Announcement

Collapse
No announcement yet.

VVenC 1.11 Brings More Performance Improvements For H.266/VVC Encoding

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • Originally posted by avis View Post
    And here's a funny thing: pretty much everyone now is entitled to H.265 support because they have HW which is licensed for H.265, that includes your GPU, iGPU or mobile SoC. But still Linux lawyers have decided not to support it for some reasons.
    Wasn't that license tied to driver binaries, not the GPU firmware blobs? I thought I remembered reading that.

    (Technically another reason a Cyber Monday'd RTX 3060 was the best upgrade for my needs, I suppose... though still mostly the CUDA, power efficiency for its price point, and, prior to upgrading the rest of my system, being compact enough to not force me to buy expensive ultra low-profile SATA cables.)
    Last edited by ssokolow; 26 February 2024, 05:14 PM.

    Comment


    • Originally posted by avis View Post

      I was 200% sure someone would post this asinine evil stupid meaningless comment. Kudos. I'm not disappointed!




      None. From 38 to 46% bitrate savings over AV1. Absolutely nothing.

      A HW decoder with a lot less complexity requiring probably half the AV1 transistor budget. A lot more power efficient decoding as a result.
      This test is biased and terribly conducted. Not representative. Make your own tests and see for yourself how VVenC is underwhelming.

      Comment


      • Originally posted by avis View Post

        AV1 is already more computationally expensive than VVC despite offering a worse compression ratio.

        AV2 will be even more computationally expensive (normally new codecs are an order of magnitude more expensive to encode and at least twice expensive to decode) and probably outside the reach of most people. Let's celebrate it! And someone will have to pay for that regardless.
        Nevermind, you lost all credibility here. You complain about people being untitled, but you are blind.

        Comment


        • Originally posted by avis View Post
          osw89

          Nice insults, very nice insults. Too bad they are 100% false. In fact that's probably the most egregious insults I've seen here in a long while.

          Good luck with your Open Source/Patent Free zealotry even if it's technically a LOT worse than comparable "closed" solutions. Don't let the fact that H.266 has an open spec and is 100% free for home use distract you.

          Speaking of the paper, I just reported it. I did not author it, not I claimed it's factual.

          You're welcome to do your own research which you won't because you're here with nothing but hatred and vengeance.
          So where are the counterarguments? Stop changing conversation because everything he said was true.

          Comment


          • Originally posted by NekoTrix View Post

            This test is biased and terribly conducted. Not representative. Make your own tests and see for yourself how VVenC is underwhelming.
            I've run my own tests, correct and VVC came on top. I'm 100% sure you've never touched vvenc but I have it on my disks right now. On two of my PCs.

            Originally posted by NekoTrix View Post

            Nevermind, you lost all credibility here. You complain about people being untitled, but you are blind.
            I don't see anything but ad hominem here. Do you have any actual counter arguments? Also, who assigned credibility to me and how I've lost it exactly? I've been known in the Linux community for over 25 years now, I'm featured in the kernel GIT log over a dozen times. What about you? I normally don't lie or make up stuff unless I ininadvertently may mistakes because I don't know something well enough or remember something wrong in which case I always apologize.

            Originally posted by NekoTrix View Post

            So where are the counterarguments? Stop changing conversation because everything he said was true.
            Are you his alter ego/the same person? I don't see anything but personal attacks from you. Funny to see this from an account which signed up a few days ago.

            You can download and compile vvenc in under a minute. You can test it against SVT-AV1 at any preset you want. You're free to share your results. I'm fine discussing them. Before doing that please don't bother replying to me. I want to speak to the point, not engage with open source zealotry and hatred which is in abundance here. Being a fanatic and believer is a lot easier than being rational and using your brain.
            Last edited by avis; 01 March 2024, 06:15 AM.

            Comment


            • Originally posted by avis View Post

              I've run my own tests, correct and VVC came on top. I'm 100% sure you've never touched vvenc but I have it on my disks right now. On two of my PCs.



              I don't see anything but ad hominem here. Do you have any actual counter arguments? Also, who assigned credibility to me and how I've lost it exactly? I've been known in the Linux community for over 25 years now, I'm featured in the kernel GIT log over a dozen times. What about you? I normally don't lie or make up stuff unless I ininadvertently may mistakes because I don't know something well enough or remember something wrong in which case I always apologize.



              Are you his alter ego/the same person? I don't see anything but personal attacks from you. Funny to see this from an account which signed up a few days ago.

              You can download and compile vvenc in under a minute. You can test it against SVT-AV1 at any preset you want. You're free to share your results. I'm fine discussing them. Before doing that please don't bother replying to me. I want to speak to the point, not engage with open source zealotry and hatred which is in abundance here. Being a fanatic and believer is a lot easier than being rational and using your brain.
              Ah yes, good ol' talking out of your arse without knowing the person in front of you. I have been using VVenC for longer than I did an AV1 encoder, that was in 2021, and I don't understand what you benefit from spouting lies. If I tell you to see for yourself, I think it is fair to assume that I've done testing myself too... Also you should improve your reading comprehension skills, my account has been created ~8 months ago, not mere days ago. That seriously makes me think you are a troll.​

              Now, seriously back to topic. Just by using the data available on this site, the benchmark you linked, and your claims, I can prove that you are full of shit.
              You claim AV1 is more computationally intensive and that's false. The benchmark you linked show that AV1 was 100x faster than VVC in this test. So of course, as I elluded to, and someone else did too a few pages back, this test used ffmpeg defaults which is a terrible and un-scientific testing point, especially if your goal is to compare the potential of encoders, and extremely misleading for everyone that cannot see that nuance and will simply assume from now on that indeed "VVC is 30% better than AV1 or whatever so why bother?". Still it shows a first contradiction to something you claimed. According to the Phoronix benchmark, at 4K, preset fast is 8 fps on a 14900K while preset 4 is 6 fps on the same CPU. There are 9 faster presets than this to choose from in the AV1 encoder and my intuition from experience is that going preset 5 would already tie fast in speed. Faster, just like in x264 and x265 is a joke of a preset that's not even worth talking about here.

              Now for a bit of context, I'm a huge codec enthusiast and only mildly care about the patent situation. I study signal processing and intend on working in the codec sphere soon. I wrote a blog post on a site called the codec wiki on thorough SVT-AV1 v1.8.0 benchmarking I ran in order to squeeze the most efficiency out of the encoder. My findings were used to tailor new defaults for a fork of SVT-AV1 with psychovisual target in mind. The fork, named SVT-AV1-PSY, thus provides double digit BD-rate improvements over mainline SVT-AV1, which is certainly not negligible. Internet person RootAtKali has made an AWCY equivalent script that computes metrics scores and calculate BD-rate gains for comparing encoders objectively. A modified version of that was used in my testing and since I know the encoders I'm testing, contrary to the person who made the benchmark you linked, I always make sure to compare the encoders at their best, and not by using some arbitrary defaults that don't represent well the full capabilities of the encoders. You can find the blog post at the following adress: https://wiki.x266.mov/blog/svt-av1-deep-dive (notice the site name, I'm sure you'll love it)

              So, I did new comparisons of VVenC v1.11.0 vs SVT-AV1 v2.0.0 as soon as a few days ago, and I'm afraid to report again (same conclusion I reached a year ago when I last made the comparison) that VVenC isn't worth it at all as it barely reaches AV1s efficiency and consistency over a wide range of content, including live action, while being significantly slower. I'm not biased unlike you can be so I'll give you my full conclusion on that testing. To be fair, VVenC manages to be slightly to moderately better than AV1 in some clips in both metrics and visually, but can at the same time be tied or slightly to significantly worse than AV1 in others, again both in metrics and visually. It is so inconsistent I measured a maximum std dev value in the twenties, which on a metric scale from 0 to 100 is insane. In comparison, AV1 at its worst presets wavers in the 10 points of deviation. I can explain parts of this inconsistency though: it comes from the QPA tool that leverages XPSNR in the analysis part of the quantizer to smartly allocate bits in places and moments where it will be most significant to an average person. However the consequence of that tool is under-allocating bits in complex scenes sporting fast movement and that results in terrible objective scores in those scenes, that put VVenC in a worst place than SVT-AV1 presets way faster than it. And the visual differences from one scene to another can be striking as well. Disabling QPA does help in that regard, but it suddenly makes VVenC way less appealing in average scores and thus in BD-rate showcases as well. VVenC tricks you in order to boost metrics and appear as more advanced than it actually is. But when your encoder at preset medium can beat SVT-AV1 at preset 2 in some clips and lose to preset 7 in some others, there's something wrong with it. This behavior is non-existent in AV1 encoders and it can easily be argued that consistency is a major argument in industrial applications, streaming situations, Blu-Ray quality concerns, and the like.

              I'm not denying the few merits of VVenC and H.266 as a whole though. The format supports film grain synthesis just as AV1 does, but no usable VVC encoder possesses the feature at the moment. When it does, VVC will become way more appealing visually, I'm personally excited to see it happen. VVenC does outperform the best AV1 encoders has to offer, including aomenc, in some cases, with extremely clean, grain-less content being one of them. VVdeC is surprisingly competitive with dav1d, and while not surpassing it, it is a very robust decoder which is great.

              I mostly didn't take encoding speed into account until now. But let's delve into it. At presets that yield similar visual results to AV1 equivalent encodes, VVenC is 2 to 10 times slower. That is with both SVT-AV1 and VVenC being compiled as release builds with native and LTO optimizations. v1.11.0 brought speed and only speed improvements over the last release, and while it is appreciated, any improvement in any direction is, it is not enough still. So it is fallacious to call VVenC faster when it absolutely is not. As for threading concerns, SVT-AV1, unlike aomenc or rav1e, has no problem scaling up to 16 threads unless going to the really slow presets, in which case tools inspired by the industry can help parallelize further the encoder to regain some of that lost potential speed if you wanted to. There's no merit to give to VVenC for still saturating CPUs with the slower presets if it often cannot compete in efficiency anyway. That just shows it's wasting CPU cycles for nothing, that's not something to brag about.

              x266 has yet to arrive, but it is extremely concerning how long it is taking to develop with how little we know about it. Internet user MartinEesmaa, an incredible person who made the VVCeasy repository that for years has shipped Windows, Linux and MacOS builds of MPV and ffmpeg with respectively VVC decoding and encoding support, has reached out to MultiCoreware last month to get some updates on x266 since it was supposed to release in H2 2023 and he basically learned nothing new about the current state of the encoder. Also, Beamr which was basically the core devs team who created x265 and also helped optimizing x264 is not involved in x266, and I heard insider information from someone else close to H.266 development stuff that x266 while very fast, would be quite underwhelming quality wise compared to VVenC. It can only improve with time, but as of now, VVC software support is extremely poor. Heck it only got demuxing support in ffmpeg in early 2024, before that, nothing at all. Some real VVC support is slatted to happen in 7.0, before that there is no way in hell to both encode or decode VVC without compiling stuff yourself, using unofficial patches,... The only container that can store H.266 streams at the moment is MP4, and MP4 muxing support could only be attained for a long while by using nighly builds of MP4Box, and even that was kind of a hack. Basically any new hardware you buy will have some sort of AV1 decoding support, from the latest Intel and AMD CPUs, to GPUs from all vendors, to most Mediatek smartphone chips, the high end Qualcomm and Apple chips, most Smart TVs... VVC has hardware support only in a select Smart TVs and even then that's 4 years after the specs have been frozen while for AV1 it only took 2 years for some initial support. Actually closer to a year and a half. So yes, more hardware support in future Intel architectures, potentially later this year, is promising, but adoption is incredibly lacking for a 4 years old codec and there's no enthusiasm​ for it altogether, which is why it currently looks very DOA.

              Ultimately I don't think you are wrong on everything, but H.266, as of now, is a less impressive generational jump over H.265 than H.265 was over H.264. And it's a damn shame. However the encoders aren't mature yet and ECM is on the way as well with great preliminary results, so the future of codecs is still exciting. BTW, I believe that for someone who's been active on Linux communities for over 25 years, which assumes you aren't just a kid, you act pretty immature and pretentious. I seriously think there are better ways to convey your opinion than being overly confrontational.​

              Comment


              • 1. Been extremely fatigued recently, confused 2023 Jun with 2024 Jan. My apologies.
                2. I've not claimed anything in this topic. I cited a paper. If you don't like it, you're welcome to contact its authors.
                3. "you are full of shit", I stopped reading after that.

                If you wanna have a productive unbiased factual rational discussion, you're welcome to edit your wall of text and remove instances of insults and egregious lies. You're also welcome to post your SSIM, VMAF, PSNR comparisons or good old PNGs to compare encodes visually on e.g. imgsli. Without any of this, I will again ignore your wall of text with prejudice.

                I get it people here hate anything that's patented or not open source but that doesn't mean that anything that's not open source or patented is "bad".

                Comment


                • Yeah yeah we get it, each time you don't have arguments, you go full diving mode, call the other a liar or claim what they say is 100% false. Bold of you to talk about insults and lies btw. Troll better next time clown.
                  Last edited by NekoTrix; 01 March 2024, 12:39 PM.

                  Comment


                  • So, nothing but ad hominem and not a single quotation, argument or counter argument. Lovely. Thanks God I skipped 90% of your rant with a decent dose of strong insults.

                    Dude, if you really wanted to show how "full of shit I am", you could at least quote whatever you felt was shitty/wrong/etc, ok? You didn't do the bare required minimum to use this insult. LMAO.

                    Comment


                    • The boldness to not read my post and then claim stuff that didn't happen because it suits you

                      I'm not losing any more time with a nobody that use his involvement in some random unrelated software as a position of authority and give away how clueless he is about the subject at hand when the testing methodology he mentionned is irrelevant. How amusing would it be to see the industry use "good old PNGs" as a way to assess encoder comparisons!

                      Comment

                      Working...
                      X