fps [36-43]


10bit HEVC 4K/8K samples

1) 1.UHD_PQ_Lovely_Swiss-51Mbps@60fps

Decode: 115 fps [74-152]
Playback [1280x720]: 76 fps [52-87]

2.Crowd_run_2160p50_Main10_100Mbps

Decode: 118 fps [112-121]
Playback [1280x720]: 75 fps [68-79]

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

Decode: 110 fps [94-124]
Playback [1280x720]: 77 fps [65-83]

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

Decode: 65 fps [53-83]
Playback [1280x720]: 55 fps [44-68]

5) 1st Bonus clip (7K):
5.Arri65_from_raw-6560x3100-263Mbps@24fps

Decode: 47 fps [40-50]
Playback [1280x720]: 35 fps [30-37]

6) 2nd Bonus clip (8K):
MBC UHD Music Core GFriend - Rough (4in1) 8k@60fps-80Mbps

Decode: 35 fps [31-41]
Playback [1280x720]: 25 fps [24-29]



Detailed results: https://docs.google.com/spreadsheets/d/1p_z3pLO_zFvcB0A57WXDILUwOjRFWbcETYG8SOIURxI/

[NikosD]

I tested Intel Pentium G4560 / HD 610. No dedicated GPU attached. 1x 8GB DDR3 RAM (1600MHz) used.


That's perfect thank you.

But where are the above 4K results ?

I think that HD 610 supports up to 8K acceleration.

Also, can you confirm using GPU-Z or HWINFO that GPU clock is 1050MHz (or something else ?)

[NikosD]

I tested Intel Pentium G4560 / HD 610. No dedicated GPU attached. 1x 8GB DDR3 RAM (1600MHz) used.


Now that I put your data on the table with the rest of the HW video decoders, they look more interesting and helpful than what I was expecting to be.

Yes, I was expecting a drop in playback performance due to only 12 EUs since HD 610 is just a GT1 version, but even pure decoding performance is lower than I was personally expecting.

Its decoding performance is better for 8bit clips than 10bit, but even in the worst case scenario of 10bit playback performance, it's still faster than Polaris UVD, but most of the times it's slower than everything else (in worst case scenario)

I have deleted the HD 630 results because they were incomplete and I'm waiting for your results on the rest 4 clips.

[Lollerke]

Now that I put your data on the table with the rest of the HW video decoders, they look more interesting and helpful than what I was expecting to be.

Yes, I was expecting a drop in playback performance due to only 12 EUs since HD 610 is just a GT1 version, but even pure decoding performance is lower than I was personally expecting.

Its decoding performance is better for 8bit clips than 10bit, but even in the worst case scenario of 10bit playback performance, it's still faster than Polaris UVD, but most of the times it's slower than everything else (in worst case scenario)

I have deleted the HD 630 results because they were incomplete and I'm waiting for your results on the rest 4 clips.

Maybe because the RAM is single channel and runs only at 1600MHz? With integrated GPU's fast dual channel RAMs are essential for gaming performance. In this case two 2400MHz DDR4 sticks in dual channel would more than double the memory bandwidth.

[NikosD]

RAM and general CPU performance has nothing to do (or shouldn't have) with HW decoding, especially in DXVA native mode.

It must be a weaker ASIC (less MFX units)

[Kadano]

That's perfect thank you.

But where are the above 4K results ?

I think that HD 610 supports up to 8K acceleration.

8bit HEVC

5) 1st Bonus clip:
Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit

Decode: 79 fps [64-96]
Playback [1280x720]: 72 fps [57-87]

6) 2nd Bonus clip (8K):
8K_Ex_sample- 7680x4320@24fps-53Mbps
[DS64] LAV Video Decoder 0.69.0.18-git Unsupported

10bit HEVC 4K/8K samples

5) 1st Bonus clip (7K):
5.Arri65_from_raw-6560x3100-263Mbps@24fps
[DS64] LAV Video Decoder 0.69.0.18-git Unsupported

6) 2nd Bonus clip (8K):
MBC UHD Music Core GFriend - Rough (4in1) 8k@60fps-80Mbps

Decode: 35 fps [31-41]
Playback [1280x720]: 25 fps [24-29]


Also, can you confirm using GPU-Z or HWINFO that GPU clock is 1050MHz (or something else ?)
GPU clock is 1050 MHz during DXVA usage, according to GPU-Z. Memory clock 800 MHz.

A different question, some anime encoders started using HEVC 12-bit. On G4560, software HEVC 12-bit 720p decoding is a bit stressing, with SVP also running. GTX 1000 series and Kaby Lake should be able to decode 12-bit HEVC, but it seems LAV Filters don't support them. Is it possible to set up 12-bit encodes to bypass LAV Filters and be decoded by Microsoft decoders (which do support 12-bit, according to DXVA Checker)?

[nevcairiel]

For iGPUs RAM still has some impact, as it needs to read and write the video frames in there - the system ram also doubles as video ram, so even if its not copied around (ie. by DXVA Copy-Back), it can matter. If it really does - well, one could benchmark some iGPU with different RAM configurations, i guess.

[NikosD]

Interesting post.

What do you mean "the system ram also doubles as video ram" ?

What is the difference between iGPU and dGPU for HW decoding using DXVA native ?

[Kadano]

Nvidia released 381.65 WHQL drivers today, new driver branch that supports Win10 Creators Update & WDDM 2.2, can Pascal users update to the latest driver and run DXVA Checker and post screenshots here?

http://www.geforce.com/whats-new/articles/quake-champions-closed-beta-game-ready-driver

Old: http://i.imgur.com/Q8zRdEP.jpg

New: http://i.imgur.com/YgZV3vJ.jpg

Seems nothing has improved.

[pacuro]

There were nothing to improve with GP107 as it has VP9.2 hardware decoder since the begining.
Unfortunatelly nothing improved with GP106, which is most expected I guess.

http://static.frazpc.pl/board/2017/04/a9a2f3e5be60fab9.jpg

[vlado08]

Day or two ago made 1 post but it didn't appear?
Repeated it again and it didn't appear?
I'm trying now again.

[vlado08]

Intel NUC7i3BNH
CPU: i3-7100U 2.4Ghz
GPU: Intel HD 620
RAM: 16GB DDR4-2132MHz - dual channel
Windows 10 x64 1607 os build 14393.969
Graphics driver 21.20.16.4590
LAV Video x64 0.69.0
DXVA Checker 3.15.0


Power Options - High Performance
(GPU @ 1000Mhz during benchmark)
video enhancement - disable; monitor@1080/59p

HEVC 8bit:

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

Decode: 156/170/195 CPU 3[0-6]
Playback: 146/160/187 CPU 5[3-10]

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

Decode: 181/188/194 CPU 3[1-5]
Playback: 172/177/180 CPU 4[2-7]

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

Decode: 153/156/158 CPU 3[1-4]
Playback: 148/149/152 CPU 5[3-7]

4) Crowd_run 4K@50fps -300Mbps

Decode: 101/110/113 CPU 3[1-6]
Playback: 96/106/109 CPU 4[1-7]

5) 1st Bonus clip:
Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit

Decode: 62/76/88 CPU 2[1-7]
Playback: 62/75/85 CPU 4[2-8]

6) 2nd Bonus clip:
8k_ex_sample 8k@24fps - 53Mbps

Decode: 42/46/48 CPU 2[0-5]
Playback: 41/44/47 CPU 3[2-5]

HEVC 10bit:

2.Crowd_run_2160p50_Main10_100Mbps

Decode: 155/157/158 CPU 4[3-6]
Playback: 137/139/141 CPU 6[5-6]

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

Decode: 117/133/151 CPU 4[2-11]
Playback: 111/123/139 CPU 6[4-11]

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

Decode: 54/67/89 CPU 4[1-9]
Playback: 53/66/85 CPU 5[2-13]

5.1st Bonus clip:
Arri65_from_raw 6560x3100@23.976fps 263Mbps

Decode: 54/56/58 CPU 3[1-7]
Playback: 48/51/53 CPU 6[4-10]

6.2nd Bonus clip:
MBC UHD Music Core GFriend - Rough (4in1) 8k
7680x4320@59.94fps 82.4Mbps

Decode: 43/48/56 CPU 3[0-8]
Playback: 38/41/47 CPU 4[0-10]

[vlado08]

when I post my testing they didn't appear. There was short message when I made the post which I couldn't read.

[NikosD]

Day or two ago made 1 post but it didn't appear?
Repeated it again and it didn't appear?
I'm trying now again.

when I post my testing they didn't appear. There was short message when I made the post which I couldn't read.

@mods
I received the same complain from another user that he was trying to post a simple text without images or attachments and he got a message that he must wait for approval.

Why?

The approval didn't come and he didn't show his message posted.

Is there some kind of restriction for new users in the forum with a small number of posts regarding their ability to post new messages ?

[NikosD]

For iGPUs RAM still has some impact, as it needs to read and write the video frames in there - the system ram also doubles as video ram, so even if its not copied around (ie. by DXVA Copy-Back), it can matter. If it really does - well, one could benchmark some iGPU with different RAM configurations, i guess.

Interesting post.

What do you mean "the system ram also doubles as video ram" ?

What is the difference between iGPU and dGPU for HW decoding using DXVA native ?

I replied you with the above post but you didn't reply.

What is the difference regarding HW decoding using DXVA native between iGPU and dGPU ?

In the meantime, Intel wants to reply to me (and you) with that new KabyLake-G:
http://wccftech.com/intel-kaby-lake-g-hbm2-gpu-multi-die/

A new Kabylake, if the rumor is true, with multi-die design.
One die a 4 core CPU with iGPU GT2 and the other die a dGPU with HBM2 memory as VRAM connected with a PCIe 3.0 x8.

A mystery around dGPU.
Is it Intel or AMD ?

[nevcairiel]

Excuse me for being out yesterday evening. :p

The difference is that a dGPU has dedicated fast video memory (ie. GDDR5 or what ever yours has), which is much faster then the normal system RAM.
An iGPU does not have this, it uses your normal system RAM as video memory. Even pure DXVA Native decoding needs to read and write video frames (ie. read reference frames, write decoded frames), so it is very well possible that the speed of the RAM can have an impact on decoding performance. How much (or if at all) would only be possible to actually be tested with different RAM configurations.

Regarding KabyLake-G, what I wonder is whos media features get used. I suppose the media features of the Intel GPU are still available, much like on an Optimus laptop? Who knows.

[NikosD]

Maybe because the RAM is single channel and runs only at 1600MHz? With integrated GPU's fast dual channel RAMs are essential for gaming performance. In this case two 2400MHz DDR4 sticks in dual channel would more than double the memory bandwidth.
Based on the above reply of nevcairiel, I realized that I had forgotten that system RAM is used as both system RAM and video RAM for iGPU.

Yes, a single channel DDR3 1600 could be a small (?) bottleneck even for pure HW DXVA native acceleration.

Only a benchmark with both configurations (single channel/dual channel) could clarify this.

[Kadano]

I tested Nvidia GTX 1050. CPU: Intel G4560. 1x 8GB DDR3 RAM (1600MHz) used. 3 samples each time, first sample dropped.

8bit HEVC

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

Decode: 180 fps [172-194]
Playback [1280x720]: 190 fps [183-201]
GPU Video clock during decoding/playback: 1607 MHz

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

Decode: 155 fps [149-162]
Playback [1280x720]: 154 fps [148-158]
GPU Video clock during decoding/playback: 1595 MHz

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

Decode: 141 fps [137-145]
Playback [1280x720]: 143 fps [138-148]
GPU Video clock during decoding/playback: 1595 MHz

4) Crowd_run 4K@50fps -300Mbps

Decode: 104 fps [92-109]
Playback [1280x720]: 106 fps [94-114]
GPU Video clock during decoding/playback: 1595 MHz

5) 1st Bonus clip:
Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit

Decode: 64 fps [49-73]
Playback [1280x720]: 64 fps [49-74]
GPU Video clock during decoding/playback: 1595 MHz

6) 2nd Bonus clip (8K):
8K_Ex_sample- 7680x4320@24fps-53Mbps

Decode: 52 fps [46-55]
Playback [1280x720]: 52 fps [45-54]
GPU Video clock during decoding/playback: 1595 MHz


10bit HEVC 4K/8K samples

1) 1.UHD_PQ_Lovely_Swiss-51Mbps@60fps

Decode: 179 fps [171-197]
Playback [1280x720]: 191 fps [185-205]
GPU Video clock during decoding/playback: 1595 MHz

2.Crowd_run_2160p50_Main10_100Mbps

Decode: 164 fps [161-165]
Playback [1280x720]: 166 fps [162-169]
GPU Video clock during decoding/playback: 1595 MHz

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

Decode: 128 fps [112-147]
Playback [1280x720]: 131 fps [114-150]
GPU Video clock during decoding/playback: 1595 MHz

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

Decode: fps [53-83]
Playback [1280x720]: 55 fps [42-74]
GPU Video clock during decoding/playback: 1582 MHz

5) 1st Bonus clip (7K):
5.Arri65_from_raw-6560x3100-263Mbps@24fps

Decode: 55 fps [52-59]
Playback [1280x720]: 55 fps [52-59]
GPU Video clock during decoding/playback: 1582 MHz

6) 2nd Bonus clip (8K):
MBC UHD Music Core GFriend - Rough (4in1) 8k@60fps-80Mbps

Decode: 54 fps [52-56]
Playback [1280x720]: 54 fps [52-55]
GPU Video clock during decoding/playback: 1607 MHz

[NikosD]

Perfect, thanks!

Most of the times your GPU Video clock is the same 1595 MHZ but a few times dropped to 1582 MHZ and for the last clip raided to 1607 MHZ

Weird.

Could it be inside the margin of error of the monitoring tool ?

Is it set to monitor every 1sec or more ?

Was the GPU Video clock steady each time for each clip during the whole process ?

[MrScientist]

I tested my four HEVC hardware encoders. I know you already have many of the results, but this was mostly for my information. For the 1050, I ran the tests three times. For the rest, I ran the tests four times and I threw out the first result. The first group of results is the decode benchmark, the second group is the playback benchmark. The first result is Nvidia 1050, second is AMD 460, third is Intel 610 (in G3930), fourth is Intel "HD Graphics" (5xx? in x5-Z8300).

1.LG_4K_View-the-Feeling.mp4

193-194-194
95-94-95
181-181-181
67-68-68

195-197-197
108-108-108
166-169-170
53-52-52

2.yvid_SAM_0235-UHD_Sample1.MP4

165-165-165
99-98-99
199-200-200
79-78-79

159-165-164
101-102-101
173-183-182
60-60-60

3.yvid_SAM_0230-ProQuality.MP4

155-155-155
96-95-95
165-164-166
62-63-63

148-156-156
96-96-97
157-157-157
53-52-853

4.crowd_run_2160p50_300Mbps.mkv

127-127-127
67-67-67
117-118-117
45-45-45

124-127-127
71-70-71
114-115-115
40-40-41

5.Netflix_TunnelFlag_4096x2160_60fps_x265_8bit_700Mbit.mkv

81-81-81
41-41-41
82-82-82
40-40-41

78-80-80
39-39-39
81-81-81
33-33-33

6.8K_Ex_sample.mp4

51-54-53
41-41-41
49-49-49
20-20-20

51-53-53
39-39-39
47-47-47
17-17-17

1.UHD_PQ_Lovely_Swiss.ts

195-197-191
73-74-73
177-178-178
ND

192-192-192
91-91-90
155-155-155
ND

2.Crowd_run_2160p50_Main10_100Mbps

163-177-176
78-76-76
167-166-166
ND

174-181-181
93-93-93
144-145-144
ND

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

141-147-148
76-76-76
142-142-143
ND

143-148-147
84-84-84
129-130-130
ND

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

74-74-74
41-41-40
71-70-71
ND

73-74-74
40-40-41
70-69-69
ND

5.Arri65_from_raw.mp4

68-68-68
ND
28-28-28
ND

60-68-68
ND
28-28-28
ND

6.MBC UHD Music Core GFriend - Rough (4in1) 8k.mkv

54-55-55
ND
51-51-51
ND

54-54-54
ND
36-36-36
ND

[NikosD]

Awesome!

Thank you for your time and effort.

[NikosD]

The first result is Nvidia 1050, second is AMD 460, third is Intel 610 (in G3930), fourth is Intel "HD Graphics" (5xx? in x5-Z8300).


Can you tell me a little more of your systems configuration ?

OS, drivers and especially for G3930 and Atom x5-Z8300 your ram configuration.
(Memory speed and single channel/dual channel)

[MrScientist]

Can you tell me a little more of your systems configuration ?

OS, drivers and especially for G3930 and Atom x5-Z8300 your ram configuration.
(Memory speed and single channel/dual channel)
Sure, sorry I didn't do that earlier.
1. Nvidia 1050, 378.92 drivers, AMD 9590, 32GB dual channel 1333MHz DDR3, Win 10 Pro
2. AMD 460, 17.4.1 drivers, AMD E1-2100, 8GB single channel 1600MHz DDR3, Win 10 Pro
3. Intel 610, 21.20.16.4534 drivers, Intel G3930, 8GB dual channel 2133MHz DDR4, Win 10 Pro
4. Intel "HD Graphics", Intel x5-Z8300, 2GB single channel DDR3, Win 10 Home (This is a tablet. I'll get the drivers and RAM speed later, it's currently not responding.)

[NikosD]

Sure, sorry I didn't do that earlier.
1. Nvidia 1050, 378.92 drivers, AMD 9590, 32GB dual channel 1333MHz DDR3, Win 10 Pro
2. AMD 460, 17.4.1 drivers, AMD E1-2100, 8GB single channel 1600MHz DDR3, Win 10 Pro
3. Intel 610, 21.20.16.4534 drivers, Intel G3930, 8GB dual channel 2133MHz DDR4, Win 10 Pro
4. Intel "HD Graphics", Intel x5-Z8300, 2GB single channel DDR3, Win 10 Home (This is a tablet. I'll get the drivers and RAM speed later, it's currently not responding.)

Your results are freaking fast.

I compare your RX 460 with my RX 470 and 460 is way faster than mine.

Your HD 610 and GTX 1050 are way faster than HD 610 and GTX 1050 of Kadano.

So, what's your secret ?

Do you cherry pick the results based on the maximum figures ?

Do you have a completely clean installation with no apps running in the background ?

One last request:
Can you post also your GPU Video Clock for HD 610, GTX 1050 and RX 460 ?

Thanks!

[MrScientist]

Your results are freaking fast.

I compare your RX 460 with my RX 470 and 460 is way faster than mine.

Your HD 610 and GTX 1050 are way faster than HD 610 and GTX 1050 of Kadano.

So, what's your secret ?

Do you cherry pick the results based on the maximum figures ?

Do you have a completely clean installation with no apps running in the background ?

One last request:
Can you post also your GPU Video Clock for HD 610, GTX 1050 and RX 460 ?

Thanks!

Hi,
I didn't cherry pick, I did the tests with either three or four passes as I described. I followed the test procedure I found here, https://www.youtube.com/watch?v=RhZCtE1P40w. Did I follow the correct protocol? I used 64bit for both benchmarks with "well-used" operating systems.

How do I report GPU video clock? On the benchmark results, it shows "Video Engine Usage." Should I do the tests again with GPU-Z running?

[NikosD]

Hi,
I didn't cherry pick, I did the tests with either three or four passes as I described. I followed the test procedure I found here, https://www.youtube.com/watch?v=RhZCtE1P40w. Did I follow the correct protocol? I used 64bit for both benchmarks with "well-used" operating systems.

How do I report GPU video clock? On the benchmark results, it shows "Video Engine Usage." Should I do the tests again with GPU-Z running?

Yes, I read that you run the tests 3 or 4 times but maybe you didn't get the average value but the best values of those runs.

The procedure in video is fine, as long as you selected DXVA native using LAV Video x64.

What version of LAV did you test ?

The "GPU Video Clock" is the "GPU clock" for Intel reported by GPU-Z.

The "GPU Video Clock" is the "GPU Video Clock" for Nvidia reported by HWiNFO

The "GPU Video Clock" is unknown for AMD, but just post GPU clock/ GPU Memory reported by GPU-Z.

I advise you and everyone else, against to run the tests with monitoring tools and benchmark procedure running at the same time.

So, please don't run at the same time tests and monitoring tools.

Run the GPU-Z and HWiNFO without reporting benchmark values of HW decoding.

I just need the GPU Video Clock from these monitoring tools while you run the benchmark in decode/playback mode.

[Lollerke]

Your results are freaking fast.

I compare your RX 460 with my RX 470 and 460 is way faster than mine.

Your HD 610 and GTX 1050 are way faster than HD 610 and GTX 1050 of Kadano.

So, what's your secret ?

Do you cherry pick the results based on the maximum figures ?

Do you have a completely clean installation with no apps running in the background ?

One last request:
Can you post also your GPU Video Clock for HD 610, GTX 1050 and RX 460 ?

Thanks!

I think it's fast because he has dual channel DDR4 2133MHz RAM for the HD610 :) Don't know about the other VGA's tho, but GPU clock shouldnt affect the decoding speed, because all of these cards decode videos purely by the video decoding engine and there's no need for the GPU's graphics processing power.

[NikosD]

Hi,
I didn't cherry pick, I did the tests with either three or four passes as I described.

I've just installed Creators Update (clean install) and latest drivers 17.4.1

I re-ran the benchmark for the first clip and I saw a HUGE performance gain ~40%.

I'll post my results later for all the clips.

Yes, your 460 figures seem reasonable right now.

Do you have creators update ?

[CruNcher]

I think it's fast because he has dual channel DDR4 2133MHz RAM for the HD610 :) Don't know about the other VGA's tho, but GPU clock shouldnt affect the decoding speed, because all of these cards decode videos purely by the video decoding engine and there's no need for the GPU's graphics processing power.

Unless you don't activate the internal PP Engines which can run over the shaders from within the driver directly or do Hybrid Compute Based Decoding/Encoding (OpenCL/DirectCompute/Cuda) that's correct.

[NikosD]

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

Decode: fps [53-83]
Playback [1280x720]: 55 fps [42-74]
GPU Video clock during decoding/playback: 1582 MHz



This one is missing the Decode Average value.

[NikosD]

Hi,
I didn't cherry pick, I did the tests with either three or four passes as I described.

I've just started posting your results on the general table and I realized that most of the times the min - avg - max values are all equal.

Also, many times the first value (min) is just one frame bigger than middle value (average)

How is it possible ?

The clips are not decoded that fast in order to have such issues.

I think your results are rounded by you in a little weird way !

Can you take a look of your results ?

[NikosD]

I have updated my RX 470 results in the general table, after the new OS+drivers.

I ran the tests after the clean installation of Creators Update + 17.4.1 drivers without any monitoring tools running or other background apps.

I re-ran the tests with all of my monitoring tools running and after finishing some basic apps installation, including background apps and there were no difference than the clean installation.

I haven't seen before such a performance gain regarding HW decoding.

The decode and playback performance of H.265 is more about 40% - 50% higher than before.

Probably the new drivers pushed the ASIC clock higher, I can't think of anything else (I don't believe it's the Creators Update)

P.S

There is also a significant gain ~25% for H.264 decoding

[P.J]

http://www.anandtech.com/show/11266/cyberlink-launches-powerdvd-17-with-uhd-bluray-and-vr-hmd-support

=)

[NikosD]

The TDP for the new RX 500 AMD cards has gone up in general, but for video playback is lowered down significantly due to a new intermediate memory state which seems to affect a lot the power consumption of the whole card due to the power hungry GDDR5 memory type.

So, instead of going from 300MHz (idle) to 2000MHz (full) during video playback it goes to 1000MHz and the power consumption is significantly lower than RX 400 series and even Pascal 1060 for high bitrate H.264 & HEVC clips.

Read more here:
http://www.anandtech.com/show/11278/amd-radeon-rx-580-rx-570-review/2

[huhn]

how they are managing to get 93-96 watt with video playback and 74 watt idle comes down is beyond me. the bad choice of PSU has something to do with it but 74 idle is just...

but yes the 1060 with the current driver is using a very high mem clock. it is usually 202.5 to 405 mhz for hardware decoding.

[pacuro]

Do we benchmark also VP9 decoding and playback in this thread? DXVA works well with 8K60 VP9 8bit 66Mbps clip well on my GTX1060 + last nightly LAV Filters, but I am not able to run DXVA Checker benchmark due to some unidentified error (empty with yellow exclamation mark)

http://static.frazpc.pl/board/2017/04/d2c705df778b4eb1.jpg

[NikosD]

...but I am not able to run DXVA Checker benchmark due to some unidentified error (empty with yellow exclamation mark)


Did you try latest 3.15.2 ?

DXVA checker had a few issues lately.

[pacuro]

Yes, 3.15.2.
Previous 3.15.0 ends with the same empty window with yellow "!"

[NikosD]

Yes, 3.15.2.
Previous 3.15.0 ends with the same empty window with yellow "!"

DXVA Checker v3.15.3 is out and fixed the problem I reported to the developer.

If the latest version doesn't fix your issue, maybe you should report that to the developer.

[sneaker_ger]

Latest LAV x64 nightlies are up to 10% faster for software decoding of HEVC compared to older versions here (Win 7x64, i5-2500K, Ducks Take Off sample). This is the biggest, if not the only, speedup I've seen to HEVC software decoding since 2014.

[clsid]

29-10-2013 | LAV 0.58.2-97 x64 | 68,6 fps
30-10-2013 | LAV 0.58.2-104 x64 | 131,2 fps
03-08-2014 | LAV 0.62.0-8 x64 | 231,0 fps
03-08-2014 | LAV 0.62.0-9 x64 | 280,3 fps
19-01-2015 | LAV 0.63.0-41 x64 | 346,5 fps
13-07-2016 | LAV 0.68.1-31 x64 | 370,4 fps
23-04-2017 | LAV 0.69.0-39 x64 | 381,7 fps

TearsOfSteal 1080p sample