Evaluation of HEVC decoders (SW, Hybrid and HW)

[CruNcher]

I have sold my Reference RX 480. So I won't be able to provide benchmarks for it sorry. In the near future I will purchase a GTX 1060 as I feel it will better suit my encoding, decoding needs.

Makes no sense and is not logic @ all without proper testing to many are brainwashed currently by the Overall Power Output results im afraid.

Can't believe you had the chance and missed it, anyways thx for the GTX 1060 VPX results :D

Which card was it exactly the Reference ?

[Trevonn]

Makes no sense and is not logic @ all without proper testing to many are brainwashed currently by the Overall Power Output results im afraid.

Can't believe you had the chance and missed it :D

I didn't care about the power output issue it had no factor in my decision at all. AMD sucks at supporting their cards media capabilities it's that simple. Their media SDK was last updated when? 01/29/2015. NVIDIA? 7/18/2016.

Already now I can encode 10-bit 4:4:4 HEVC, decode VP9 up to 8K. The applications I use like staxrip, hybrid, OBS all have first class support for NVENC whereas with VCE, support is less than satisfactory or non-existent.

Finally based on the benchmarks I've seen the 1060 is way ahead of the 480 in terms of decode performance.

[huhn]

if you reached a decoding speed of way more than real time than it doesn't matter if a card can decode faster or not.

and because a video card has an encoder that doesn't mean it is any good...

[CruNcher]

jep especially if the encoder/decoder can operate more efficiently from the heart of the system instead of shuffling data inefficiently around the pci-e bus ;)

Though some would very well argue here being the VRAM closer to the GPU for Recording and keeping the overall latency low.

Thus also why the new AMD SSG concept is interesting, moving the whole data as close as possible to the GPU for decoding/transcoding avoiding most of the latency :)

@huhn
though it can very well be a indication of better overall efficiency doesn't have to but can be which overall can result in lower power requirements at realtime playback.

And Nvidias whole architecture since the inclusion of their Tegra R&D and major introduction with Maxwell is based on Power Efficiency and High Frequencies Serial Execution at it's very core.

[Trevonn]

if you reached a decoding speed of way more than real time than it doesn't matter if a card can decode faster or not.

and because a video card has an encoder that doesn't mean it is any good...

Fair points but for what I do it is simply the better card.

[CruNcher]

true to be told for Video Playback i would prefer a Nvidia Shield and not use my PC @ all ;)

Same most probably for game recording i would build myself a Tegra Based Recording Platform and not use the PC, directly ;)

Only when we come to stuff that can't be really executed from the outside and that would be as main applications for example video editing result acceleration from within a PC Workflow as with the SSG Model for example :D

Hard todo a 4k playback platform as efficient as the Nvdia Shield in 11W, without windows bloat ;)

Which basically reflects the idle requirement of the P8 State on the GTX 1060 very well ;)

[Trevonn]

true to be told for Video Playback i would prefer a Nvidia Shield and not use my PC @ all ;)

Same most probably for game recording i would build myself a Tegra Based Recording Platform and not use the PC ;)

Only when we come to stuff that can't be really executed from the outside and that would be as main applications video editing result acceleration from within a PC Worflow as with the SSG Model for example :D

Hard todo a playback platform as efficient as the Nvdia Shield in 11W ;)

Which basicly reflects the idle requirement of the P8 State on the GTX 1060 very well ;)

I Just did a quick run through of some gameplay clips I've recorded and looked at the power usage in the application HWINFO. The whole time it was in the P8 State consuming 9-11 Watts.

Extra: With the 4K 10-bit 400 Mb/s jellyfish clip it was around 25 watts.

[CruNcher]

Gameplay clips aren't really complex at all they only carry enough complexity @ constant 60 FPS

4K 60 FPS

H.264
H.265

are far more interesting ;)

Though i would be very well interested in a 4K 60 FPS Encoding result of Star Wars Battlefront Endors Level (Speeder Bike Chase) with the improved Encoder ;)

Also be advised the drivers you have in your signature have some severe Nvdiia acknowledged DPC issues with Pascal and its PWM which can cause problems for Encoding/Decoding related tasks.

https://forums.geforce.com/default/topic/941579/geforce-1000-series/gtx-1080-high-dpc-latency-and-stuttering/post/4944053/#4944053

there is a hotfix driver for this and a new driver coming though the results look rather odd

calling this good seems like blasphemy from that Manuel ;)

WDDM 2.0 gives them big headaches currently ;)

https://a70ad2d16996820e6285-3c315462976343d903d5b3a03b69072d.ssl.cf2.rackcdn.com/18f8f7b290aafc463e12e29f5c6688d5


PS: You didn't answered which card yet if you are a "Gamer" it's sure no Reference and it sounds like you are mainly using it for that Purpose ;)

[huhn]

Fair points but for what I do it is simply the better card.

i'm most likely going for the 1060 myself.

but decoding speed is not a point for the nvidia card in my book.

[Trevonn]

I am using the Zotac GTX 1060 Mini. I also have the high DPC latency but it does not seem to affect anything from what I can see audio/visually

[Yups]

if you reached a decoding speed of way more than real time than it doesn't matter if a card can decode faster or not.


Polaris might struggle to hold 60+ fps in 4k with very high bitrate content. Based on this test: https://www.computerbase.de/2016-06/radeon-rx-480-test/7/#diagramm-videos-abspielen-h265-codec-lav-filters

Pascal is factor 2.7 faster than the RX 480 there and also slower than the shader based solutions from AMDs prior generation. And they obviously don't use a demanding sample. That would mean the H.265 10bit -2160p-400Mbps sample is too demanding when even Pascal can barely hold 60+ fps over the entire video (#1226). RX 480 might struggle with the H.265 10bit - 2160p-140Mbps sample assuming there is a factor 2.7 gap in this sample as well. Although we need more high bitrate 4k decoding tests to see how it copes with demanding 4k videos.

[JohnLai]

I didn't care about the power output issue it had no factor in my decision at all. AMD sucks at supporting their cards media capabilities it's that simple. Their media SDK was last updated when? 01/29/2015. NVIDIA? 7/18/2016.

Already now I can encode 10-bit 4:4:4 HEVC, decode VP9 up to 8K. The applications I use like staxrip, hybrid, OBS all have first class support for NVENC whereas with VCE, support is less than satisfactory or non-existent.

Finally based on the benchmarks I've seen the 1060 is way ahead of the 480 in terms of decode performance.

Nice, so, I have question on NVENC for you.
Using staxrip default CQP and lookahead 32, what is the percentage difference between HEVC 8bit and 10bit file size? (Make sure ignore the audio size) If possible, average FPS for 8bit and 10bit encode result too (and resolution of encoded video)?

[easyfab]

Nice, so, I have question on NVENC for you.
Using staxrip default CQP and lookahead 32, what is the percentage difference between HEVC 8bit and 10bit file size? (Make sure ignore the audio size) If possible, average FPS for 8bit and 10bit encode result too (and resolution of encoded video)?

+1 I'm also interested in encoding performance.
If possible some samples and/or PNSR/SSIM.
You have some samples here : https://media.xiph.org/video/derf/
park_joy would be a good choice

[huhn]

Polaris might struggle to hold 60+ fps in 4k with very high bitrate content. Based on this test: https://www.computerbase.de/2016-06/radeon-rx-480-test/7/#diagramm-videos-abspielen-h265-codec-lav-filters

Pascal is factor 2.7 faster than the RX 480 there and also slower than the shader based solutions from AMDs prior generation. And they obviously don't use a demanding sample. That would mean the H.265 10bit -2160p-400Mbps sample is too demanding when even Pascal can barely hold 60+ fps over the entire video (#1226). RX 480 might struggle with the H.265 10bit - 2160p-140Mbps sample assuming there is a factor 2.7 gap in this sample as well. Although we need more high bitrate 4k decoding tests to see how it copes with demanding 4k videos.

400 mbit has nothing to do with real world performance.

edit: my 960 plays it fine too...
hardware decoder have a "max" speed and they only loose speed with really high mbit. you can assume bit rate about the 100 mbit are no problem for 2016 GPU of cause a test would be better.

BTW. the quality of that 400 mbit file is just terrible. it has a lot of banding it is super unsharp and i can see something like macro blocks...
just a bad source pumped with way to many mbit. even a hardware encoder would do a way better job at 1/4 the mbit if the source was fine.

[CruNcher]

Polaris might struggle to hold 60+ fps in 4k with very high bitrate content. Based on this test: https://www.computerbase.de/2016-06/radeon-rx-480-test/7/#diagramm-videos-abspielen-h265-codec-lav-filters

Pascal is factor 2.7 faster than the RX 480 there and also slower than the shader based solutions from AMDs prior generation. And they obviously don't use a demanding sample. That would mean the H.265 10bit -2160p-400Mbps sample is too demanding when even Pascal can barely hold 60+ fps over the entire video (#1226). RX 480 might struggle with the H.265 10bit - 2160p-140Mbps sample assuming there is a factor 2.7 gap in this sample as well. Although we need more high bitrate 4k decoding tests to see how it copes with demanding 4k videos.

Assumptions don't bring us far we need Real results ;)

though i would agree 400 mbps surely not but that is also a extreme target no avg user needs to decode @ all ;)

based on the Computerbase test i would say most current consumer content will be fine as well.

Though i wouldn't expect todo multistream decoding on UVD as high as possible as with Nvidia VPX ;)

Normal PIP usage but most probably that's it resources left on VPX are massiv no doubt :)

http://i1.sendpic.org/i/5V/5Va4r9EajGDMMOYvX8z7IcczjoU.jpg


Now the question is how the improved 2 pass encoder does :D

And who knows maybe AMD could improve it's decoding Performance in an entire DX 12 DXVA2 Render scenario over time.

We need raw frame decoding results to no render surface @ all.

[Yups]

400 mbit has nothing to do with real world performance.



Nobody denied this.


Assumptions don't bring us far we need Real results ;)



For both sides. Huhn assumed that Polaris is more than fast enough which has not been proven. We only have a Computerbase and Notebookcheck decoding test where Polaris does much much worse for H265. And VP9 decoding is still not supported by the driver.

GTX 1080:

The new video decoder was also convincing in our review. Both 2K H.264 as well as 4K H.264 playback was smooth at high frame rates and the CPU load was low (4K 100 Mbps 138 fps, 4% CPU). HEVC was no problem, either (1080p at 565 fps, 4% CPU and 4K 267 fps at 1% CPU).

RX 480:

The new video engine can actually convince in the test. A 4K video (H.264, 100 Mbps) was played at 84 fps and with 3 % CPU load in the DVXAChecker Playback benchmark. In comparison: The GTX 1080 managed 138 fps and 4 % CPU load.

The new H.265 decoder handles a 4K 10-bit video at 72 fps and 2 % CPU load.

http://www.notebookcheck.net/Nvidia-GeForce-GTX-1080-Desktop-Review-Pascal-has-arrived.165500.0.html
http://www.notebookcheck.net/AMD-Radeon-RX-480-Review-The-fastest-Polaris-desktop-card-at-launch.168250.0.html


Factor 3.7 in the Notebookcheck test, most likely a higher bitrate video, maybe 100 Mbit like H264? The gap seems to grow with higher bitrate content.

[CruNcher]

So all in all it just doesn't certifies the DRM Requirements Nvidia can better handle this with their Falcon Core :)

So we don't know yet how much additional overhead the encrypted transmission will cost AMD though the Blu-Ray test seems to give some insight into this.

Also the AMD DXVA Checker list of uknown GUIDS seems pretty damn long compared to Nvidia :D

Polaris 10

http://www.notebookcheck.net/fileadmin/Notebooks/Sonstiges/AMD/RX_400/dxvac.png

15 uknowns

GP104

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

10 uknowns


@Trevonn

Could you post a updated shoot from your GTX 1060 GP106 and current driver release :)

[Trevonn]

http://i.imgur.com/vsV8VZv.jpg

That GP104 shot looks to be from an older DXVA Checker

[P.J]

400mbps is useless for H.265 but Prores

[CruNcher]

yes but in Trevonns updated shoot you can pretty well see that the Performance overhead is most likely for Simple 8K Decoding goals and Multi Decoding 2 Simple 4K Streams.

Nothing of this has @ all relevance for the target audience not even in the distant future ;)

Though we most probably not only see Nvidias VPX Core advancements but also their Internal Framebuffer Compression Advancements that allow to reach this Goal, which is pretty interesting the topic of their DCE R&D and how much they are away from AMD here saving bandwith cutting costs and still be more efficient or at least as efficient with a cut down Hardware Design on the Memory side of things i find that a very impressive achievement of Nvidias R&D :D

Its like they have a ultra efficient CABAC running internally over everything and in Pascal they started to improve on that further with 3D Data :D

Now imagine Nvidia is a AOM Member, so if they might have a lucky day sometime and think "hey why not contributing our research for the greater good" that would be quiet cool ;)

Most Probably Nvidia started so many Neural Network Algorithm improvement stuff by now that their own GPU is improving it's Design Partly already so to speak itself ;)

[Trevonn]

Now I just need to win the lottery so I can buy an 8K display

[Trevonn]

Nice, so, I have question on NVENC for you.
Using staxrip default CQP and lookahead 32, what is the percentage difference between HEVC 8bit and 10bit file size? (Make sure ignore the audio size) If possible, average FPS for 8bit and 10bit encode result too (and resolution of encoded video)?

Percentage difference: 5.07%

I hope this has all the information you need

http://i.imgur.com/odoq2IL.jpg

[huhn]

no b frames AGAIN...

[JohnLai]

Percentage difference: 5.07%

I hope this has all the information you need



:thanks:
6% decrease in encoding performance and 5.07% file size saving. Reasonable tradeoff.

[kral2008]

Tested by i5 6500 Desktop cpu (intel HD 530 graphics)

Astra.SES.Demo.HEVC.ts
Size: ~ 350 MB
Resolution : 3840x2160
Overall bit rate : ~ 24 Mb/s
Bit depth : 10 bits
Frame rate : 50 fps

Codec LAV filter : DXVA2 Native
Test intervals: 3 times
FPS: 28 - 39
CPU usage: 39 - 47 %

I wonder why still CPU usage is under 60% and FPS is lower than 50.
In my opinion it should load more CPU to give higher FPS than 39 frames per second.

All other acceleration technologies over loads the CPU and delivers unacceptable results.

Edit:
I couldn't install any VGA to HP ProLiant DL380 G9 or others. Only specific cards can be compatible with these servers. This means I can test xeon v4 series. Maybe I can benchmark only the 1225 V5 processor in near future.

[nevcairiel]

I wonder why still CPU usage is under 60% and FPS is lower than 50.
In my opinion it should load more CPU to give higher FPS than 39 frames per second

Not everything multi-threads fully to use all of your cores - hybrid hardware decoders are one of those things.

[CruNcher]

Yeah and it takes time ;)

See for example CoreAVC vs FFMPeg it took some time before they reached the Assembly Optimization state of Picards CPU Mastery ;)

i wonder if Mainconcept is still running behind the deblocking performance today :D

ANd still Intels result is nice compare it to the Nvidia Hybdrid H.265 8 bit Decoder and you would be shocked how bad that is and especial how good ffmpeg is on 4 CPU cores vs it by now :D
y
Though to be fair Nvidia doesn't really invest to much time into the hybdrid decoder Performance just always low complexity goals and moves on fast then to the DSP Core which we saw first in Tegra X1 and later the GTX 950 and now every Pascal could theoreticaly do 8K low complexity and also Intel always has the advantage of faster data transfer with their hybrid ;)

You can guess which is which no words or even numbers really needed and yeah you don't want to see that spikes end (even can see it at all) @ 60 fps Broadcast Target ;)

http://i1.sendpic.org/t/2y/2yvEcUL2Vt5Eb7rGfivinfdNX7x.jpg (http://sendpic.org/view/1/i/55UW30l8Oq3DEbmBUu7m3LKXAIl.png)

http://i1.sendpic.org/t/uM/uMFXoYafHHD0qz9ef2sHYqpzwIm.jpg (http://sendpic.org/view/1/i/Agsiq4mgBUQeEgEyF5ds6bH0lED.png)

CUVID looks a little better but overall even the shaders don't really seem to be utilized that much and the higher clock alone doesn't do that much more magic @ all so totaly starving :(

http://i1.sendpic.org/t/8D/8DZ8fYtygrAGCM4NFZcjm3MWtek.jpg (http://sendpic.org/view/1/i/dsRVO7fZEMzfYW8cSEK0Ld9s9k2.png)

[alfianadjeh]

anyone have been tested RX460 for hardware decoding HEVC 4k 60fps ?
I really confusing, buying GTX950 or Rx460 2gb.

[huhn]

if you want to display at UHD you should use a card with 4 GB Vram.

2GB is at it's limited when used with UHD. i'm just telling you this because i made the huge mistake and got a 2GB 960 and not the 4GB version and i got a lot of problem thanks to the 2GB Vram.

[nsnhd]

if you want to display at UHD you should use a card with 4 GB Vram.

2GB is at it's limited when used with UHD. i'm just telling you this because i made the huge mistake and got a 2GB 960 and not the 4GB version and i got a lot of problem thanks to the 2GB Vram.

That means iGPU with max 1GB shared Vram can't display 4K ? Can you explain a little more about your problems.

[clsid]

Windows itself already uses a few hundred MB of GPU memory.

If you use software decoding then only a few frames are buffered in GPU memory. Then 1 GB might be enough. If you use hardware acceleration then much more frames are being buffered, since there can be up to 16 reference frames needed in the decoding process. As an example, here a 4k 10bit video uses 1100MB on top of 200 MB used by Windows 7. But most CPUs are not fast enough for 4k HEVC, so hwaccel is usually needed.

If you use madVR then you need even more memory since that by default uses larger queues/buffers than EVR. Then you can easily go over 2GB.

[CruNcher]

The only time where Nvidias Hybrid Decoder on the GTX 970 so far is more efficient then the 4 Sandy Bridge CPU Cores with ffmpeg is with DivX Hevc 4K Stream complexity @ 8000 Kbps , there it actually can save power :)

Some glitches still persist though but much better efficiency wise especialy when you switch from DXVA Native to CUVID there the Driver Render overhead seems also lower compared to DXVA Native on the CPU with their Hybrid but surely even better manageable with DX 12 and Vulkan in Future Media Player Render Pipelines CUVID already shows no glitches in this example testcase :)

Lower HEVC Complexity Bitstream Playback Stability Test (Nvidia Hybdrid H.265 Decoder R&D)

Nvidia Hybrid (DXVA)

http://i1.sendpic.org/t/7K/7KC18t3CqIVWOShmf9DEYgrGzC3.jpg (http://sendpic.org/view/1/i/hsPUTbmSf6PoDxCkoougEXdw9K3.png)

Nvidia Hybrid (CUVID)

http://i1.sendpic.org/t/jU/jUfcD9Wm5mwcQ7rZHsmm4Ud57aX.jpg (http://sendpic.org/view/1/i/dNNyDlfTjqP4kFNyKj8Hru1ImPo.png)

FFMPEG

http://i1.sendpic.org/t/vO/vOVpYJgoTdDluYmdytOYMcDEuwS.jpg (http://sendpic.org/view/1/i/947B3s8goI0K6I2bxL919BEcGsh.png)

I guess pretty much this was also Nvidias Primary Development goal for the Hybrid Decoder Web Streaming support before the Full Hardware Core was ready ;)

and now they go up to 8K with their IP Core in Tegra and Discrete ;)


And AMD doesn't really seem fully yet to be up to it but they made advances none the less

[P.J]

if you want to display at UHD you should use a card with 4 GB Vram.

2GB is at it's limited when used with UHD. i'm just telling you this because i made the huge mistake and got a 2GB 960 and not the 4GB version and i got a lot of problem thanks to the 2GB Vram.

Would you explain more? you mean madVR?

[chros]

But most CPUs are not fast enough for 4k HEVC, so hwaccel is usually needed.
Yep, my i73630QM can't keep up with >=25fps 4k 10bit hevc (~40Mbps) material (using latest stable LAV).

[CruNcher]

On Mobile Hardware it makes almost always sense to use the fixed function parts instead of hybrid or Software but on edge cases desktop it can be quiet different which path to go (and which in the end is the most beneficial for the users use case) based on the whole system configuration especially in hybrid cases and you need todo smart dynamic decisions based on the system configuration data ;)

[kral2008]

Intel Xeon 1225 V5 :sly: on HP ML10 Gen9 Server and Windows 10 Ent x64

Still stupid results :eek:

Astra.SES.Demo.HEVC.ts
ID : 255 (0xFF)
Menu ID : 10202 (0x27DA)
Format : HEVC
Format/Info : High Efficiency Video Coding
Format profile : Main 10@L5.1@Main
Codec ID : 36
Duration : 2 min
Width : 3 840 pixels
Height : 2 160 pixels
Display aspect ratio : 16:9
Frame rate : 50.000 FPS
Color space : YUV
Chroma subsampling : 4:2:0
Bit depth : 10 bits



Decoder: LAV Video Decoder with DXVA2 Native Enabled:
Decoder Device: HEVC_VLD_Main10
Processor Device: BF752EF6-8CC4-457A-BE1B-08BD1CAEEE9F
Frames: 6400
Time: 184.460 sec
FPS: 34.696 [28-42] fps
CPU Usage: 36 [22-45] %
GPU Usage: -

Decoder: LAV Video Decoder
Decoder Device: -
Processor Device: BF752EF6-8CC4-457A-BE1B-08BD1CAEEE9F
Frames: 6400
Time: 214.116 sec
FPS: 29.890 [27-33] fps
CPU Usage: 89 [13-98] %
GPU Usage: -

Decoder: Microsoft H265 Video Decoder MFT
Decoder Device: HEVC_VLD_Main10
Processor Device: BF752EF6-8CC4-457A-BE1B-08BD1CAEEE9F
Frames: 6397
Time: 184.580 sec
FPS: 34.657 [28-43] fps
CPU Usage: 35 [23-47] %
GPU Usage: -

[wanezhiling]

Xeon 1225 V5 belongs to Skylake family which doesn't support HEVC 10-bit decoding at all

[nsnhd]

Intel Xeon 1225 V5 :sly: on HP ML10 Gen9 Server and Windows 10 Ent x64

Still stupid results :eek:

Astra.SES.Demo.HEVC.ts
ID : 255 (0xFF)
Menu ID : 10202 (0x27DA)
Format : HEVC
Format/Info : High Efficiency Video Coding
Format profile : Main 10@L5.1@Main
Codec ID : 36
Duration : 2 min
Width : 3 840 pixels
Height : 2 160 pixels
Display aspect ratio : 16:9
Frame rate : 50.000 FPS
Color space : YUV
Chroma subsampling : 4:2:0
Bit depth : 10 bits



Decoder: LAV Video Decoder with DXVA2 Native Enabled:
Decoder Device: HEVC_VLD_Main10
Processor Device: BF752EF6-8CC4-457A-BE1B-08BD1CAEEE9F
Frames: 6400
Time: 184.460 sec
FPS: 34.696 [28-42] fps
CPU Usage: 36 [22-45] %
GPU Usage: -

Decoder: LAV Video Decoder
Decoder Device: -
Processor Device: BF752EF6-8CC4-457A-BE1B-08BD1CAEEE9F
Frames: 6400
Time: 214.116 sec
FPS: 29.890 [27-33] fps
CPU Usage: 89 [13-98] %
GPU Usage: -

Decoder: Microsoft H265 Video Decoder MFT
Decoder Device: HEVC_VLD_Main10
Processor Device: BF752EF6-8CC4-457A-BE1B-08BD1CAEEE9F
Frames: 6397
Time: 184.580 sec
FPS: 34.657 [28-43] fps
CPU Usage: 35 [23-47] %
GPU Usage: -

Try CPU decode with MPC-HC Lav 0.68.1 in EVR mode, you may get 50 fps

[kral2008]

Xeon 1225 V5 belongs to Skylake family which doesn't support HEVC 10-bit decoding at all

It supports HEVC_VLD_Main10 as other Skylake family members, but only Hybrid I think.

[huhn]

That means iGPU with max 1GB shared Vram can't display 4K ? Can you explain a little more about your problems.
with widnows you are easily using 300 mb Vram without doing anything.

if i remember correctly 1.4 GB Vram usages was the lowest i have seen with hardware/software UHD decoding on UHD screen and that with everything on min. and other programs can and will use Vram too.

Would you explain more? you mean madVR?

i have to use a GPU buffer of 4 with madVR. madVR doesn't really use more than EVR with default settings. but i'm still pretty much maxed out.

i can re test this maybe next week sorry.

[chros]

with widnows you are easily using 300 mb Vram without doing anything.

if i remember correctly 1.4 GB Vram usages was the lowest i have seen with hardware/software UHD decoding on UHD screen and that with everything on min. and other programs can and will use Vram too.

i have to use a GPU buffer of 4 with madVR. madVR doesn't really use more than EVR with default settings. but i'm still pretty much maxed out.
Thanks for the info, huhn!

[wanezhiling]

It supports HEVC_VLD_Main10 as other Skylake family members, but only Hybrid I think.

'Hybrid' means null :)

[CruNcher]

with widnows you are easily using 300 mb Vram without doing anything.

if i remember correctly 1.4 GB Vram usages was the lowest i have seen with hardware/software UHD decoding on UHD screen and that with everything on min. and other programs can and will use Vram too.



i have to use a GPU buffer of 4 with madVR. madVR doesn't really use more than EVR with default settings. but i'm still pretty much maxed out.

i can re test this maybe next week sorry.

Yes this part of the VRAM is for Rendering the Desktop but you hardly realize this loss and you get a lot of benefits from it and the (LDDM) WDDM DWM Render Model since it was introduced with Longhorn back then in efforts to gain Apples GPU Desktop Efficiency Level.

Though the overall Memory usage up today is dependent on more factors like if you use Variable Refresh or a fixed Refresh Rate 10bit 8 bit hdr or not since Windows 8.1 and 10 there is no real differentiation anymore and it's a fully fixed part integrated and in Windows 10 it drives the whole UWP experience between Windows Devices :)

No 3rd Party Media Player yet uses it fully efficient on Windows 10 and it's many Display changes yet but also drivers are still in a changing stability state ;)

There is no 3rd Party Player yet even existing with a full DX 12 Render Pipeline not even a simple ported one ;)

[P.J]

Don't waste time on 'Hybrid'

with widnows you are easily using 300 mb Vram without doing anything.

if i remember correctly 1.4 GB Vram usages was the lowest i have seen with hardware/software UHD decoding on UHD screen and that with everything on min. and other programs can and will use Vram too.



i have to use a GPU buffer of 4 with madVR. madVR doesn't really use more than EVR with default settings. but i'm still pretty much maxed out.

i can re test this maybe next week sorry.

Then it happens only with madVR not EVR :)

[CruNcher]

Don't waste time on 'Hybrid'



Then it happens only with madVR not EVR :)

No surprise EVR is tuned for System Resource efficiency ontop of DWM not Extreme Quality ;)

[huhn]

UHD playback Vram usages using a UHD source HDR processing is disabled in madVR because EVR doesn't know what HDR is:
lavfilter 0.68.1
mpc-hc 1.7.10.252
clip sony 4K HDR Camp
gpu 960 with a boost clock of 1430 mhz

copyback
madVR GPU queue at 4 1547-1750 mb
EVR CP queue 4 1922-2000 mb

native
madVR GPU queue at 4 present queue 3 1600 mb BLACK SCREEN in full screen.
EVR CP queue 4 1780 mb

a difference of 100 of mb each time are pretty normal.
get more than 2 gb that's all i have to say to this.

the default GPU queue for madVR is 8 which is unusable for UHD with 2 GB Vram. the default GPU queue for EVR is 5 or 6.

GPU usages was ~70% with both EVR and madVR. madVR was at default except queue and FSE mode.

[chros]

get more than 2 gb that's all i have to say to this.
Thanks for the detailed tests!

[leonccyiu]

http://www.notebookcheck.com/Kaby-Lake-Core-i7-7500U-im-Test-Skylake-auf-Steroiden.172422.0.html

Notebookcheck has a DXVA checker screenshot from the Kaby Lake HD650 Graphics

HEVC Main 10 as well as VP9 Profile 0 and 10 bit are now decoded up to 8k.

http://www.notebookcheck.com/fileadmin/Notebooks/MSI/CX72-7QL/dxva_kaby.png

[NikosD]

Nice info for Kaby from Anandtech.

It's all Media update the Kabylake update.
The gain of CPU is the gain of MHz due to the advanced 14+ nm process.
The GPU is the same Gen 9.

But the media capabilities are impressive.

http://www.anandtech.com/print/10610/intel-announces-7th-gen-kaby-lake-14nm-plus-six-notebook-skus-desktop-coming-in-january

[chros]

But the media capabilities are impressive.
Indeed! :)