What is current status for hardware H.265 encoding.

[TEB]

Anyone know if the Intel HW encoder or the Nvidia HW encoders support HEVC with Constant Quality for realtime encoding?

[cakuhnen]

Anyone know if the Intel HW encoder or the Nvidia HW encoders support HEVC with Constant Quality for realtime encoding?

Encoding with Intel Media SDK 2018 R1, encoding with Intel Media SDK GPU Accelerated plugin i get 19 fps for 1080p vídeo and the quality are great

[RanmaCanada]

Anyone know if the Intel HW encoder or the Nvidia HW encoders support HEVC with Constant Quality for realtime encoding?

That is all they support, as CRF is not in their capabilities, AFAIK.

[JohnLai]

K, everyone. Slightly unrelated....but here nvidia stuff for reading.

http://on-demand.gputechconf.com/gtc/2018/presentation/s8601-nvidia-gpu-video-technologies.pdf

[ShogoXT]

I have been treating this thread as an all in one encoding hardware implementations thread anyway.

I read that PDF,a couple of things. It's nice that they still are expanding features, but most of it seems to be about cuda accellerated encoding in the later half. Wasn't opencl and cuda acceleration shown to be worse for quality? Also those YouTube pictures are funny is it trying to demonstrate blocking artifacts.

https://blog.parsecgaming.com/nvidia-nvenc-outperforms-amd-vce-on-h-264-encoding-latency-in-parsec-co-op-sessions-713b9e1e048a

More evidence that vce is very far behind. On Reddit I try to help people with encoding settings on OBS and most people don't believe me that nvenc is decent and vce is not so good compared to say x264 very fast.

https://en.wikipedia.org/wiki/Video_Core_Next

Has anyone seen this? Apparently it was put into the Raven Ridge APU.

[JohnLai]

I have been treating this thread as an all in one encoding hardware implementations thread anyway.

I read that PDF,a couple of things. It's nice that they still are expanding features, but most of it seems to be about cuda accellerated encoding in the later half. Wasn't opencl and cuda acceleration shown to be worse for quality? Also those YouTube pictures are funny is it trying to demonstrate blocking artifacts.

https://blog.parsecgaming.com/nvidia-nvenc-outperforms-amd-vce-on-h-264-encoding-latency-in-parsec-co-op-sessions-713b9e1e048a

More evidence that vce is very far behind. On Reddit I try to help people with encoding settings on OBS and most people don't believe me that nvenc is decent and vce is not so good compared to say x264 very fast.

https://en.wikipedia.org/wiki/Video_Core_Next

Has anyone seen this? Apparently it was put into the Raven Ridge APU.

So far, nvidia adaptive GOP, adaptive IPB frame placement, lookahead and AQ (spatial and temporal) which technically are based on CUDA seem to work well. It depends on the implementation I guess.

Intel also has some sort of adaptive GOP, lookahead and IPB placements.

The greatest feature added by Nvidia and Intel is "constant quality" rate control mode (vbr-quality = nvidia , ICQ = intel)

No idea what is AMD doing with its VCE.

The youtube example is funny indeed, but the second picture sets with a picture of a man = Elecard HEVCAnalyzer? Before it was discontinued and acquired by adobe? Hmm....

[ShogoXT]

Looks like the trend of hardware accelerated encoding is not stopping. Adobe Premiere has added Quicksync encoding acceleration, not the asic chip, but using the IGPU for processing for speeding up encoding times, since it is otherwise completely unused on higher end rigs. If I still had a Intel system I would myself keep the IGPU enabled on Windows 10 since having 2 gpus enabled is no longer an issue, for more options.

This had fueled fanboy wars of course in the AMD vs Intel vs Nvidia world:

https://www.gamersnexus.net/guides/3310-adobe-premiere-benchmarks-rendering-8700k-gpu-vs-ryzen

I was going to start commenting about it in the other forums, but I thought x264 and cpu encoding in general got rid of OpenCL and CUDA acceleration because it is only usable for lookahead and usually makes the quality worse anyway.

Both Intel and Nvidia just in the last few months have put forth these features again. Did they fix the downsides? Does Adobe just say "they wont notice the quality difference anyway"?

Id really like to know if this is the future again or a temporary fad. Or is it mainly for newer encoders like VP9?

Also AMD needs to catch up, this was literally the whole reason they came up with the HSA platform idea, so what are they doing?

EDIT: Id like to bring up these discussions on reddit and am hoping to learn a few things here first.

EDIT2: https://forums.adobe.com/thread/2473774
Looks like it is for h264.

[benwaggoner]

It's nice that they still are expanding features, but most of it seems to be about cuda accellerated encoding in the later half. Wasn't opencl and cuda acceleration shown to be worse for quality?
There's nothing intrinsically worse about CUDA and OpenCL encoding. The problem has been the latency between GPU and CPU got in the way of lots of fast-feedback loops, which are increasingly important as the number of tools a codec can use expands, so more things need to be tried quickly in parallel with most options quickly terminated. CABAC is historically hard to multithread (although HEVC WPP makes it much more feasible; one CABAC thread per 64 pixel vertical row), so peak single-thread performance was/is a key limiting factor (CABAC can take up a good fraction of encode and decoding total MIPS).

GPU gives great SIMD support with hundreds of parallel cores. But a modern Intel CPU has much better per-core SIMD performance with the AVX family, and are getting ever more cores per CPU, all tied together with fast caches in unified memory

CUDA and OpenCL work great for more waterfall-like processes, where the GPU doesn't need to constantly report back to the CPU and get modified instructions.

FPGA looks like it has some good potential in video encoding.

The only grail for CPU/GPU is to have a fast CPU AND a fast GPU on the same die, sharing the same shared memory and shared caches. Intel has fast CPUs but not powerful OpenCL integrated GPUs; just using AVX-2/512 is generally faster for the compute part (although using GPU for decode, preprocessing, and lookahead can help).

It seems like AMD could do something amazing here, although they do multiple dies in the same package; the Zen microarchitecture doesn't have shared memory, or even have the HBM from prior AMD processors. Not a big issue for gaming, but it is for encoding and some other kinds of GPU compute tasks.

[foxyshadis]

It's kind of funny that the ideal video encoder would look something like the PS3's Cell; a beefy CPU, a bunch of lightweight SPEs directly connected for specialized workloads, and a GPU to farm out the most repetitive tasks to. Internally, a lot of CPUs do seem kind of like that these days, it's just not exposed to the programmer as explicitly and thus harder to take advantage of. The PS3 was just a little too ahead of its time.

[NikosD]

@benwaggoner

Great post, your last one.

[ShogoXT]

https://www.nvidia.com/content/dam/en-zz/Solutions/design-visualization/technologies/turing-architecture/NVIDIA-Turing-Architecture-Whitepaper.pdf

Bitrate efficiency and quality has gone up! This is a big deal see page 29.

[ReinerSchweinlin]

That changed last week, Intel released their full HEVC GPU module in the Windows 2018 R1 community release. They haven't yet done it in the Linux version for some reason.

So the community edition now includes the advanced HEVC-Encoder, not only the QSVENC SDK which is available in handbrake and others?

[zub35]

Test new NVENC-HEVC encoder (+ B frames) RTX2070

x264, x265, nvenc, qsv

test 1 https://rigaya34589.blog.fc2.com/blog-entry-1069.html
test 2 https://rigaya34589.blog.fc2.com/blog-entry-1070.html

New HEVC encoder - very good for game streams (twitch etc.)
Of course, this will not replace x264 (medium+) on a second computer, but will improve the quality of broadcasts, especially for low bitrates.

p.s. I not the author of the test's

[ReinerSchweinlin]

Thank you, very interesting.

[hajj_3]

Nvidia codec SDK 9.0 has just been released, it has some nice improvements for their new turing architecture: https://developer.nvidia.com/nvidia-video-codec-sdk

[videoh]

Your link says "coming soon" and only 8.2 is linked. Did I miss something?

[Yups]

Next year could be interesting.


Other improvements include a new HEVC Quick Sync Video engine that provides up to a 30% bitrate reduction over Gen9 (at the same or better visual quality)

For the media block, Intel says that the Gen11 design includes a ground up HEVC encoder design, with high quality encode and decode support.

https://www.tomshardware.com/reviews/intel-sunny-cove-gen11-xe-gpu-foveros,5932-3.html
https://www.anandtech.com/show/13699/intel-architecture-day-2018-core-future-hybrid-x86/3

[ReinerSchweinlin]

thank you for the interesting news :)
I assume this will require new hardware then? If Intel puts these engines in the lower end CPUs (Like G4560 Pentiums Kaby lake), building a cheap and fast encoding machine will be possible :)

[Yups]

Yes it requires a new hardware which includes the new encoder, this is a Gen11 presentation, means Icelake or Lakefield.

[NikosD]

@JohnLai
@CruNcher

Anyone else here with a Turing card ?

I used NVEncC v4.38 today just for a few H.265 encodings and I'm really impressed by the speed and quality.

Any particular settings for H.265 on Turing cards ?

[NikosD]

CUVID can work in two modes, either direct hardware access or through DXVA. In DXVA mode, CUVID can access the hybrid DXVA decoders, but that mode doesn't work on Windows 10, only the pure CUVID mode works on 10, which only has fixed-function decoders.

It's been more than 2 years (!) since this post.

Does CUVID have support in both modes on Win 10 nowadays ?

[nevcairiel]

No. But they also haven't made a hybrid decoder in a long time.

[NikosD]

Actually they don't need to.

nVidia accelerates everything in HW (MPEG1, MPEG2, MPEG4 ASP, H.264, H.265 (8/10/12 bit), VC-1, VP8, VP9 (8/10/12 bit)

We could expect a hybrid AV1 decoder sometime, I guess.

[NikosD]

So, I did a small benchmark regarding speed, not quality, using this source:
ftp://helpedia.com/pub/multimedia/x264/testvideos/2011%20-%2002%20-%20H.264%20CPU%20DXVA%20codec%20comparison%20-%20Core2Duo%20vs%20UVD%202.2/6.Cat-1080p60fpsRef4-25Mbps.m2ts

It's an H.264 1080p60fps file with an average of 25Mbps bitrate using two platforms:

1) Skylake Core i5 6500 under Win 10 x64 v17763 using 6576 drivers (API v1.28) and the HD 530 iGPU

Using QSVEncC v3.20 - GPU Clock/ Video Clock -> 1050 MHz

Encoding speeds:

H.265

ICQ/ Best -> 33,25 fps

ICQ/ Balanced -> 66,76 fps

ICQ/ Fastest -> 211,59 fps


H.264

ICQ/ Best -> 79,35 fps

ICQ/ Balanced -> 165,60 fps

ICQ/ Fastest -> 190,45 fps


2) nVidia GTX 1660 - Win 10 x64 v17763 - Drivers 430.39 (NVENC API v9.0 - CUDA 10.1)

Using NVEncC v4.38 - GPU Clock -> 1965 MHz - Video Clock -> 1815 MHz

Encoding speeds:

H.265

VBRHQ/ Quality -> 165,14 fps

VBRHQ/ Default -> 298,40 fps

VBRHQ/ Performance -> 441,89 fps


H.264

VBRHQ/ Quality -> 241,52 fps

VBRHQ/ Default -> 413,76 fps

VBRHQ/ Performance -> 667,61 fps

Using other encoding modes (VBR/ CBR/ CBRHQ/ CQP) the encoding speed didn't change.

As you can see the performance difference is huge, because Turing encoder is really fast.

I'm not sure if HD 630 is a lot faster than HD 530 and if Pascal encoder is even faster than Turing due to more encoding engines (2 vs 1)

[Forteen88]

@NikosD. Could you please also do a small test on metric quality (SSIM, and maybe VMAF on a proper video-source), if possible, on the Turing-GPU encoder vs x265@slower (with the same file-size)?
I saw that there are just a few Nvidia-cards that supports Turing HEVC-encoding with B-frames,
https://developer.nvidia.com/video-encode-decode-gpu-support-matrix

[NikosD]

@Forteen88
Sorry, but I don't have/use tools and apps regarding quality metrics and benchmarks.

I'm here mainly for speed benchmarks.

[ReinerSchweinlin]

Encoding with Intel Media SDK 2018 R1, encoding with Intel Media SDK GPU Accelerated plugin i get 19 fps for 1080p vídeo and the quality are great
I just ran across this ... Do you have any links with a "how to" ? I´d like to try out the GPU Accelerated HEVC Encoder in the free edition.

[RanmaCanada]

So, I did a small benchmark regarding speed, not quality, using this source:
ftp://helpedia.com/pub/multimedia/x264/testvideos/2011%20-%2002%20-%20H.264%20CPU%20DXVA%20codec%20comparison%20-%20Core2Duo%20vs%20UVD%202.2/6.Cat-1080p60fpsRef4-25Mbps.m2ts

It's an H.264 1080p60fps file with an average of 25Mbps bitrate using two platforms:

1) Skylake Core i5 6500 under Win 10 x64 v17763 using 6576 drivers (API v1.28) and the HD 530 iGPU

Using QSVEncC v3.20 - GPU Clock/ Video Clock -> 1050 MHz

Encoding speeds:

H.265

ICQ/ Best -> 33,25 fps

ICQ/ Balanced -> 66,76 fps

ICQ/ Fastest -> 211,59 fps


H.264

ICQ/ Best -> 79,35 fps

ICQ/ Balanced -> 165,60 fps

ICQ/ Fastest -> 190,45 fps


2) nVidia GTX 1660 - Win 10 x64 v17763 - Drivers 430.39 (NVENC API v9.0 - CUDA 10.1)

Using NVEncC v4.38 - GPU Clock -> 1965 MHz - Video Clock -> 1815 MHz

Encoding speeds:

H.265

VBRHQ/ Quality -> 165,14 fps

VBRHQ/ Default -> 298,40 fps

VBRHQ/ Performance -> 441,89 fps


H.264

VBRHQ/ Quality -> 241,52 fps

VBRHQ/ Default -> 413,76 fps

VBRHQ/ Performance -> 667,61 fps

Using other encoding modes (VBR/ CBR/ CBRHQ/ CQP) the encoding speed didn't change.

As you can see the performance difference is huge, because Turing encoder is really fast.

I'm not sure if HD 630 is a lot faster than HD 530 and if Pascal encoder is even faster than Turing due to more encoding engines (2 vs 1)

Turing should actually be faster than Pascal and better quality as Pascal can only use 1 encoding engine at a time. What it allows for is you to have 2 encodes going, using 1 engine each (consumer and low end Quadro cards are limited to 2 simultaneous encodes at a time though there are fixes for that) vs 1 engine doing 2 encodes, like on the cards with only 1 engine (anything below the 1070Ti.)

[tyee]

Just got a 1660ti and have started encoding 4k today. I use avisynth+ (latest version) for decoding and NVencC64.exe for encoding. I'm getting about 150fps for 1080p and only about 25fps for 4k. For 4k, my cpu is only 50% occupied. Any way to speed up everything??

[RanmaCanada]

Just got a 1660ti and have started encoding 4k today. I use avisynth+ (latest version) for decoding and NVencC64.exe for encoding. I'm getting about 150fps for 1080p and only about 25fps for 4k. For 4k, my cpu is only 50% occupied. Any way to speed up everything??
Stop piping through avisynth as it's actually slowing down your encode. It can't feed the asic fast enough. Also, you're encodes are going to be garbage compared to just using your processor. Hardware encoding still hasn't caught up to software, at least with HEVC.

[sneaker_ger]

I got myself a GTX 1660 Ti. Just a short test in case anyone is interested. NVEncC64 -c hevc --output-depth 10 -b 5 --bref-mode middle --aq --aq-temporal --vbrhq 0 --vbr-quality 30 -u quality vs. x264 --preset veryslow --tune film 2pass at ~2200 kbps.

Input and output files:
https://mega.nz/#F!BlNDGIgJ!7lBUBs61l1oiIIBScrMyYg
(Source is some part of Netflix' "Meridian" encoded to 1080p lossless H.264.)

Screens (s=source, nv=1660 Ti, x=x264):
https://abload.de/img/s014jj13.png
https://abload.de/img/nv01w7j73.png
https://abload.de/img/x01s0kjv.png

https://abload.de/img/s02cxjdu.png
https://abload.de/img/nv02ctkup.png
https://abload.de/img/x025nj5e.png

https://abload.de/img/s03eqj3a.png
https://abload.de/img/nv0383jk4.png
https://abload.de/img/x03hhj8k.png

https://abload.de/img/s04dcj3m.png
https://abload.de/img/nv04pfjaq.png
https://abload.de/img/x046njag.png

https://abload.de/img/s05itkot.png
https://abload.de/img/nv05b9j42.png
https://abload.de/img/x05lnkfh.png

[Sharc]

I got myself a GTX 1660 Ti. Just a short test in case anyone is interested. NVEncC64 -c hevc --output-depth 10 -b 5 --bref-mode middle --aq --aq-temporal --vbrhq 0 --vbr-quality 30 -u quality vs. x264 --preset veryslow --tune film 2pass at ~2200 kbps.

x264 is still superior w.r.t. reproduction of details, but NV (NVEncC) has improved significantly. My choice except for "archiving" purpose.

[NikosD]

I got myself a GTX 1660 Ti.
Just a short test in case anyone is interested.

NVEncC64 -c hevc --output-depth 10 -b 5 --bref-mode middle --aq --aq-temporal --vbrhq 0 --vbr-quality 30 -u quality
vs.
x264 --preset veryslow --tune film 2pass at ~2200 kbps.
Nice...Did you keep notes regarding encoding speed (fps) and what is the CPU ?

[sneaker_ger]

CPU is an old i5-2500K. It cannot even decode the source sample in real-time. I didn't keep notes but just did a short test just for encoding speed (with sample recoded to H.264 and decoded by HW to focus on pure encoding speed). x264 CRF is in the ball-park of 7 fps. NVENC HEVC 1080p maxed out everything about 160 fps. With HEVC default settings about 330 fps. H.264 default setting between 500 and 600 fps, may need a longer sample to measure reliably. Not sure if decoding or encoding is the bottleneck.

[Sharc]

Not sure if decoding or encoding is the bottleneck.
.. or data transfer rate from/to storage device (HDD, SSD, USB …)?

[sneaker_ger]

For the speed test I encoded a lossy sample, H.264, 10 MB, SSD.

[Sharc]

Source is some part of Netflix' Meridian encoded to 1080p lossless H.264. I can upload later if anyone is actually interested but be warned cause it's 2.7 GB.

May I ask you to upload the source? I am interested in doing some AVC 8-bit tests with my Pascal GTX1050Ti. I think that Pascal/NVEncC have also improved a lot recently.
Thanks.

[sneaker_ger]

https://mega.nz/#F!BlNDGIgJ!7lBUBs61l1oiIIBScrMyYg (full source is available at xiph.org (https://media.xiph.org/video/derf/meridian/MERIDIAN_SHR_C_EN-XX_US-NR_51_LTRT_UHD_20160909_OV/))

[Sharc]

Thank you.

[pacuro]

Where is JohnLai? Did I miss something? Hope for his update with best hevc nvenc settings for Turing. Is it enough just to enable B-frames with his last best settings?

[ReinerSchweinlin]

I gave myself a christmas present and got a RTX 2060. First tests with NVENC in different GUIs give me good enough quality for my 1080p recodes when speed is important (around 140fps).
So far, I only used default settings. Like pacuro, I was wondering what settings I could use to max out quality (enabling b-frames gave me a bigger filesize than defaults in hybrid). Is there some detailied documentation of the NVENC in RTX Cards I could look into?

[tyee]

I'm using this command line app and it has very nice quality and I also get around 150fps for 1080p, about 50fps for 4k

https://github.com/rigaya/NVEnc

Is John still around the forums? I saw his settings somewhere, he was using Staxrip.

[pacuro]

@ReinerSchweinlin you probably doing something wrong if enabling B-frames makes your encodes bigger. In my case just enabling b-frames + b as ref each makes files lighter - as it should be in theory.

https://bpccdn.fra1.digitaloceanspaces.com/original/3X/9/7/972e44a699e4bd0cec8c883a2a98eea8f4b7c608.png

[craigpro]

These are the settings I'm using and my 720p encodes are always a few hundred MB larger than similar shows from the scene HEVC encodes. Does anyone have any suggestions for better settings please? Thank you.

--cqp 18:20:22 --codec h265 --preset quality --level 5.1 --output-depth 10 --qp-init 20 --qp-max 22 --qp-min 18 --max-bitrate 5000 --aq --aq-temporal --gop-len 240 --lookahead 16 --slices 2 --multiref-l0 4 --multiref-l1 4 --strict-gop --nonrefp --weightp

[benwaggoner]

These are the settings I'm using and my 720p encodes are always a few hundred MB larger than similar shows from the scene HEVC encodes. Does anyone have any suggestions for better settings please? Thank you.

--cqp 18:20:22 --codec h265 --preset quality --level 5.1 --output-depth 10 --qp-init 20 --qp-max 22 --qp-min 18 --max-bitrate 5000 --aq --aq-temporal --gop-len 240 --lookahead 16 --slices 2 --multiref-l0 4 --multiref-l1 4 --strict-gop --nonrefp --weightp
Having a QP range of just 18-22 seems like a really narrow band if your content is ever VBV constrained. 4 QP isn't enough for even a 2x difference in bitrate.

[RanmaCanada]

These are the settings I'm using and my 720p encodes are always a few hundred MB larger than similar shows from the scene HEVC encodes. Does anyone have any suggestions for better settings please? Thank you.

--cqp 18:20:22 --codec h265 --preset quality --level 5.1 --output-depth 10 --qp-init 20 --qp-max 22 --qp-min 18 --max-bitrate 5000 --aq --aq-temporal --gop-len 240 --lookahead 16 --slices 2 --multiref-l0 4 --multiref-l1 4 --strict-gop --nonrefp --weightp

With current tech you will not get close to what software encoding can do. If you want better results, use a CPU. I know it's not what you want to hear, but it's the truth. GPU is quick and dirty.

[benwaggoner]

With current tech you will not get close to what software encoding can do. If you want better results, use a CPU. I know it's not what you want to hear, but it's the truth. GPU is quick and dirty.
Another way to frame that is that the quality @ perf for GPU is higher at very high speeds, but when encoding time/resources can be more, SW pulls ahead in quality as GPU encoding's peak quality is low.

GPU encoding is great for things like Twitch where it is recording screen activity while gaming. It takes minimal CPU away from the game, and is able to use the frames in the GPU as source without having to copy the pixels to main memory. Just writing 4Kp60 RGB frames to the CPU on top of everything else further stresse the memory bandwidth.

It's also good for very low power embedded solutions where a Tegra or Atom processor can be used with little CPU but with a HW encoder available.

There's been work done on hybrid encoders, which use the GPU for the first pass to be refined in software, which can speed up encoding 20-25%.

[Yups]

I'm currently testing Quicksync on Iris Xe LP. Compared to Gen9.5 they have greatly improved the fixed function encoder, both quality and speed. On Gen9.5 graphics the fully fixed function encoder was more or less useless (no b-pyramid, no bframes), means the hybrid was the better choice and for H265 Gen9.5 didn't even support FF. On Iris Xe LP I don't really see a quality difference between hybrid and the fully fixed function (H265) encode, the difference is minor I would say but the speed difference is really huge. So at this point it makes the h265 hybrid somehow obsolete.

[RanmaCanada]

I'm currently testing Quicksync on Iris Xe LP. Compared to Gen9.5 they have greatly improved the fixed function encoder, both quality and speed. On Gen9.5 graphics the fully fixed function encoder was more or less useless (no b-pyramid, no bframes), means the hybrid was the better choice and for H265 Gen9.5 didn't even support FF. On Iris Xe LP I don't really see a quality difference between hybrid and the fully fixed function (H265) encode, the difference is minor I would say but the speed difference is really huge. So at this point it makes the h265 hybrid somehow obsolete.

Are you using the Intel Media SDK? That is what is recommended to get the best encodes out of Quicksync. At least last time I checked.

[Yups]

Are you using the Intel Media SDK? That is what is recommended to get the best encodes out of Quicksync. At least last time I checked.


As for the hardware encoder it doesn't matter, it just needs the graphics driver, it has all the required files. The Media SDK is required for the software encoder.