Evaluation of HEVC decoders (SW, Hybrid and HW)

[CruNcher]

Doesn't mean much read the footnotes and it also doesn't explain how the Decoding is actually implemented that's entirely up to Nvidia ;)

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

Thats important but they don't say exactly which they keep it in the dark ;)

Even as a Member you cant seem to access

https://developer.nvidia.com/video-encode-decode-gpu-support-matrix

currently ;)

[wanezhiling]

Can someone with GP104 or GP106 install the latest Nvidia 375.95 or later drivers and run the latest DXVA Checker to see if those GPUs support VP9 Profile2?
http://i.imgur.com/m20eY2v.jpg
No luck, same 375.95 driver

[CruNcher]

** Max resolution support is limited to selected Pascal chips
*** VP8 decode support is limited to selected Pascal chips

It doesn't say anything about VP9 10bit Profile2 support, don't need to be snarky and a troll.

You really don't understand how VPX works.

Though 1070/1080 GP104 is not so important GP106 1060 is more important ;)

[NikosD]

It looks like GP107 has the most advanced video decoder right now, even more than Kabylake because i have never seen anywhere that Kaby supports 12bit HEVC which is the only thing missing though.

On the other hand, we have to see also HW encoding capabilities especially for HEVC and of course speed.

My opinion is that Kaby is faster than Pascal in all formats supported by both.

Also desktop Kabylake supports HDCP 2.2, it's not like Kabylake for laptops.

We'll see in less than two months.

[CruNcher]

Of course it will be faster it doesn't have to fight like Pascal for Performance on the PC but wait for the new Pascal Tegra and how it will kill it ;)

Though on Mobile Pascal it could look slighty different also ;)

i want to see the new Kabylake based Atom in Action for 200$ complete system :D


I find it great that Nvidia by the way as mostly the first advises this

Performance represents measured average performance and quality of different classes of videos (camcorder, gaming, screen, synthetic, and telepresence). Performance may vary based on OS and software versions, and motherboard configuration.

[eddman]

Doesn't mean much read the footnotes and it also doesn't explain how the Decoding is actually implemented that's entirely up to Nvidia ;)



Thats important but they don't say exactly which they keep it in the dark ;)

Even as a Member you cant seem to access

https://developer.nvidia.com/video-encode-decode-gpu-support-matrix

currently ;)

I think those footnotes are in regard to GP100 chips which are feature set G.

If GP102, 104, 106 and 107 chips had different hardware accelerating capabilities, nvidia would not have used the same feature set designation, H in this case, for all of them.


http://developer.download.nvidia.com/designworks/video-codec-sdk/secure/7.1/01/NVDEC_VideoDecoder_API_ProgGuide.pdf?autho=1480256815_a4ba43531ca6c92c72e50a50086f8eaf&file=NVDEC_VideoDecoder_API_ProgGuide.pdf

Page 8. GP100 is listed separately from others. It seems the only difference is that GP100 cannot do 8K coding.

http://i.imgur.com/mqGI6Up.png


Tesla P100 devices, which use GP100 chips, are listed as feature set G:

http://us.download.nvidia.com/XFree86/Linux-x86_64/375.20/README/supportedchips.html

http://i.imgur.com/82ZLUqj.png


I think the reason for 1050 cards being VP9 10-bit enabled, is down to drivers. Perhaps nvidia is yet to enable it for other Pascal cards.

Either that, or nvidia made a mistake in feature set naming and GP107 is actually feature set I or something.

[CruNcher]

If GP104, 106 and 107 chips had different hardware accelerating capabilities, nvidia would not have used the same feature set designation, H in this case, for all of them.

This is exactly what suprises me currently but anyway it could support it but maybe not fixed function but via Nvidias CUDA Research Decoder though it's unlikely they would hit the same performance then on AVG Joe Consumer Cards.

I think the reason for 1050 cards being VP9 10-bit enabled, is down to drivers. Perhaps nvidia is yet to enable it for other Pascal cards.

Possible but Review time is long over ;)

Either that, or nvidia made a mistake in feature set naming and GP107 is actually feature set I or something.

Would be a funny Intern error again of course ;)

[eddman]

maybe not fixed function but via Nvidias CUDA Research Decoder

You mean VP9 10-bit? Hmm, yes, maybe it's just hybrid decoding, but even if that's the case, it should be possible for nvidia to enable it on all feature set H cards with future drivers unless, again, they made a mistake and GP107 is a higher feature set chip.

[CruNcher]

yep it's not bad performing on the HEVC side of things at all up of course until a certain complexity ratio and they surely could have improved it even further but they just stooped.

So it would be nice to see how they do with VP9 Decoding on GP104/GP106 with it :)

http://i1.sendpic.org/t/yK/yKIrDmtcdnUzg6RaBHUXd1buSSb.jpg (http://sendpic.org/view/1/i/5uOA8BKxrxKzM8Euj6JiNWFTvRv.png)

It makes me very curious what they reach so far on their entirely own GPU optimized Codec R&D :)

[NikosD]

Please don't forget that GP107 is not like the other Pascal processors.

They are manufactured in different process 14nm, instead of 16nm, by Samsung.

So it is very much possible to have really different fixed-fuction video decoder inside, that Nvidia should mention of course.

[Yups]

I wonder if the decoding/encoding speed/quality differs from bigger Pascal cards, beside VP9 profile2 of course. As a GTX 1080 owner it's a bit disapppointing that the bigger cards are less capable once again, same with Maxwell v2. Do we have DXVA checker screenshots from Polaris and Bristol Bridge with current drivers?

[eddman]

Please don't forget that GP107 is not like the other Pascal processors.

They are manufactured in different process 14nm, instead of 16nm, by Samsung.

So it is very much possible to have really different fixed-fuction video decoder inside, that Nvidia should mention of course.

One has nothing to do with the other. Two chips can be built on entirely different processes and by different manufacturers and yet have the exact same hardware capabilities.

It's all up to the designer of those chips. If nvidia made them to have the same features, then they have the same features.

If turns out that GP107's video capabilities are in fact higher than other Pascal chips, then it was a deliberate design decision by nvidia to upgrade the video block, and not because GP107 is 14nm and the others are not.

So far all Pascal chips, except for GP100, are feature set H, so they SHOULD be the same, but since nvidia sometimes messes things up for no reason, I wouldn't be surprised if GP107 turns out to be feature set H2 or I or something.

We have to wait and see.

[CruNcher]

Im especially interested how it looks on the Mobile Parts now get more and more curius of that because of the weak reviews on such very specific things :D

Can't wait to see NikosD GCN UVD Polaris evaluation there at least i know someone with the capabilities of understanding it does it right even if im not such a fan of a simple synthetic benchmark and we have the winner approach like NikosD follows in it's core and shares unfortunately with most reviewers these days ;)

[NikosD]

One has nothing to do with the other. Two chips can be built on entirely different processes and by different manufacturers and yet have the exact same hardware capabilities.




As a GTX 1080 owner it's a bit disapppointing that the bigger cards are less capable once again, same with Maxwell v2.


It is a very well known tactic since ATI HD 2000 series and probably in older cards too, that smaller GPU chips by manufacturing process (14nm vs 16nm) or by number of shaders (1080 vs 1050) can hold bigger/better video decoders.

It's all about size and space for video decoding inside the GPU chip.

Of course from a commercial point of view, very few people really care about video transcoding features of a card like 1080.

[nevcairiel]

As a GTX 1080 owner it's a bit disapppointing that the bigger cards are less capable once again, same with Maxwell v2

(Assuming its actually a newer decoder in GP107: )

1080/1070 is about half a year older now, so advances can be expected eventually, even if technically still in the same "series".
They could delay these additions until the next full gen, but we wouldn't really benefit from that either, would we.

Its not really about the 1050 being smaller and therefor having more space, its just 6 month newer. The next gen "big" cards will have the same new "bigger" decoder as well.

[CruNcher]

And because of the Higher Shadercount they calculate if it wouldn't be more feasible to use these Shaders for Decoding and stress the CPU therefore a bit more ;)
Nevcariel these decisions are way over a year old what we see is the turn out of simple economic balance and Design decisions about the VPX target and practical goals for specific release schedule it's like partly planing into the future with predicted and existing knowledge of how it will turn out ;)

[eddman]

It is a very well known tactic since ATI HD 2000 series and probably in older cards too, that smaller GPU chips by manufacturing process (14nm vs 16nm) or by number of shaders (1080 vs 1050) can hold bigger/better video decoders.

It's all about size and space for video decoding inside the GPU chip.

Not really. As nevcairiel pointed out, smaller chips usually come, sometimes by a few months, after their bigger counterparts, so it's not surprising that they usually come with better video blocks too.

Manufacturing processes dictate very little when it comes to video engines.

I really doubt that GP107 is different (it's probably a driver feature difference), but I've been wrong about nvidia before, so who knows.

[NikosD]

Its not really about the 1050 being smaller and therefor having more space, its just 6 month newer. The next gen "big" cards will have the same new "bigger" decoder as well.

Not really. As nevcairiel pointed out, smaller chips usually come, sometimes by a few months, after their bigger counterparts, so it's not surprising that they usually come with better video blocks too.


No. It's not that.

Guys, for me it's really obvious that you don't have a clear picture of the subject, lack of knowledge I think.

Since you insist, I have to remind you the first Fermi card of Nvidia that had VP5 decoder was the smallest one - the low entry budget GT520.

Just one month before the release of GT520 the Fermi cards had VP4 inside.

Also, one month AFTER the release of GT520, the 560 Ti card had VP4 decoder.

Release time to market has nothing to do with the new video decoders.

I hope you don't need more proofs.



And because of the Higher Shadercount they calculate if it wouldn't be more feasible to use these Shaders for Decoding and stress the CPU therefore a bit more ;)
Nevcariel these decisions are way over a year old what we see is the turn out of simple economic balance and Design decisions about the VPX target and practical goals for specific release schedule it's like partly planing into the future with predicted and existing knowledge of how it will turn out ;)

Well said, once again.

[huhn]

first problem is these cards are 5 years old and the "new" 560 ti was just a GF110/GF114 these chips where produced already before the 560 ti so they had the same VP.

the GT520 was so slow that the only thing it could do was video decoding.

[NikosD]

I hope some time such completely irrelevant and wrong posts to stop, taking space from the thread and confusing people.

The only purpose besides trolling is to prove me wrong, which is a very very very difficult thing to do :D

[huhn]

you know what happens last time his ego took over him...

[nussman]

I hope some time such completely irrelevant and wrong posts to stop, taking space from the thread and confusing people.

Just stop posting ... :cool:

[NikosD]

Hello world!

We have a troll party here.

May 2011 Nvidia released 560 and November 2011 release 560 ti with 448 cores.

Both AFTER GT520 and both with VP4 and not VP5 decoder

[NikosD]

Hope you guys are no more than 20 or 25 years old.

Because if you are older than that...

[NikosD]

You should have been banned long time ago at least from this thread permanently and probably from the whole forum too.

If it was in my hand I would definitely do it, but lucky you I'm not the admin or even a moderator.

This is the last time ever that I even refer to you.

[eddman]

No. It's not that.

Guys, for me it's really obvious that you don't have a clear picture of the subject, lack of knowledge I think.

Since you insist, I have to remind you the first Fermi card of Nvidia that had VP5 decoder was the smallest one - the low entry budget GT520.

Just one month before the release of GT520 the Fermi cards had VP4 inside.

Also, one month AFTER the release of GT520, the 560 Ti card had VP4 decoder.

Release time to market has nothing to do with the new video decoders.

I hope you don't need more proofs.


IINM, the first GF119 chips were launched on Jan 5, 2011, in the form of 410M and GT 520M.

The first GF114 card, 560 Ti, was launched on Jan 25, 2011.

Yes, GF114 launched 20 days after GF119 and still came with a VP4 engine, but do we know for sure it was only because of die space issues and no other reason?

I remember reading at anandtech that back then nvidia simply would not put better video blocks in mid to high-end cards becasue there was no need for them in the market and to maybe save some money in the process by going with the old design. Video decoding capabilities were focused on low-end cards which were the main go-to cards for HTPC users.

I've been following GPUs since, I think, 2002 and don't remember reading that it was solely because of die space constraints but it's possible I missed that information somewhere.

It's plausible that die space could play a role here, but is it the only player?

If you have a source or something, would you please link it here? It never hurts to read new information.

The only purpose besides trolling is to prove me wrong, which is a very very very difficult thing to do :D

Let's not call the fellow posters trolls, shall we? There is also no reason to bring in ego. I'm looking for truth and facts, not proving others wrong.

[huhn]

and both are old chip with disabled shader parts GF110 and GF100...

[nevcairiel]

I hope you don't need more proofs.

One key point you are ignoring is that its based on the chip - if they re-release an older chip a year later its clearly not going to have more features, and the 560 Ti is the same chip was the 560, and the 560 Ti 448 (the one released in november) is an even older chip (a downclocked GF110)

GF114 was released in January 2011, that includes 560 and 560 Ti.
GF110 was released in December 2010, thats the 560 Ti 448 (which was released much later, in November 2011)

GF117 was released in April, thats the GT520

So even though its only 3 month or so, there is still a clear timeline.
But clearly you prefer calling people trolls.

[eddman]

and both are old chip with disabled shader parts GF110 and GF100...

GF114 and GF119 were new designs; not disabled GF100 and GF110.

One key point you are ignoring is that its based on the chip - if they re-release an older chip a year later its clearly not going to have more features, and the 560 Ti is the same chip was the 560, and the 560 Ti 448 (the one released in november) is an even older chip (a downclocked GF110)

GF114 was released in January 2011, that includes 560 and 560 Ti.
GF110 was released in December 2010, thats the 560 Ti 448 (which was released much later, in November 2011)

GF117 was released in April, thats the GT520

So even though its only 3 month or so, there is still a clear timeline.
But clearly you prefer calling people trolls.

Actually, NikosD is not wrong. GF114 came after GF119 and there are other such examples in other generations, not all of them though.

The point of the debate was that this is done due to die space constraints and limitations. I argued that it's not just that, but mostly a market/cost decision. It is possible to put the same video engine in a big die if there is a need for it.

[NikosD]

Guys please, do me a favor.

Ask a hardware guy to explain you how many YEARS need a new ASIC like a video decoder for R&D and how many million dollars are needed for that and then write your opinions.

Or just read CruNcher's replies who obviously knows about HW.

It's been fun, thank you.

[eddman]

Guys please, do me a favor.

Ask a hardware guy to explain you how many YEARS need a new ASIC like a video decoder for R&D and how many million dollars are needed for that and then write your opinions.

Or just read CruNcher's replies who obviously knows about HW.

It's been fun, thank you.

Implying I don't know anything? Let's not assume things like that about each other. I did not call you anything so I expect the same.

I did mention costs and market in my comment up there. You and CruNcher are very much right about those.

I was just saying that die sizes alone don't dictate everything. It's not a "won't fit on the die" situation.

It's been fun, thank you.

[pirlouy]

Or just read CruNcher's replies who obviously knows about HW.
This guy...
I've been hiding these CruNcher's posts for long time. As far as I remember, the guy was doing very big posts with 0 useful info inside (and useless big signature obviously). Talking about this guy as a reference is the worth you could do.

But even if you're right or wrong, your arrogance is boring to read. I really tried to find a particular sense of humour, but there's definitively not.

[Trevonn]

GTX 1060 also doesn't have VP9 10-bit decode. 375.95

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

[CruNcher]

Which becomes economic wise really interesting as if Nvidia locked the requirements for a newer VPX now basically to a very specific Performance range since Maxwell actually.

As the 1050 (Ti) GP107 is the same Performance range as the old 960 GM206 ;)

since everything shifted with the 16/14 nm introduction by 1 chip and +-10 Card number in overall favor of the Energy Efficiency.

I become more and more curious about the actual Mobile Parts VPX status :D

Mobile Part Decisions have very much longer impact on the Avg Consumer user and can't be really changed at all so decisions there have to be taken with greater care for the right time or else you might lose a possible content consumer as your target.

[Trevonn]

The 1050 ti came 4 months after the previous VPX version (June - Oct).

The GTX 960 did the same thing (Sep - Jan).

It's just a natural progression of technology nothing more. NVIDIA starts at the top end first so of course the weaker cards which are released months after the big GPUs will possibly have improved decoding/encoding capabilities

[NikosD]

Some people here really don't understand the concept, while others really trying to make them understand.

[pacuro]

The PureVideo VP8 decoder of Pascal GP107 (GTX 1050 & 1050Ti) has already added 10-bit VP9 HW decoding support!!!:p:p:p

Could you please show us DXVA Checker screenshot with GUID visible or just paste string.

[wanezhiling]

NVIDIA GeForce GTX 1050 Ti
MPEG2and1_VLD: DXVA2/D3D11, SD / HD / FHD / 4K
MPEG2_VLD: DXVA2/D3D11, SD / HD / FHD / 4K
MPEG1_VLD: DXVA2/D3D11, SD / HD / FHD / 4K
VC1_VLD2010: DXVA2/D3D11, SD / HD / FHD
VC1_VLD: DXVA2/D3D11, SD / HD / FHD
32FCFE3F-DE46-4A49-861B-AC71110649D5: DXVA2/D3D11
H264_VLD_Stereo_Progressive_NoFGT: DXVA2/D3D11, SD / HD / FHD / 4K
H264_VLD_Stereo_NoFGT: DXVA2/D3D11, SD / HD / FHD / 4K
H264_VLD_NoFGT: DXVA2/D3D11, SD / HD / FHD / 4K
HEVC_VLD_Main: DXVA2/D3D11, SD / HD / FHD / 4K / 8K
HEVC_VLD_Main10: DXVA2/D3D11, SD / HD / FHD / 4K / 8K
20BB8B0A-97AA-4571-8E99-64E60606C1A6: DXVA2/D3D11
15DF9B21-06C4-47F1-841E-A67C97D7F312: DXVA2/D3D11
MPEG4pt2_VLD_Simple: DXVA2/D3D11, SD / HD / FHD
MPEG4pt2_VLD_AdvSimple_NoGMC: DXVA2/D3D11, SD / HD / FHD
9947EC6F-689B-11DC-A320-0019DBBC4184: DXVA2/D3D11
MJPEG_VLD_PureVideo: DXVA2/D3D11, SD / HD / FHD / 4K
VP9_VLD_Profile0: DXVA2/D3D11, SD / HD / FHD / 4K / 8K
VP9_VLD_10bit_Profile2: DXVA2/D3D11, SD / HD / FHD / 4K / 8K
DDA19DC7-93B5-49F5-A9B3-2BDA28A2CE6E: DXVA2/D3D11
6AFFD11E-1D96-42B1-A215-93A31F09A53D: DXVA2/D3D11
914C84A3-4078-4FA9-984C-E2F262CB5C9C: DXVA2/D3D11
4B2E068F-396A-43DF-B86B-7DC6D2B6726A: DXVA1
312B6A3A-41DE-4A43-9C0F-FB3D5F0DFB9A: DXVA1

[pacuro]

@Wanezhiling
I rather thought about GUID corresponding to VP9_VLD_10bit_Profile2 - try to point over this entry and rectangle with info will appear. This number I am interested in :)

[sneaker_ger]

DXVA_ModeVP9_VLD_10bit_Profile2 Profile
[...]
The associated GUID definition for the corresponding entry in the dxva.h header file is as follows:
// {A4C749EF-6ECF-48AA-8448-50A7A1165FF7}
DEFINE_GUID(DXVA_ModeVP9_VLD_10bit_Profile2,
0xa4c749ef, 0x6ecf, 0x48aa, 0x84, 0x48, 0x50, 0xa7, 0xa1, 0x16, 0x5f,
0xf7);
https://www.microsoft.com/en-us/download/details.aspx?id=49188

[wanezhiling]

http://i.imgur.com/IyNix27.png

[NikosD]

@wanezhiling

If you compare the video decoding speed of your card 1050 Ti with 1060 card and the benchmark results of Trevonn, does it have any performance difference ?

[wanezhiling]

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

I just checked this clip right now

Decode:
GTX 1050ti -59/73/93

Playback:
GTX 1050ti - 60/74/93

While this system is slow, X4 860K, 8g ram*1, no SSD...

I dont think GP107 is different with other Pascal in terms of performance, its just been added VP9 main10 ability by accident....

[NikosD]

Yes, the performance is exactly the same.

It looks like they enabled VP9 10bit only for GP107.

It's more like a commercial/marketing policy of Nvidia rather than a new decoder.

I can't believe they made two video decoders differing only in VP9 10bit decoding.

That would be insane.

But if I was a Nvidia owner of 1060/1070/1080 I would definetely post to their forums to find out what is going on.

[CruNcher]

We need to know what each off the Mobile Parts shows with the newest SDK based Driver.

The decision would be relatively strange indeed if none of the Mobile Parts supports it but surely it's not.

I dont think GP107 is different with other Pascal in terms of performance, its just been added VP9 main10 ability by accident....

The 3.5 GB decision for the GTX 970 wasn't either by accident it was planed also to keep it out of the press to allow a very competitive priced product to enter no one expected to see from Nvidia since the GTX 460.

Nothing happens by accident or you doing the wrong business and have failing strategies.

And i wouldn't be surprised if we see a coming Tegra Pascal supporting everything we talk about maybe a Shield 2 Nvidia needs to update their Android TV Platform it's a very important target and CPU is out of question for the Decoding complexity we are talking about here.

It is so much higher it's crazy and it will get even crazier in the future.

[NikosD]

If there are massive complaints in Nvidia's forums regarding the VP9 10bit addition to GP107 only, then maybe Nvidia could be forced to add it to the other Pascal cards or explain why it doesn't do it.

[CruNcher]

There is a simple mindset behind this and that is lying in the target system requirements in correlatation to the needed CPU Power for a given user target.

Gamer invest heavily into CPU Power and that is what 1080/1070 mainly target.

With this knowledge ahead you can plan wonderful Efficient strategies on your Design and Release Schedule choice and be very efficient with it, that's a very big strength of Nvidia ;)

Also intel is not so dumb as you might think intel releases a feature just in time the market needs it ;)

Nvidia calculates that the lower Cards wont be supported by a that Powerfull CPU and Hardware Decoding more essential and then they have their CUDA (Hybrid supported strategy they can at every time employ) to get at least minimal support (mostly VOD complexity requrements) on even the Gamer Cards ;)

Nvidia does it actually in a very very efficient way of thinking about it to save resources and Hardware space wherever they can to exactly reroute it where needed for their target customers.

Though currently it can be predicted that they mostly gonna to fail vs AMDs new Strategy if they don't adapt fast enough and people start to realize the problems of this approach.

Also as we see a continuously complexity increase and especially faster cycles it becomes questionable if the old ASIC approach still will hold up, thus AMD and Intel already experimenting with a FPGA approach (more release control faster time to markets).

People all the time arguing about coil whine though they don't realize how complex the whole thing really is economic/phsysical/technical and on what for a Edge we stand now.

[NikosD]

This is just an excuse.

If the decoder is exactly the same and you only need a driver to enable VP9 10bit, then the marketing segmentation of the GPUs by the drivers is awful and humiliating for the owners.

Why not a 1060 owner have VP9 10bit acceleration if he has already the hardware inside ?

[aufkrawall]

And the idea that it has been activated by mistake for a single product doesn't seem legitimate to your mind?

I'd rather have better VDPAU support (VP9 & HEVC 10 bit...) than VP9 10 bit for Windows, as long YT uses it only for HDR.

[CruNcher]

It is a possibility of course for a "bang look our customers get that support we are ready for it and now we have a surprise every Pascal supports it" ;)

I don't believe in it seeing all these footnotes though they bring you back to reality ;)

And also it would need the reviewers to redone reviews for bigger impact and yet i see none of this planned press feedback by Nvidia appearing, doesn't mean it not might show up soon ;)

But i would say there might be a chance that the Mobile Pascal parts especially in the lowest target range actually do support it fully in Hardware as well already.

And just about it it might be even that the whole Lineup actually has it inside but Nvidia decides where to make it available through the Driver sure they can do this they have the control but this is heavy speculation without someone proving it ;)

Nvidia has employed a lot of Software Locking into their Product Decisions and Strategies over the years so i wouldn't say this would be even impossible especially as hiding such a capability is a easy if switch but until no one proves that it doesn't exist ;)


If the decoder is exactly the same and you only need a driver to enable VP9 10bit, then the marketing segmentation of the GPUs by the drivers is awful and humiliating for the owners.

If you find that you should be rather scared or fascinated about the future ;)