What is current status for hardware H.265 encoding.

[Yups]

Yeah, I'm fairly disappointed in how it turned out. I thought AMD VCE/VCN being rubbish was just a meme but I guess it's true. Almost tempted to dig a Sandy/Ivy Bridge CPU with QSV out and compare it to 8 Mbps AVC on that thing.


I have an older Kabylake CPU, I'm sure it's much better at 8 Mbps using Quicksync. Actually 8 Mbps is too high on a half decent solution for this sample, that's why I went down to 2.5 Mbps in my testing.

[PatchWorKs]

Here are my settings for HB. I use ist for UHD sources. But it's look not bad with FHD sources. Not Perfekt, but not bad for this time.
I'm testing very low QSV encoding parameters (35-40) for FHD source and I strongly suggest to use rigaya's QSVEnc instead: mutch better results.

You can easily test it through FastFlix (https://fastflix.org/):
https://raw.githubusercontent.com/cdgriffith/FastFlix/master/docs/gui_preview.png

[lt8nk]

Why not just upgrade your processor to a zen3 like a 5600? You will get a significant increase in encode speed without having to spend much money on a new platform.

Sorry for the late answer. I thought I will receive notifications for new posts...

In fact, I hadn't given all the info so as not to overload my post. The 1500x is on the classic PC. But I also have an old Synology NAS of almost 10 years, and a mini windows PC based on Atom x5-z8350 (which doesn't allow to encode in h265, only in reading), connected to the TV. This mini PC also hosts a Jellyfin server, a TV card (so NextPVR) and a bunch of other servers. Honestly, even if I tried to optimize everything, this little Atom surprises me. As I'm running out of space with 4TB in Raid 1, I was thinking of making a big change to replace both the mini PC and the NAS in a single device that would allow me for example to encode TV recordings in H265 automatically, and to increase space with a kind of Raid 5 with SnapRaid.

I think I will wait beginning of september before buying it. Actually, it would cost me about 240 € for the whole without the new HDD.

Now you know everything.

[RanmaCanada]

Sorry for the late answer. I thought I will receive notifications for new posts...

In fact, I hadn't given all the info so as not to overload my post. The 1500x is on the classic PC. But I also have an old Synology NAS of almost 10 years, and a mini windows PC based on Atom x5-z8350 (which doesn't allow to encode in h265, only in reading), connected to the TV. This mini PC also hosts a Jellyfin server, a TV card (so NextPVR) and a bunch of other servers. Honestly, even if I tried to optimize everything, this little Atom surprises me. As I'm running out of space with 4TB in Raid 1, I was thinking of making a big change to replace both the mini PC and the NAS in a single device that would allow me for example to encode TV recordings in H265 automatically, and to increase space with a kind of Raid 5 with SnapRaid.

I think I will wait beginning of september before buying it. Actually, it would cost me about 240 € for the whole without the new HDD.

Now you know everything.
Ah makes sense. What I would do in your situation then is make an UNRAID server for storage and use the docker functions to host your media server. As long as you have an Intel CPU that supports quicksync, you're pretty much good to go. An i3 or even a Pentium Gold should more than do the job. I would personally pick at least 10th gen, but if you can swing it, 12th gen would be superior.

Good luck!

[lt8nk]

At the beginning I also thought about using UNRAID but I am a bit afraid of compatibility of my USB TV card and external sound card with passthrough. I am even not sure of compatibility with Windows 11... Concerning CPU, I will definetely go for 12th generation. But I still hesistate between Pentium G7400 and i3-12100T.

[butterw2]

Dual-core G7400 with UHD-710 (av1 hw-dec and h265 hw-enc) seems interesting for a low power machine running on stock cooler (46W TDP, no boost). I would like to see some quick sync testing vs 12100.

For software encoding (ex: x264) with only 4 threads, it will not be able to compete with a 12100. But from what I figure it does beat old quadcore i5s (ex: i5-6500) in terms of performance.

[Yups]

Here's a preliminary look at how AMD fares with the Intel test clip, and it doesn't look too good. CQP 22:24 is VCEEncC's default, and you need RDNA2 for B-frames, but I'm not sure they'd help. Pre-analysis isn't supported for HEVC.


B-frames would help, how much no idea. There are some b-frames tests for H264 but not for h265. On Intel Iris Xe CQP is very effective with many bframes. I tried what bitrate I need to match or beat your WX5100 scores with a bitrate of almost 8 Mbit. I need around 4.5 Mbit.


https://abload.de/img/irisxeo5jak.png


Settings

--avhw --codec hevc --quality best --profile main --ctu 64 --bframes 14 --gop-len 240 --b-pyramid --open-gop --no-tskip --sao none --cqp 19:22:24
https://drive.google.com/file/d/1Jp3Ns8-XsQB47NDsxKPcqCyGz_KdAIUs/view?usp=sharing

[ReinerSchweinlin]

Dual-core G7400 with UHD-710 (av1 hw-dec and h265 hw-enc) seems interesting for a low power machine running on stock cooler (46W TDP, no boost). I would like to see some quick sync testing vs 12100.

For software encoding (ex: x264) with only 4 threads, it will not be able to compete with a 12100. But from what I figure it does beat old quadcore i5s (ex: i5-6500) in terms of performance.
The G7400 looks interesting as a cheap HW-Encoding machine. Do we have any info whether the Encoder in the smaller Intel CPUs are the same as the above mentioned XE ones that did well?

I am still looking for some power efficient System that could do two things:

x265 encode with a very good power efficiency (speed is not important, efficiency is king - it can run for days)

HW-Encode with the best quality efficiency..

Notebooks or NUCs with I5-1135.. CPUs came to mind. Do you guys know of any other cheap candidates?

Intel Core i3-1210U seems very promising in terms of power efficiency, but I guess its too new, can´t find any systems with it....

[perrinpages]

The G7400 looks interesting as a cheap HW-Encoding machine. Do we have any info whether the Encoder in the smaller Intel CPUs are the same as the above mentioned XE ones that did well?

I am still looking for some power efficient System that could do two things:

x265 encode with a very good power efficiency (speed is not important, efficiency is king - it can run for days)

HW-Encode with the best quality efficiency..

Notebooks or NUCs with I5-1135.. CPUs came to mind. Do you guys know of any other cheap candidates?

Intel Core i3-1210U seems very promising in terms of power efficiency, but I guess its too new, can´t find any systems with it....

Framework is selling their mainboards with an i5-1240P for $449, but you need a case to house it in and the connectivity situation might not be ideal. It is much cheaper than any nuc I've seen that's actually in stock...

https://frame.work/products/mainboard-12th-gen-intel-core?v=FRANGACP04

[Mister XY]

B-frames would help, how much no idea. There are some b-frames tests for H264 but not for h265. On Intel Iris Xe CQP is very effective with many bframes. I tried what bitrate I need to match or beat your WX5100 scores with a bitrate of almost 8 Mbit. I need around 4.5 Mbit.


https://abload.de/img/irisxeo5jak.png


Settings

--avhw --codec hevc --quality best --profile main --ctu 64 --bframes 14 --gop-len 240 --b-pyramid --open-gop --no-tskip --sao none --cqp 19:22:24
https://drive.google.com/file/d/1Jp3Ns8-XsQB47NDsxKPcqCyGz_KdAIUs/view?usp=sharing

You can e.g. from minute 6:10 wonderfully recognize the artifacts or the block formation.
Otherwise the picture looks very good for the size.

[Yups]

You can e.g. from minute 6:10 wonderfully recognize the artifacts or the block formation.
Otherwise the picture looks very good for the size.



Keep in mind it's a low bitrate. Did you try --no-tskip --sao none? I think it's better, the size and VMAF score are definitely.

[benwaggoner]

Keep in mind it's a low bitrate. Did you try --no-tskip --sao none? I think it's better, the size and VMAF score are definitely.
Have you found a source/example where --no-tskip actually improves quality? Not that --tskip is enabled in any of the presets.

[ReinerSchweinlin]

Framework is selling their mainboards with an i5-1240P for $449, but you need a case to house it in and the connectivity situation might not be ideal. It is much cheaper than any nuc I've seen that's actually in stock...

https://frame.work/products/mainboard-12th-gen-intel-core?v=FRANGACP04
Thanx, thats a good price for the given performance, indeed.

IF the smaller CPUs with built in XE or UHD7xx provide the same quality, even a smaller system could be an option (for the HW encoding part...)

[RanmaCanada]

Thanx, thats a good price for the given performance, indeed.

IF the smaller CPUs with built in XE or UHD7xx provide the same quality, even a smaller system could be an option (for the HW encoding part...)

AFAIK the asics are the same for given generations as per the wikipedia (https://en.wikipedia.org/wiki/Intel_Quick_Sync_Video) entry for Intel Quicksync. Meaning Alder Lake Mobile should have the same ASICS as the desktop variant as they are not separated between mobile and desktop.

[lt8nk]

Maybe the Intel ARC A310 card could be a good investment if associated with a low power cpu. As it can also encode in AV1, I hope that the HW encoding part for H265 would at minimum be the same (if not better) than 12th gen . But the power consumption of the card has to be low.

[perrinpages]

Thanx, thats a good price for the given performance, indeed.

IF the smaller CPUs with built in XE or UHD7xx provide the same quality, even a smaller system could be an option (for the HW encoding part...)

Intel's processors generally share the same graphics architecture for a given CPU generation...the only difference would be performance for non-fixed-function encoding (dependent on the number of EUs in the SKU).

https://www.anandtech.com/show/17278/intel-launches-alder-lake-u-and-p-series-processors-ultraportable-laptops-coming-in-march

[Yups]

Have you found a source/example where --no-tskip actually improves quality? Not that --tskip is enabled in any of the presets.


I've tried two samples and in both of them it was slightly better according to the VMAF score. It seems to be enabled in all Quicksync presets by the way.

I hope I can try out Alchemist soon. HEVC encoder seems to be improved slightly. There is only one downside, they removed the Hybrid mode. Only the full fixed function mode is supported.

[ReinerSchweinlin]

AFAIK the asics are the same for given generations as per the wikipedia (https://en.wikipedia.org/wiki/Intel_Quick_Sync_Video) entry for Intel Quicksync. Meaning Alder Lake Mobile should have the same ASICS as the desktop variant as they are not separated between mobile and desktop.
Yes, I assume the same, but I am not sure...

I assume that the encoding performance of h265 has not changed after GEN11, so assuming the generation of the first XE`s offsprings like Pentium Gold 7505 SHOULD be cheap HW-Encoding horses...

[RanmaCanada]

Yes, I assume the same, but I am not sure...

I assume that the encoding performance of h265 has not changed after GEN11, so assuming the generation of the first XE`s offsprings like Pentium Gold 7505 SHOULD be cheap HW-Encoding horses...

If you want to go to the extreme, Framework is selling their laptop motherboards with Adler Lake processors, and they have created a bios that allows it to work without a battery. Meaning that with a 3d printed case, you can have an extremely low power hardware encoding machine.

[perrinpages]

Thanx, thats a good price for the given performance, indeed.

IF the smaller CPUs with built in XE or UHD7xx provide the same quality, even a smaller system could be an option (for the HW encoding part...)

You might be interested in the LattePanda 3 Delta. It has a low power 11th gen cpu (N5105).

https://www.lattepanda.com/lattepanda-3-delta

You probably won't find readily available low power 12th gen for a while. The low end is on a slower refresh cycle.

[benwaggoner]

I've tried two samples and in both of them it was slightly better according to the VMAF score. It seems to be enabled in all Quicksync presets by the way.
I wouldn't expect VMAF to be a particularly acute metrics for comparing those differences. It can really bias towards smoothing out flatter areas.

I hope I can try out Alchemist soon. HEVC encoder seems to be improved slightly. There is only one downside, they removed the Hybrid mode. Only the full fixed function mode is supported.
Last I looked Hybrid didn't really have a compelling quality @ speed advantage over the faster x265 presets.

[Yups]

Last I looked Hybrid didn't really have a compelling quality @ speed advantage over the faster x265 presets.


I'm talking about Quicksync. Hybrid encoding mode has a higher quality than pure FF, although the pure FF from Arc seems to be better than Hybrid on Iris Xe. I will test it out soon.

CQP with many bframes should have a superior quality to the faster x265 presets by the way. And sure hybrid runs relatively slow on a iGPU due to low clock speed and shader count.

[ReinerSchweinlin]

You might be interested in the LattePanda 3 Delta. It has a low power 11th gen cpu (N5105).

https://www.lattepanda.com/lattepanda-3-delta

You probably won't find readily available low power 12th gen for a while. The low end is on a slower refresh cycle.
Thanx. I hust looked up prices. The LattePanda is around the same ballpark as a 11GEN full laptop on sale with i5-1135G7, 8GBRAM, SSD, etc... Around 350 euros seems to be the entrypoint at the moment for at least GEN11 Intel GPU full working systems.

Addition: The N5105 seems more like a GEN10 CPU... The internal GPU is one generation older than the XE ones, we talked about intensively. Its hard to get a grasp of which quality can be achieved with which version, I was under the impression that at least "XE Generation" was needed...
If the N5105 is a candidate, there is a very low cots Intel NUC :)

[Yups]

N5105 comes with a Gen11 iGPU. The hybrid encoder should be the same, low power is one generation behind. Also this Atom SKUs might only support low power mode. In this case it's much worse.

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).


I have something new to test:


https://abload.de/img/a380m2dl1.png

Quick comparison versus Iris Xe by using this (https://drive.google.com/file/d/1YX1V0SeSkYaq6Ui41vv1wcOatbLnuLSL/view?usp=sharing) sample and CQP highest quality. As already mentioned Intel Arc only supports low power mode, there is no hybrid mode.


Intel Demo Clip VMAF speed bitrate

Quicksync H265
A380 CQP FF best 92.26 309 fps 2351 kbit

Iris Xe CQP best 91.96 67 fps 2312 kbit
Iris Xe CQP FF best 91.60 249 fps 2382 kbit


Exact same settings for both:

--avhw --codec hevc --quality best --profile main --bframes 14 --gop-len 120 --b-pyramid --open-gop --weightb --weightp --no-tskip --d3d11 --sao none --fallback-rc --fixed-func --cqp 23:24:30


Arc doesn't need Hybrid mode, it's as good or slightly better than Iris Xe hybrid with the advantage of the much higher performance.

[Kurtnoise]

Could you try the AV1 hardware encoder please ?

[Yups]

Could you try the AV1 hardware encoder please ?


I haven't tested much but in this video and this bitrate AV1 is not as good as HEVC. AV1 reviews from rigaya or THG should be accurate. The software seems immature at the moment, they may improve it over time. Some things doesn't work like mbbrc or extbrc and CQP even on lowest quantizer parameter can't go to low bitrates, so I couldn't even properly compare CQP H265 to CQP AV1. However a tweaked H265 Quicksync with 14 bframes CQP is hard to beat, even with software x265 it would require the slower presets to really make a difference.


Intel added bframes+bpyramid support for the low power mode on h264, previously it was only supported in hybrid mode. This is a very important improvement for game streaming on Arc dGPUs since hybrid mode is no real option if it's the primary GPU because it lowers the gaming performance quite a bit due to the shader usage. The fully fixed H264 encoder is on the same level as hybrid on Iris Xe now. There is no Lookahead mode though, it required GPU support.

Of course AV1 is much better than h264, twitch should support AV1 as soon as possible.

[easyfab]

Intel Demo Clip VMAF speed bitrate

Quicksync H265
A380 CQP FF best 92.26 309 fps 2351 kbit

Iris Xe CQP best 91.96 249 fps 2312 kbit
Iris Xe CQP FF best 91.60 67 fps 2382 kbit


Are the speed in good order for Iris Xe ?

A380 speed isn't so imprerssive.

[Yups]

Intel Demo Clip VMAF speed bitrate

Quicksync H265
A380 CQP FF best 92.26 309 fps 2351 kbit

Iris Xe CQP best 91.96 249 fps 2312 kbit
Iris Xe CQP FF best 91.60 67 fps 2382 kbit


Are the speed in good order for Iris Xe ?

A380 speed isn't so imprerssive.




I have a PCIe 3 system without rBAR, I cannot rule out that it effects the encoding performance. The speed is impressive imho. H264, H265, AV1 all very similar in performance. In the sample above it goes over 500 fps in balanced preset with a small quality loss.

Iris Xe was very fast already in low power mode: https://forum.doom9.org/showpost.php?p=1940526&postcount=451


It's faster than a RTX 3090 Ti: https://www.tomshardware.com/reviews/intel-arc-a380-review/5

[ReinerSchweinlin]

Intel Demo Clip VMAF speed bitrate

Quicksync H265
A380 CQP FF best 92.26 309 fps 2351 kbit

Iris Xe CQP best 91.96 249 fps 2312 kbit
Iris Xe CQP FF best 91.60 67 fps 2382 kbit


Are the speed in good order for Iris Xe ?

A380 speed isn't so imprerssive.
309 fps is quite impressive, most other cards only get this fast when run in "Fast" settings with much lower quality...

[Yups]

I finished another test including VP9, H264, H265, AV1 CBR and also ICQ+CQP from my A380. For this test I'm using QSVEnc 7.21 and I decided to use out of the box settings, I only changed the gop length to 120 and bframes/gop-ref-dist depending on the codec and bitrate mode. gop-ref-dist 8 for AV1, bframes 5 for H265 CBR etc. H265 CQP is the only exception, I tweaked it (bframes 14, open gop, SAO+tskip off), basically this is the max possible quality. However, this optimized CQP quality isn't reachable for most tools, Handbrake doesn't even support open gop.




ToS_1920x800_xdither.y4m (https://forum.doom9.org/showthread.php?p=1853595#post1853595) PSNR SSIM VMAF speed bitrate

Quicksync H265
Intel Arc A380 CBR H264 38.965 0.9636 91.98 210 fps 1898 Kbit
Intel Arc A380 CBR VP9 38.430 0.9611 92.67 148 fps 1901 Kbit
Intel Arc A380 CBR H265 40.461 0.9710 94.92 208 fps 1898 Kbit
Intel Arc A380 CBR AV1 40.358 0.9713 94.24 205 fps 1899 Kbit

Intel Arc A380 ICQ AV1 41.570 0.9736 94.29 215 fps 1900 Kbit
Intel Arc A380 CQP AV1 41.584 0.9740 94.18 219 fps 1895 Kbit
Intel Arc A380 CQP H265 (optimized) 42.163 0.9756 95.59 218 fps 1909 Kbit



Iris Xe Hybrid CBR H264 37.625 0.9622 91.42 206 fps 1897 Kbit
Iris Xe FF CBR H264 37.441 0.9549 90.24 284 fps 1897 Kbit
Iris Xe FF CBR H265 40.166 0.9700 94.00 285 fps 1898 Kbit



VMAF scores are odd, CQP/ICQ difference to CBR is usually a lot higher, PSNR+SSIM differ a lot more.

Hardware decoding doesn't work on this RAW sample and that's why the A380 is relatively slow and missing rBAR doesn't help.

Intel improved H265 slightly over Iris Xe and apparently they improved H264 quite a bit as well which surprised me. I mean even Iris Xe was really good on h264 compared to Nvidia.

[benwaggoner]

VMAF scores are odd, CQP/ICQ difference to CBR is usually a lot higher, PSNR+SSIM differ a lot more.
In terms of perceptual correlation, VMAF > SSIM > PSNR (although even VMAF still has mediocre correlation).

But for more than a few seconds of video, taking the mean of the scores of individual frames is only broadly indicative of overall quality. A video that's a stable VMAF of 80 is a lot more perceptually pleasant than one that oscillates between 60 and 100. Visual stability and interframe coherency are a huge deal, although hard to capture in simple metrics.

This is a particularly big deal for CBR versus VBR comparisons. In essence, CBR maintains a stable bitrate by varying quality, and VBR maintains a stable quality by varying bitrate. Even if CBR yields the same bitrate and average VMAF as a VBR, the VBR will often look markedly better overall.

The closing gap between CBR and VBR could come from better CBR rate control with a longer lookahead so the encoder has more flexibility to optimally allocate bits over a several second period.

This can have a big impact. Try comparing two identical CBR x265 encodes, one where --rc-lookahead=keyint, and another where --rc-lookahead=1. The former will generally look quite a bit better due to better rate control. Or compare 1-pass CBR to 2-pass CBR.

[Yups]

Intel Flex 140 with 4 decoder+encoder, twice as much as A380 :D


https://abload.de/img/intel-innovation-flexrhciv.jpg
https://www.nextplatform.com/2022/10/04/different-gpu-horses-for-different-datacenter-courses/

[rwill]

8/16.8 FLOPs is a bit low. Still faster than Zuse Z3 but the Z3 is somewhat old.

[Blue_MiSfit]

Quite promising for dense and power efficient live encoding!

[TEB]

Is the Videopipeline on the A770 the same as the FLEX 170 ?

[Yups]

Is the Videopipeline on the A770 the same as the FLEX 170 ?


By the looks of it they have the same media unit. DG2/ATSM have the same encoding/decoding/processing features, they are in the same table: https://github.com/intel/media-driver/blob/9de86486b373575e1ab1270443a52104edeab61a/docs/media_features.md

[Yups]

Actually there is a github page with lots of informations and tests from Flex series: https://github.com/intel/media-delivery/blob/ddfbad8bc5d3134b7d2ee2ae17d8aacb226a717b/doc/benchmarks/intel-data-center-gpu-flex-series/intel-data-center-gpu-flex-series.rst

[Yups]

Currently testing Sol Levante on Arc A380 in CBR mode on different bitrates. AV1 is a lot better than HEVC in this case, at least in these very low bitrate settings I have tried.


https://abload.de/img/levantetrflg.png

[benwaggoner]

Currently testing Sol Levante on Arc A380 in CBR mode on different bitrates. AV1 is a lot better than HEVC in this case, at least in these very low bitrate settings I have tried.
Have you done a visual comparison? Libaom received specific VMAF tuning, so encoders based on it tend to yield somewhat higher VMAF scores than a double-blind comparison would yield.

That said, AV1's strength has primarily been in <<1 Mbps bitrates. I've not found example much >500 Kbps where AV1 offers a consistent perceptual improvement over well-tuned HEVC.

[Yups]

Have you done a visual comparison? Libaom received specific VMAF tuning, so encoders based on it tend to yield somewhat higher VMAF scores than a double-blind comparison would yield.

That said, AV1's strength has primarily been in <<1 Mbps bitrates. I've not found example much >500 Kbps where AV1 offers a consistent perceptual improvement over well-tuned HEVC.


Yes I made some visual comparisons and it looks noticeably better on AV1, less blocky. This is basically default CBR, nothing special.


AV1= https://abload.de/img/av1n3de3.png
https://abload.de/img/av1_2qpcqf.png

HEVC= https://abload.de/img/hevcwlcyc.png
https://abload.de/img/hevc_20qcm0.png

[Yups]

I could improve HEVC with open gop and further improve with VBR, it comes closer to AV1 but the gap is still relatively big. I also improved AV1 slightly with gop-ref-dist 4 instead of 8. AV1 VBR isn't better and CQP/ICQ is worse for this test case. AV1 QSV is better than x265 slow ABR in this test.



SolLevante_SDRv2_1080p24_8bit (https://forum.doom9.org/showthread.php?p=1853595#post1853595)

Quicksync QSVEnc PSNR SSIM VMAF VQM Bitrate
Arc A380 CBR H264 32.763 0.9018 63.52 1.956 1380 Kbit
Arc A380 CBR VP9 33.052 0.9032 68.43 1.930 1375 Kbit

Arc A380 CBR H265 33.996 0.9173 71.66 1.676 1384 Kbit
Arc A380 CBR H265 open gop 33.997 0.9174 71.86 1.672 1384 Kbit
Arc A380 VBR H265 open gop 34.169 0.9202 73.17 1.630 1381 Kbit

Arc A380 CBR AV1 gop-ref-dist 8 34.328 0.9221 74.21 1.578 1379 Kbit
Arc A380 CBR AV1 gop-ref-dist 4 34.416 0.9223 74.56 1.572 1383 Kbit

x264 slow ABR Handbrake nightly 33.170 0.9086 64.62 2.083 1381 Kbit
x265 slow ABR Handbrake nightly 34.361 0.9187 72.48 1.729 1379 Kbit

[tormento]

I could improve HEVC with open gop
Can we say that ARC is encoding HEVC better than x265 (at least slow)?

[excellentswordfight]

Can we say that ARC is encoding HEVC better than x265 (at least slow)?
I doubt that its true as a general statement. I did a test recently with hevc on a HD770 (xe igpu) and i preferred x264 over it in that case.

I've not use vmaf that much, but isnt low 70 rather low quality either way?

[rwill]

I doubt that its true as a general statement. I did a test recently with hevc on a HD770 (xe igpu) and i preferred x264 over it in that case.

I've not use vmaf that much, but isnt low 70 rather low quality either way?

70 is eye cancer territory. And the sequence has around 28% black/white text only credits so VMAF for normal scenes is probably even lower. The numbers Yups published here cannot and should not be used to compare the different encoders realistically. Besides the unrealistic crap quality you also cannot compare CBR and ABR encodes, thats not how the objective metrics he used work. The results are useless.

Using some clean Tears of Steel 1080p source and target ABR rates from 1500k to 5000k, in maybe 4 or 5 steps, with a tune for PSNR and using global PSNR for comparison might give a better indication of the encoder implementations compression efficiency.

[Yups]

Can we say that ARC is encoding HEVC better than x265 (at least slow)?


No I wouldn't say. To make a statement like this it requires much more samples and much more variation in bitrate, video samples etc. I would expect that x265 slow is better in most cases because it's more stable. I may post other videos where x265 slow will be better.

I doubt that its true as a general statement. I did a test recently with hevc on a HD770 (xe igpu) and i preferred x264 over it in that case.

I've not use vmaf that much, but isnt low 70 rather low quality either way?



What software did you use? The software is a big problem for hardware encoder. Most of the software is very basic without any options and some software like Handbrake are using poor settings. For example Handbrake is using a very low gop size of 1s, on a 24fps video like this the gop size of 24 is low when x265 default max gop is 250 afaik. It can be changed in Handbrake with custom settings but nobody does it. Open gop isn't even supported in Handbrake, for x265 it's enabled on default afaik.


70 is eye cancer territory. And the sequence has around 28% black/white text only credits so VMAF for normal scenes is probably even lower. The numbers Yups published here cannot and should not be used to compare the different encoders realistically. Besides the unrealistic crap quality you also cannot compare CBR and ABR encodes, thats not how the objective metrics he used work. The results are useless.

Using some clean Tears of Steel 1080p source and target ABR rates from 1500k to 5000k, in maybe 4 or 5 steps, with a tune for PSNR and using global PSNR for comparison might give a better indication of the encoder implementations compression efficiency.


I can post a 5000 kbit bitrate result with tuned PSNR. From what I can see it doesn't change much though.

[Yups]

Here the result. It's not fundamentally different to the low bitrate results.


SolLevante_SDRv2_1080p24_8bit (https://forum.doom9.org/showthread.php?p=1853595#post1853595)

Quicksync QSVEnc PSNR SSIM VMAF VQM Bitrate

Arc A380 VBR H265 open gop 37.546 0.9570 90.98 0.754 5256 Kbit
Arc A380 VBR AV1 gop-ref-dist 4 38.259 0.9602 91.67 0.724 5255 Kbit
x265 slow ABR Handbrake PSNR tune 38.270 0.9568 90.68 0.774 5262 Kbit

[tormento]

Here the result.
Can you link the resulting videos?

Are you using the best parameters on ARC or "medium/default" ones?

[Yups]

Can you link the resulting videos?

Are you using the best parameters on ARC or "medium/default" ones?


I use best preset/TU1. Command line looks like this: --avsw --codec hevc --output-depth 8 --quality best --gop-len 120 --vbr 6500 --bframes 5 --open-gop


Here the samples: https://drive.google.com/file/d/1BMAEsC10KPdD6eTuElq5upWGi45iVifS/view?usp=sharing

edit: Davinci Resolve Quicksync AV1 is broken, it stutters. Maybe same issues as in Handbrake: https://github.com/HandBrake/HandBrake/issues/4570

I guess they are not even aware of this lol. Also it's slow because hardware decoding doesn't work. keyframe 1 on default poor quality, it can be changed though. Gop ref dist can't be changed, this is a problem assuming they are using dist 1. As I said the software is a big problem. edit2: frame ordering is the culprit for Davinci Resolve. It's enabled by default, the stutter is gone without frame ordering.

[outhud]

I'm close to buying a cheap Intel mini PC for realtime 1080p H265 encoding.
I'm trying to decide between Gemini Lake N4020 and newer Jasper Lake N5105. Both have Intel UHD Graphics.

Considering both can do realtime 1080p H265 HW encodes, is there any reason to think that there would be a quality improvement using the newer processor?

[tormento]

Here the samples
Can you please tell us the encoding speed?

Can you increase/decrease quality factor to see how the samples scale?