Evaluation of HEVC decoders (SW, Hybrid and HW)

[eddman]

There are a few possibilities:

1. GP107 uses a newer feature set video block, and nvidia ...
a) made a mistake in their documents and still named it H (instead of, say, H2 or I)
or
b) is repeating a 970-memory-situation and are giving GP107 a better video block without disclosing it to the public.

2. GP107 has the same video block as other Pascal cards and nvidia is ...
a) simply testing its VP9 10-bit functionality with 1050 cards before also enabling it for other cards
or
b) trying to give GP107 based cards an, IMO, unfair edge through drivers and locking other cards out.

Personally, I think and hope it's number 2.a), but they do have a history of doing sneaky things. We won't know for sure until someone asks in their geforce and developer forums or, better yet, a reputable website contacts them and asks for explanation.

I am thinking of maybe buying a 6GB 1060 and wouldn't like to have less video features when nvidia advertises same H video blocks and equal capabilities for all geforce pascal cards.

[NikosD]

OK. That's a good post.

I'm closer to 2b, unless there is pressure from the users.

Go to Nvidia forums!

[aufkrawall]

b) is repeating a 970-memory-situation and are giving GP107 a better video block without disclosing it to the public.

Did Nvidia ever promise VP9 10 bit for the big Pascal GPUs?
I don't think I've ever seen it on any feature slide.

[sneaker_ger]

Nope. That's why everyone is so surprised to have it show up in DXVA checker.

[CruNcher]

There is also another possiblity a very specific one actually can you guess which ;) ?

Okay Okay

how about the

Polaris Mobile vs Pascal Mobile review shootout

that will come sooner or later and should be very close actually and both preparing up ;)

b) trying to give GP107 based cards an, IMO, unfair edge through drivers and locking other cards out.

it would be only unfair if it really exists in the other if it doesn't it would be a perfect valid business decision based on the cards to low shadercount to compensate with CUDA and Systems with to low CPUs using that target

From a Tech Economic Geeks point it would be a ultra efficient decision, that every shareholder would be proud off :D

[NikosD]

We have been warned by Nvidia actually from the previous Maxwell generation.

The first surprise was with GM206 with a new decoder for Maxwell and full fixed-fuction acceleration of HEVC (8bit/10bit).

And the 2nd surprise was again with GM206 (an internal surprise inside the first surprise) that supported VP9 8bit too.

Sneaky Nvidia like eddman accurately said and tricky.

[CruNcher]

That is the gist of it since Maxwell Perf/Watt is crucial in every decision they do.

And that is firiing partly back on terms of stability when it comes to a really optimized multithreaded input/output pipeline.

Nvidia started to extremely bruteforce to get that stability back and the funniest thing about it they made it a trend to see higher clocks as something cool ;)

also the GM206 VP7 was introduced in Tegra X1 (Shield Android TV) first and then later arrived in GM206 and X1 Simultaneously ;)


Im pretty unsure if Nvidia would release another Shield TV now with AV1 so close to finish though ;)

Actually the foundation that we might see VEGAs UVD already with AV1 support might be not bad at all as the first Consumer Chip, though it's a very tight schedule to hit :)

[MisterXDTV]

I have been following this thread for months:

I just have some questions: I just bought a 1050ti that will be my christmas gift this year (haven't opened it yet).

Did I make the right choice for "next-gen" video decoding and eventual UHD Blu-ray playback? Will it support all DRM and copy protection required by windows 10 for streaming 4K on video services?

I read about another codec being released (AV1?). Will it become as important as HEVC and VP9 right away??

[CruNcher]

For now and sometime in the future you are safe ;)

[sneaker_ger]

1050 ti is a very good, if not the best choice at the moment. AV1 will not become as important as HEVC and VP9 right away because no one has any hardware for it right now. It doesn't exist yet. Netflix and Amazon 4K use HEVC, Blu-ray uses HEVC. Even Youtube pushed out VP9 profile 2 and Chromecast Ultra just a few months before the sheduled release of AV1. These companies won't block all their customers out just to push a new video codec. All of them still use H.264 as fall back, for example.
As for DRM: 1050 ti and other Pascal cards are supposed to work with PlayReady 3.0 which is the very latest DRM. We don't yet know why Netflix and Microsoft don't support them yet for 4K on Windows 10. Maybe just a new driver is needed.

So I will say:
- There is no better bang/buck for video decoding than 1050 ti at the moment. If you need a card right now: go for it.
- if you don't need one right now: wait. It isn't working with Netflix 4K yet. UltraHD Blu-ray doesn't work on PC yet. There isn't Amazon 4K on PC yet at all. Pretty much only Youtube 4Kp60 requires it (unless maybe if you have a very fast CPU). No need to rush. (But also: AV1 hardware could be another year or two off. If you are only waiting for AV1 but don't plan on waiting that long: buy the 1050 ti now)

[MisterXDTV]

Thanks for your answers....

In february I also got a Blu-ray drive that can actually read (but not play) UHD Blu-rays

Do you think an updated PowerDVD is all I need for playback, or will I have to wait for AACS 2.0 to be broken?

[sneaker_ger]

Probably "only" need a PowerDVD update.
But:
1.) they haven't announced any ETA yet
2.) they haven't announced prices yet
3.) they haven't announced hardware requirements yet

So again: lots of "ifs" and "whens". Similar for AACS 2.0: no one knows when it will be broken - or if at all. We don't even know whether UltraHD Blu-ray will survive as a format long enough...

[MisterXDTV]

On that subject, I know MadVR already converts HDR10 to SDR. But what about Dolby Vision in the future?

I read that Pascal can/will support 12 bit decoding but I still don't know whether DV is just a "software" thing or needs dedicated hardware not found on Pascal cards

[pacuro]

Update on https://developer.nvidia.com/nvidia-video-codec-sdk

" **** VP9 10/12 bit decode support is limited to select Pascal chips "

[nevcairiel]

Update on https://developer.nvidia.com/nvidia-video-codec-sdk

" **** VP9 10/12 bit decode support is limited to select Pascal chips "

That still doesn't answer any questions however, because it has a similar comment about 8K support, which we know is available on GP104, for example (just not on GP100)

[pacuro]

Still no VP9 10bit hw decoder detected in new 376.09 drivers in GP106. Due to same branch probably so nothing new.

[CruNcher]

As im very interested in Nvidias CUDA Video Research here some more information about their HEVC CUDA Decoder behavior and how it fails with DXVA on EVR-CP very heavily

http://forum.doom9.org/showpost.php?p=1787887&postcount=21387

This is also interesting in General when Benchmarking the Hybrid Decoders :)

Source Nvidia SDK

While Kepler and first generation Maxwell GPUs had one NVENC engine, certain variants of the second generation Maxwell GPUs and Pascal Generation GPUs have two/three NVENC engines physically present on the silicon.

Also they added (updated) footnotes again


* HEVC/VP9 10/12 bit decoding 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
**** VP9 10/12 bit decode support is limited to select Pascal chips

Support Matrix now available as well we seem now to be in a semi stable state with information available officially :)

[CruNcher]

Nvidia Official Support Matrix Kepler,Maxwell and Pascal

NVDEC (not taking resolution/performance abilities into the matrix @ all)

http://i1.sendpic.org/i/9Y/9YyblcoNV26IRHljJ2Uyss0YBfi.png

NVENC

http://i1.sendpic.org/i/uG/uGW0ZLjYj2Cxa9LQxvilqQnF0qX.png

GP107 not added yet (no pro Quadro/Tesla/Grid product based on it yet official) most probably all ready official that it has 3 encode/decode units and matrix will be updated when pro product arrives ;)

[P.J]

http://www.notebookcheck.net/fileadmin/Notebooks/Acer/Aspire_ES1-332-P91H/dxvachecker.PNG

Intel Apollo Lake's fixed function hardware decoder supports HEVC Main10 8K & VP9 Profile0 4K.

:goodpost:

[CruNcher]

This combination might become a crazy nice testing platform for analyzing Intel and Nvidia in a very similar environment :)

Kabylake + 1050 M

If the rumors are true and the 1050 Mobile will be released (based on GP107) it could have the new VPX like the 1050/TI.

Actually i really like the SOA of the 1050 TI it is a Perf/Watt monster with SOA Decoding capabilities of all for now important solutions (pre AV1) :)

This combined with Kabylake in a Mobile Platfom (yummy yummy)

[CruNcher]

Don't we love Marketing :D

http://cdn.wccftech.com/wp-content/uploads/2016/12/Radeon-Software-Crimson-ReLive-NDA-Only-Confidential-v4-page-043-copy-840x473.jpg

Its definitely VP9 Decoding over AMD's OpenCL accelerated Decoder R&D

http://cdn.wccftech.com/wp-content/uploads/2016/12/file-page10-copy-649x840.jpg

GTX 1060 (GP104) most probably wont do much more on the Encoding side as the GTX 970 (GM204) also does the 30 fps HEVC Limit (at reasonable 15 Mbit/s)

But the GTX 1050 TI GP107 should maybe really reach the 60 FPS Encoding with 3 units :)

So Vega could be also the first Consumer 4K Game 4K 60 FPS Recording :D

[nevcairiel]

That VP9 decode image makes no sense o.o

So Vega could be also the first Consumer 4K Game 4K 60 FPS Recording :D

ShadowPlay can record 4K at 60 FPS, if thats what you are trying to say. That line is rather confused.

[Trevonn]

Finally a free game recorder with hardware HEVC encode support. Your turn NVIDIA

[Trevonn]

That VP9 decode image makes no sense o.o



ShadowPlay can record 4K at 60 FPS, if thats what you are trying to say. That line is rather confused.

I think he means in HEVC

[CruNcher]

GameDVR
Shadowplay

and now ReLive

Massively hurt 3rd Parties as for example products like Action or Bandicam.

We will surely see a lot of these Products dieing in the Future if they don't buildup exciting features not presented by the OS and IHV solutions.

I think he means in HEVC

Nvidia could introduce it in no time they didn't yet which is quiet strange ;)

There would be 0 issues in their Ecosystem doing so up to Shield TV streaming no problem.

And the new Encoder core would even allow for High Quality Lossless HEVC for acquisition

Like you can already do H.264 Lossless for that purpose with Maxwell :)


Which both AMD doesn't support yet.

[Trevonn]

The only decent 3rd party option is OBS I don't think any of other 3rd party options will be missed especially the paid for ones.

The only problems with shadowplay is the lack of granularity in replay/record (Let me input a number!), variable frame rate encoding and the default colour space.

[CruNcher]

variable frame rate recording is no problem it's a feature that AMD neither supports ;)

It will be interesting to see if they push the new AMF with HEVC Encoder support right into GIT on the 8th December so if they sheduled it right on or if they still lose time on mass implementation ;)

@Trevonn

Why should the default colorspace be a problem with the target of Shadowplay being strictly Gamers as a Product and the playout happening to services which only support it :D

[Trevonn]

Regarding CFR, NVIDIA supports it in OBS and it helps with editing in premiere which doesn't like VFR video

I looked into the colour space issue a little more turns out it's not an issue haha nvm.

[aufkrawall]

Like you can already do H.264 Lossless for that purpose with Maxwell :)

Well, I messed around with all available encoder options in OBS with my 1070, and everything except of lossy 4:2:0 H.264 is just broken garbage.

[huhn]

decoding nvidia lossless h264 could be a simple decoding problem.

[aufkrawall]

Does it even support decoding H.264 lossless?

[sneaker_ger]

What is "it"?

[aufkrawall]

DXVA2, I guess.
I haven't payed attention if it was using hardware decoding.

[sneaker_ger]

If it was using software decoding: IIRC LAV H.264 lossless is hacky to account for broken x264 versions. If you upload a sample I can check.

[nevcairiel]

Does it even support decoding H.264 lossless?

NVIDIA hardware can support 4:2:0 8-bit lossless decoding on some GPUs, but its not exposed through any dxva2 decoder I know of.
4:4:4 or 4:2:2 is not supported through any hardware.

The 4:4:4 lossless decoding difference sneaker_ger talks of would result in mild image corruption only, and only with one particular type of coding tool in use, so its not guaranteed to be an issue with every lossless stream.

[sneaker_ger]

Yes, it is unlikely to be used. But also not difficult to check.

[huhn]

i guess the problem is more from the encode. nvidia hardware support the encoding of 4:4:4 lossless and if this stream results in garbage than one can assume nvidia is doing there own thing or OBS can't handle it properly.

[CruNcher]

i guess the problem is more from the encode. nvidia hardware support the encoding of 4:4:4 lossless and if this stream results in garbage than one can assume nvidia is doing there own thing or OBS can't handle it properly.

Im pretty sure BtbN never took High 4:4:4 Predictive Lossless into account really but the Profile works without issues exposed through the SDK and also Decoding works through CUVID at least, also decoding x264 High 4:4:4 Predictive Lossless works without problems.

But it's only really interesting comparing it for archival purposes and they're better solutions like FFV1 or Lagarith.

Though it will be interesting if in the future we can replace it with a VPx or AVx Lossless based option for archival and standardize those :)

[NikosD]

By far better drivers from AMD this year compared to Nvidia which had a lot, lot of problems.

Crimson Relive are extraordinary good, as I read to technical sites/ reviews.

This afternoon I will get a Sapphire Nitro+ RX 470/8GB and test the Polaris HW decoder in the next few days, even in VP9 (hybrid) after Crimson's Relive activation.

Stay tuned ;)

[aufkrawall]

By far better drivers from AMD this year compared to Nvidia which had a lot, lot of problems.

I think I can judge this better than you (owned a R9 390 for almost half of this year and else I've been with Nvidia) and I kinda disagree.

[NikosD]

There were numerous problems with Nvidia drivers and some of them were unusually "brutal" bugs for Nvidia.

For the majority of the users out there, AMD made a far better impression than Nvidia this year regarding drivers.

[huhn]

i running nvidia and AMD system over the past 5 years.

i have a "brutal" bug with AMD for 4 month plus now.

i have to disagree on the driver side too... this was a messy year for AMD driver...

[NikosD]

Nvidia was famous, once upon a time, for its drivers.

Actually Nvidia is a software company more than a HW company.

As the whole Internet says all over the planet, besides people here in this thread :) , this year was one bad or probably the worst year ever for Nvidia regarding drivers.

But, what the Internet and the technical sites know when we have so wise and skilled people around this thread...

[huhn]

nvidia had more flaws than usual that doesn't mean AMD was doing great at all.

there wasn't a major new driver bug that directly affected me using nvidia this whole year not a single one.

but i can instantly recall multiply problems and times where i had to skip AMD drivers.

openGL games like planetary annihilation where broken for month the bug was never added to the known issues.
madVR D3D9 windowed fullscreen is now broken for months on win 10.
and don't let me start about NGU polaris "performance" my R9 270 is most likely out performing my RX 480...
the removal of d3d9 openCL interop from the driver... great times!

and it is nice to read some where on the internet without personal experience.

[NikosD]

The opposite is quite true actually.

Most of the Nvidia users just skipped drivers or were waiting for the next one to fix bugs of the previous that was meant to fix bugs over a previous one.

It was a nasty situation and I'm not talking about video exclusively.

[huhn]

well them tell me your personal experience. it is not hard to find people with driver problems on the internet and there are currently way more nvidia user out there than AMD user so there are more people with problems.

and yes i know there where broken driver that got usually replaced in a couple of days.

i'm not here to tell "everything is all right in the green house" because it clearly isn't.

i mean for how many more years needs the CUVID decoder to die when any openCL program is used at the same time?

the difference is i don't have to skip one driver i have to wait months for a FIX from AMD or no FIX at all and a work around is created.

13 something is most likely still the fastest openCL AMD driver... it is just 3 years old...

[NikosD]

You have a serious issue concentrating on the specific subject we are talking about - 2016 drivers - like all the trolls actually.

So...

[Asmodian]

The opposite is quite true actually.

Most of the Nvidia users just skipped drivers or were waiting for the next one to fix bugs of the previous that was meant to fix bugs over a previous one.

It was a nasty situation and I'm not talking about video exclusively.

I didn't skip any Nvidia drivers, there were a few versions that had minor flickering on the desktop while at 144 Hz but it wasn't a problem at 120Hz and I don't remember any other issues. This flickering has been solved now. :)

I don't have any AMD cards to compare too but calling it a "nasty situation" is exaggerating it a lot, at least in my experiance.

But NGU is oddly slow on my Titan X (pascal) too, being slightly slower than NNEDI3 128. Maybe it is that NNEDI3 is oddly fast on this GPU but I doubt it. :p

[NikosD]

... on my Titan X (pascal) too...



Oh my God!

You threw 1500€ to buy Titan Pascal ?

For how much performance over 1080 ?

Do you also burn money to make heat in the winter ?

AMD needs desperately to release the top version of VEGA and Zen 8c/16t, there is no other to make real profit.

[huhn]

it not hard to find a broken NVIDIA driver: http://www.ghacks.net/2016/11/17/nvidia-375-86-driver-issues/

i mean nikosd tested them all! he must know! i mean this argument is "the internet said so". without links of cause. people that are using hardware in question are simply trolls no question.