What is current status for hardware H.265 encoding.

[Yups]

I finished my Tigerlake speed test and also some other results just for reference. QSVEnc/NVEnc (h265) used for the GPUs and Handbrake nightly for the x265 run. Test sample is a 1080p 24fps 25 Mbit video which is converted down to ~2500 Kbit.



Iris Xe LP 1100 Mhz GPU+fixed function
CBR slowest= 68 fps
CBR balanced= 131 fps
CBR fastest= 226 fps

Iris Xe LP 1100 Mhz fixed function
CBR slowest= 260 FPS
CBR balanced= 325 FPS
CBR fastest= 507 FPS


Iris Xe LP 1100 Mhz GPU+ fixed function
CQP slowest= 75 fps
CQP balanced= 144 fps
CQP fastest= 263 fps


Iris Xe LP 1100 Mhz fixed function
CQP slowest= 271 fps
CQP balanced= 344 fps
CQP fastest= 518 fps


For reference some other GPUs and x265 ultrafast:

HD 630 1150 Mhz GPU + fixed function
CBR slowest= 43 fps
CBR balanced= 100 fps
CBR fastest= 302 fps

i7-1165G7 2800 Mhz (Turbo disabled)
x265 ultrafast= 58 fps

GTX 1080 2000 Mhz
CBR quality= 317 fps
CBR fastest= 502 fps

[RanmaCanada]

But what was the quality comparison to NVENC? Speeds mean nothing if the output is garbage (Looking at you AMD VCE) :) If you have not done a comparison, would it be possible to do one?

[Yups]

In this speed test NVENC quality is worse than Iris Xe but I will use different settings when I go for quality on both, for example I didn't use lookahead on NVENC and keep in mind this isn't Turing. I will do a quality comparison later and also I have to test Turing which I don't have for now. I'm testing on Iris Xe mainly at the moment. As for Iris Xe CQP is the best bitrate control method when it comes to quality, better than ICQ and much better than CBR/VBR. CQP works very good with many bframes+bpyramid on Iris Xe unlike ICQ.

[RanmaCanada]

In this speed test NVENC quality is worse than Iris Xe but I will use different settings when I go for quality on both, for example I didn't use lookahead on NVENC and keep in mind this isn't Turing. I will do a quality comparison later and also I have to test Turing which I don't have for now. I'm testing on Iris Xe mainly at the moment. As for Iris Xe CQP is the best bitrate control method when it comes to quality, better than ICQ and much better than CBR/VBR. CQP works very good with many bframes+bpyramid on Iris Xe unlike ICQ.
Thanks!

[benwaggoner]

In this speed test NVENC quality is worse than Iris Xe but I will use different settings when I go for quality on both, for example I didn't use lookahead on NVENC and keep in mind this isn't Turing. I will do a quality comparison later and also I have to test Turing which I don't have for now. I'm testing on Iris Xe mainly at the moment. As for Iris Xe CQP is the best bitrate control method when it comes to quality, better than ICQ and much better than CBR/VBR. CQP works very good with many bframes+bpyramid on Iris Xe unlike ICQ.
The Turing NVENC added B-frames, IIRC, so is quite a lot better.

[utack]

I am honestly surprised how much praise nvenc gets "around the internet"
The H.264 encoder certainly isn't bad for a hardware encoder, but people comparing it to x264 medium or even slow need to tone it down a notch.

In a "non realtime" scenario with aq, slowest preset and full bframes it did not seem competitive to x264 medium preset in quality in some gaming footage I've tested

[Asmodian]

It is from comparing it to the previous hardware encoders, Turing was a big upgrade.

It is actually somewhat watchable now! Credit where it is due. ;)

[aegisofrime]

Does Ampere have any upgrades over Turing? I have a 3080 and 3060Ti coming, so I'm curious about this.

[foxyshadis]

Does Ampere have any upgrades over Turing? I have a 3080 and 3060Ti coming, so I'm curious about this.

Would you be willing to give the HEVC challenge a shot once they arrive? https://forum.doom9.org/showthread.php?t=175776 At least then we have a thing we can compare to. Staxrip can handle the encoding to nvenc quite well.

[nevcairiel]

Does Ampere have any upgrades over Turing? I have a 3080 and 3060Ti coming, so I'm curious about this.

No, the encoder is the same. Only the decoder was enhanced to support AV1.

[benwaggoner]

I am honestly surprised how much praise nvenc gets "around the internet"
The H.264 encoder certainly isn't bad for a hardware encoder, but people comparing it to x264 medium or even slow need to tone it down a notch.

In a "non realtime" scenario with aq, slowest preset and full bframes it did not seem competitive to x264 medium preset in quality in some gaming footage I've tested
Yeah, HW encoders haven't ever really approached the "medium" speed of the best SW encoder, and the gap gets bigger as codecs get more complex. CPUs with lots of cores, lots of cache, and powerful SIMD are really good at trying lots of different ways to encode each block, with early exits.

We've seen high-end broadcast encoders move from ASIC-based encoding to SW and SW-defined encoding over the last decade or so for this reason.

The sweet spot for HW encoders these days seems to be parallel encoding of CPU-intensive stuff with minimal perf hit. Streaming games is the key scenario; a SW encoder taking "only" 30% of CPU when gaming could have a big negative impact on fps. Plus encoding frames already in GPU memory doesn't add to memory throughput pressure much.

[aegisofrime]

Would you be willing to give the HEVC challenge a shot once they arrive? https://forum.doom9.org/showthread.php?t=175776 At least then we have a thing we can compare to. Staxrip can handle the encoding to nvenc quite well.

Sure, can do, I have the 3060Ti already just waiting to install it for Cyberpunk :D Not to doubt you nevcairiel, but it's worth a shot eh?

[varnav]

Some benchmarks done by me recently. I'm getting surprisingly good results from nvenc and even Intel compared to xlib265. Any comments?

System configuration

CPU: Intel Core i7 10750H
GPU 1: Intel UHD Comet Lake GT2
GPU 2: Mobile nVidia GeForce RTX 2060 (Turing microarchitecture)

Transcoding:
Windows 10 20H2
ffmpeg 2020-12-01-git-ba6e2a2d05-essentials_build-www.gyan.dev
HEVC encoder version 3.4+27-g5163c32d7

VMAF:
Ubuntu 20.04 WSL2
ffmpeg N-100180-gdbf8a16
VMAF 1.5.3

Source: http://ftp.nluug.nl/pub/graphics/blender/demo/movies/ToS/ToS-4k-1920.mov

File size: 704 MB
h264 (High) (avc1 / 0x31637661), yuv420p, 1920x800 [SAR 1:1 DAR 12:5], 7862 kb/s, 24 fps, 24 tbr, 24 tbn, 48 tbc (default)

VMAF command line:

./build.ffmpeg.sh
ffmpeg -i result.mkv -i ToS-4k-1920.mov.mkv -lavfi libvmaf="model_path=/usr/local/share/model/vmaf_v0.6.1.pkl" -report -f null -

Software

ffmpeg -xerror -i .\ToS-4k-1920.mov -map_metadata 0 -movflags use_metadata_tags -vcodec libx265 -crf 20 -preset slow -acodec copy ToS_cpu.mov

encoded 17620 frames in 1695.38s (10.39 fps), 4568.08 kb/s, Avg QP:23.91

File size: 416 MB
VMAF score: 98.601471

nVidia

ffmpeg -xerror -vsync 0 -i .\ToS-4k-1920.mov -rc-lookahead 25 -map_metadata 0 -movflags use_metadata_tags -preset p6 -spatial-aq 1 -temporal_aq 1 -cq 26 -vcodec hevc_nvenc -acodec copy ToS_nvenc.mov

fps=220 q=19.0 Lsize=416202kB time=00:12:14.12 bitrate=4644.3kbits/s speed=9.15x

File size: 406 MB
VMAF score: 97.881867

Intel

ffmpeg -xerror -hwaccel auto -i .\ToS-4k-1920.mov -vcodec hevc_qsv -b:v 6M -acodec copy ToS_intel.mov

frame=17620 fps=82 q=-0.0 Lsize= 440303kB time=00:12:14.09 bitrate=4913.5kbits/s speed=3.43x

File size: 429 MB
VMAF score: 97.592410

[benwaggoner]

VMAF >95 isn't that sensitive to subjective quality differences. I wouldn't predict actual real-world quality from scores between 97.6 and 98.6. Especially for clips more than 20 seconds or so, a 98.6 can look worse than a 97.6 (although all should look quite good). Reencoding from a source only 50% higher bitrate is also an odd case since the source artifacts could easily be more prominent than artifacts added in reencoding.

Testing with all around 3 Mbps would probably give some more interesting quality comparisons. Although it all depends on the scenario in the end. There are absolutely use cases where a 50% higher bitrate for 10x faster encoding is a great tradeoff.

[varnav]

I agree transcoding from already compressed video is not a clean thing, I will grab raw source someday and try this with it.

For subjective part, here is set of screenshots. I mixed them for the blind test. Who can tell what's what? Several guys could not.

https://drive.google.com/open?id=1ZCV2Xb9p15uqP4OTYMBnUWRGf1Ah_XEM

My scenario is recompression of home collection of short (but still 1080p 60 fps) videos, I am making a tool for this:

https://github.com/varnav/filmcompress

[foxyshadis]

I agree transcoding from already compressed video is not a clean thing, I will grab raw source someday and try this with it.

For subjective part, here is set of screenshots. I mixed them for the blind test. Who can tell what's what? Several guys could not.

https://drive.google.com/open?id=1ZCV2Xb9p15uqP4OTYMBnUWRGf1Ah_XEM

My scenario is recompression of home collection of short (but still 1080p 60 fps) videos, I am making a tool for this:

https://github.com/varnav/filmcompress

We've pretty much settled on the idea that comparing screenshots is pointless, and only clips is valid, despite how much more difficult that is.

That said, in each triplet of yours, one is noticeably softer, but whether that's innate to the coder or an artifact of the options chosen is impossible to tell. The level of softness is well below what I'd consider a threshold for noticing in motion, though. Overall, 4.5Mb/s is probably just not nearly enough to stress any of these chips anymore.

[hajj_3]

VMAF 2.0 was released a few days ago btw: https://github.com/Netflix/vmaf/releases/tag/v2.0.0

[varnav]

I can cut clips if needed. I can run more tests too if someone wants.
But IMO NVENC in later hardware revisions of NV chips produces very good results. Software is still better, but speed difference is 10-20 times.

[Yups]

On Intel both HD630 and Iris Xe the constant bitrate modes seem to work a lot better with a low bitrate budget. For now only a PSN-Y test using this (https://drive.google.com/file/d/1YX1V0SeSkYaq6Ui41vv1wcOatbLnuLSL/view?usp=sharing) sample, all of them HEVC main 8 bit 1920x1080 at ~2.5 Mbit 24 fps/240 gop. varna, you might use this sample.


Iris Xe:
CBR 2500 Hybrid quality= 37.97 dB (60 FPS, 3 bframes+pyramid)
CBR 2500 FF/low power quality= 37.60 dB (265 fps, 5 bframes+pyramid)

ICQ 25 Hybrid quality= 39.52 dB (66 fps, 5 bframes +pyramid)
ICQ 25 Hybrid speed= 39.34 dB (213 fps, 5 bframes +pyramid)

CQP 22_23_26 Hybrid quality= 40.32 dB (67 fps, 16 bframes+pyramid, offset 1_6_8)
CQP 23_23_26 FF/low power quality= 40.06 dB (272 fps, 16 brames+pyramid, offset 2_6_8)

Best on HD630 (Gen9.5):

CQP 23_24_28 Hybrid quality= 38.98 dB (42 fps, 16 bframes+pyramid, offset 2_6_8)

Iris Xe CQP vs HD630 CQP

https://s10.directupload.net/images/201212/cs73djrc.png


The difference is really big, Gen9 looks much worse in the video.

[Yups]

Same video on my GTX 1080 (8 bit HEVC, 2.5 Mbit, 240 gop....)

CBR quality= 36.93 dB
CBR HQ quality= 37.37 dB (Lookahead 32)
VBR HQ quality= 38.05 dB (Lookahead 32)
CQP quality= 38.84 dB (CQP 22_33_36, Lookahead 32)

I've also tried other settings but the PSNR-Y went down or didn't improve.

[easyfab]

I got my new nuc 11 with i7-1165G7 CPU.

I did some tests only with 1 file for the moment ( Kimono1_1920x1080_24.y4m ) but it give me a 1st idea.

SW encoding :


enc.time s. size ko vmaf psnr ssim


x264 fast 4.25 3954 93.76 39.94 12.70
x264 veryslow 31.7 3618 90.47 39.65 12.54

x265 superfast 10b 5.58 3698 95.49 40.63 13.01
x265 superfast 10b 5.74 4042 96.20 40.86 13.16
x265 medium 10b 11.24 3202 93.91 40.29 12.85
x265 medium 10b 12.26 4232 96.16 41.03 13.35
Svt-hevc preset 8 10b 5.97 3970 95.81 40.61 13.17

rav1e preset 6 q100 8b 84.24 4020 97.01 41.34 13.68
rav1e preset 6 q110 10b 238.35 3493 96.25 41.05 13.45

aom cpu 6 crf 32 10b 69.61 3733 97.70 41.61 13.76

svt-av1 speed 7 q34 10b 29.08 4012 97.30 41.31 13.55
svt-av1 speed 8 q35 10b 27.03 3782 96.94 41.14 13.43


other HW encoding :

enc.time s. size ko vmaf psnr ssim


Nvencc(1050 gtx) quality 10b 2.65 3721 92.72 39.85 12.49

qsvencc (m3-7y30) best 10b 9.00 3826 93.28 40.07 12.69

veencc (4800u) slow 10b 4.00 3794 93.22 39.71 12.38

and new i7-1165G7 tiger lake :

enc.time s. size ko vmaf psnr ssim


Qsvencc h264 best 8bit 1.00 3658 93.59 39.45 12.32

Qsvencc hevc best 8bit 4.00 3746 96.97 40.93 13.21
Qsvencc hevc best 8bit ff 1.00 3894 96.69 40.76 13.11


Qsvencc hevc best 10bit 4.00 3719 96.85 40.91 13.17
Qsvencc hevc best 10bit ff 1.00 3845 96.50 40.75 13.07

QSV on Tiger is a great jump in quality vs my older HW and can compare to SW encoding (but 10x faster/ less than 40w).
I don't have newer nvidia to compare.

[Yups]

Did you use VBR/CBR for i7-1165G7? FF/Hybrid quality difference is rather low with a big speed advantage in favour of FF.

[easyfab]

only cqp e.g. :
-c hevc --cqp 36:38:39 --profile main10 --fixed-func -b 6 --ref 6

[Yups]

Ok thanks, can you test it with --qp-offset 2:6:8 --b 16

You have to lower your cpq values as well with this offset to get a similar output size.

[easyfab]

without :
size 3845 psnr 40.75 ssim 13.07

with --qp-offset 2:6:8 -b 16:
size 3806 psnr 40.74 ssim 13.19

And qp-offset seem to be in on by default but don't know with what values.

for info :

cop3.DirectBiasAdjustment value changed off -> auto by driver
cop3.GlobalMotionBiasAdjustment value changed off -> auto by driver
QSVEncC (x64) 4.13 (r2009) by rigaya, Feb 17 2021 13:49:19 (VC 1928/Win/avx2)
OS Windows 10 x64 (19042) [UTF-8]
CPU Info 11th Gen Intel Core i7-1165G7 @ 2.80GHz [TB: 4.69GHz] (4C/8T) <Tigerlake>
GPU Info Intel Iris(R) Xe Graphics (96EU) 100-1300MHz [28W] (27.20.100.9313)
Media SDK QuickSyncVideo (hardware encoder) FF, 1st GPU, API v1.34
Async Depth 6 frames
Buffer Memory d3d9, 1 input buffer, 46 work buffer
Input Info avqsv: H.264/AVC, 1920x1080, 24/1 fps
VPP Enabled ColorFmtConvertion: nv12 -> p010
AVSync cfr
Output HEVC main10 @ Level 5 (high tier)
1920x1080p 1:1 24.000fps (24/1fps)
avwriter: hevc => matroska
Target usage 4 - balanced
Encode Mode Constant QP (CQP)
CQP Value I:32 P:35 B:36
QP Limit min: 22, max: 63
Trellis Auto
Ref frames 6 frames
Bframes 6 frames, B-pyramid: on
Max GOP Length 240 frames
Ext. Features WeightP WeightB QPOffset tskip ctu:64 sao:all

[Yups]

It's on by default but they are worse. SSIM jumped from 13.07 to 13.19 with a smaller file size, clearly better I would say. You did use b frames 6 by the way unless this log is old. From my test in #369 Tigerlake CQP with this settings reached PSNR/SSIM results which are comparable to x265 medium/slow, CQP on Tigerlake is superb. And because Iris Xe is so new we may see driver improvements in the future.

[benwaggoner]

It's on by default but they are worse. SSIM jumped from 13.07 to 13.19 with a smaller file size, clearly better I would say. You did use b frames 6 by the way unless this log is old. From my test in #369 Tigerlake CQP with this settings reached PSNR/SSIM results which are comparable to x265 medium/slow, CQP on Tigerlake is superb. And because Iris Xe is so new we may see driver improvements in the future.
The subjective correlation of PSNR and SSIM aren't really good enough to confidently state that different encode parameters yielding a 0.12 dB difference is a material quality improvement. There are perceptual optimizations that can reduce PSNR and SSIM by >1 dB while improving subjective quality.

[Yups]

I have tested these settings a lot. They are better subjective and objective. These settings allow lower CQP values over the default ones. Of course there is a limit, a too big offset may cause problems like more noise depending on the content. And in this case the SSIM/PSNR may increase while subjective there are some downsides.

[Yups]

Another Tigerlake CQP test with yvid_SAM_0235-UHD_Sample1 (HEVC main 8 bit 3840x2160 80 Mbit 30 fps)



PSNR-Y SSIM-Y speed bitrate

Tigerlake CQP FF quality 29.79 0.875 67 fps 3317 kbit
Tigerlake CQP quality 29.94 0.878 35 fps 3344 kbit

x265 medium 29.40 0.871 5.4 fps 3335 kbit
x265 slow 29.92 0.876 2.4 fps 3330 kbit

x264 slow 26.80 0.811 9.5 fps 3480 kbit



I'm using these Tigerlake CQP settings (Staxrip+QSVEnc):

--codec hevc --quality best --profile main --bframes 16 --gop-len 240 --b-pyramid --open-gop --qp-offset 2:6:8 --d3d9 --fixed-func --cqp 35:38:40

[ukmark]

I don't have Tiger Lake, but it's predecessor Ice Lake i7-1065G7 and am very impressed with CQP fixed-function encoding. I have posted my findings on StaxRip forum here:-

http://forum.doom9.net/showthread.php?p=1936660#post1936660

It appears the main improvement with Tiger Lake is the ability to use b-frames and b-pyramid with fixed-function encoding. With Ice Lake these are only available with GPU EU encoding (which is MUCH slower).

Suffice it to say that I no longer use x265. With blu ray (or reasonably high bit-rate) sources, I see no advantages for x265. I really find it very difficult to see any quality differences (that's after re-encoding about 50 of my blu ray collection using HW QuickSync encoding.)

I also posted a few screen grabs comparing x265 with QSV CQP (at 720p resolution) here (scroll down the page to see):-
https://github.com/rigaya/QSVEnc/issues/46

[Yups]

With Ice Lake these are only available with GPU EU encoding (which is MUCH slower).



It is but you can try non FF TU7/fastest preset where you can use 16 bframes+bpyramid, it's fast and the quality is better than TU1/best FF no bframes+bpyramid according to my tests. Same video from #379:

TU7 16 bframes+pyramid PSNR-Y: 29.45---3337 kbit---116 fps
TU1 FF no bframes/pyramid PSNR-Y: 28.95---3432 kbit---81 fps

[RanmaCanada]

Thanks for these tests, as it is extremely difficult to find quicksync encode information on forums. Everything is usually NVENC or x265.

[ukmark]

It is but you can try non FF TU7/fastest preset where you can use 16 bframes+bpyramid, it's fast and the quality is better than TU1/best FF no bframes+bpyramid according to my tests. Same video from #379:

TU7 16 bframes+pyramid PSNR-Y: 29.45---3337 kbit---116 fps
TU1 FF no bframes/pyramid PSNR-Y: 28.95---3432 kbit---81 fps

Thanks - I'll give that a go!

[Nico8583]

Very interesting, I'm a SW user but I would like to try Intel HW. Is it possible with an Intel 8500 (UHD 630) ? Is the quality will be the same than newer CPU, just the speed slower ? Thank you.

[Tenkei]

I'd love to see some tests at higher bitrates, like 10 Mbps+.

[Yups]

Very interesting, I'm a SW user but I would like to try Intel HW. Is it possible with an Intel 8500 (UHD 630) ? Is the quality will be the same than newer CPU, just the speed slower ? Thank you.


No it won't be the same unfortunately, HD630 is Gen9.5 based which is old gen. Intel upgraded their HEVC encoder with Icelake (Gen11): https://forum.doom9.org/showpost.php?p=1860006&postcount=317

With Tigerlake there is another upgrade to the HEVC FF encoder (bframes+bpyramid support).

[Nico8583]

Thank you, I'll try it in the future if I buy a Gen11 Intel CPU

[Yups]

Thank you, I'll try it in the future if I buy a Gen11 Intel CPU


If you buy new, you should better buy something with Gen12/Iris Xe graphics.

[Nico8583]

If you buy new, you should better buy something with Gen12/Iris Xe graphics.
Thank you, I don't plan to buy a new setup now but thanks for the information ;)

[benwaggoner]

If you buy new, you should better buy something with Gen12/Iris Xe graphics.
Are those available for anything other than laptops yet?

[Yups]

Are those available for anything other than laptops yet?


Desktop CPU generation called Rocketlake-S is coming this month with Xe based iGPU, the desktop models have only 32 EUs instead of 96, even though the media capabilities should be the same. Also there are some NUC devices in the market based on Tigerlake-U.

[hajj_3]

Are those available for anything other than laptops yet?

30th march: https://wccftech.com/intel-rocket-lake-11th-gen-desktop-cpus-officially-launching-on-30th-march/

[Yups]

It's a pity GPUs are so expensive these days, even a small TU116 Turing with newest NVENC costs 350-500€ which is ridiculous, it makes modern iGPUs much more valuable.

[Nico8583]

Desktop CPU generation called Rocketlake-S is coming this month with Xe based iGPU, the desktop models have only 32 EUs instead of 96, even though the media capabilities should be the same. Also there are some NUC devices in the market based on Tigerlake-U.
It will be Gen 11, not 12 :confused:
Intel H410 and B460 will not be compatible with Intel Gen 11 CPU, it's a shame.

[Yups]

It will be Gen 11, not 12 :confused:
Intel H410 and B460 will not be compatible with Intel Gen 11 CPU, it's a shame.

Rocketlake-S iGPU is based on Xe architecture:


Enhanced Intel UHD graphics featuring Intel Xe Graphics architecture.
https://newsroom.intel.com/wp-content/uploads/sites/11/2020/10/Intel-Rocket-Lake-S-Architecture.pdf

[Yups]

Another test. I'm using ToS_1920x800_xdither from this thread (https://forum.doom9.org/showthread.php?t=175776) and a low bitrate of ~1 Mbit with max 120 gop (Quicksync 121, Intel recommended 5*fps+1 for Quicksync some years ago)

VMAF speed bitrate

Quicksync H265 (27.20.100.9316)
HD630 CQP 86.78 52 fps 1023 kbit

Iris Xe CQP FF best 90.04 344 fps 1028 kbit
Iris Xe CQP best 90.55 97 fps 1014 kbit

NVENC H265 (461.81)
GTX1080 CQP best 85.86 215 fps 1023 kbit


x265 (Staxrip 2.1.8.4)
i7-1165G7 ultrafast 87.23 81 fps 1017 kbit
i7-1165G7 ultrafast CRF 84.80 83 fps 1023 Kbit
i7-1165G7 ultrafast QP 83.90 83 fps 1001 Kbit

i7-1165G7 medium 90.13 28.5 fps 1013 Kbit
i7-1165G7 slow 92.47 10.7 fps 1014 Kbit



I did us CQP 27_27_29 offset 2_3_7 bframes 16 for Iris Xe if someone is curious, same on HD 630 except for the CQP parameter.

[Tenkei]

Another test. I'm using ToS_1920x800_xdither from this thread (https://forum.doom9.org/showthread.php?t=175776) and a low bitrate of ~1 Mbit with max 120 gop (Quicksync 121, Intel recommended 5*fps+1 for Quicksync some years ago)

Can you test higher bitrates (around 10 or more)?

[Yups]

Yes I can but 10Mbit on this video is pointless, it's too much for making a difference. In general the differences will narrow down with more bitrate. I may try a 4K 50 Mbit sample converted to 20 Mbit later one.

[Yups]

I'm using this this sample (https://4kmedia.org/samsung-phantom-flex-uhd-4k-demo/) with a bitrate between 18.0-18.8 Mbit and 10 bit. All of them have a good quality with this high bitrate, a faster Quicksync preset makes sense in this case. I think I could even try fastest FF.

VMAF speed bitrate

Quicksync H265 (27.20.100.9316)
HD630 CQP 97.45 15 fps 18.5 Mbit

Iris Xe CQP FF best 98.33 67 fps 18.8 Mbit
Iris Xe CQP fastest 98.09 107 fps 18.2 Mbit
Iris Xe CQP best 98.46 29.5 fps 18.7 Mbit

NVENC H265 (465.51)
GTX1080 CQP best 96.43 87 fps 18.0 Mbit


x265 (Staxrip 2.1.8.4)
i7-1165G7 medium 97.70 4.8 fps 18.4 Mbit
i7-1165G7 slow 98.44 2.1 fps 18.5 Mbit

[Tenkei]

Iris Xe really holds up against software encoding, at least in metrics. Do you see any visual differences between Iris Xe best and x265 slow? NVENC tends to remove details to get better compression for the image as a whole. Could you upload the encoded clips?