Slashdot Mirror
Netflix Finds x265 20% More Efficient Than VP9 (streamingmedia.com)
Reader StreamingEagle writes (edited): Netflix conducted a large-scale study comparing x264, x265 and libvpx (Google-owned VP9), under real-world conditions, and found that x265 encodes used 35.4% to 53.3% fewer bits than x264, and between 21.8% fewer bits than libvpx, when measured with Netflix's advanced VMAF assessment tool. This was the first large-scale study to use real-world encoder implementations, and a large sample size of high quality, professional content.A Netflix spokesperson explained why they did the test in the first place; "We wanted to understand the current state of the x265 and libvpx codec implementations when used to generate non-realtime encodes optimized for OTT use case. It was important to see how the codecs performed when testing on a diverse set of premium content from our catalog. This test can help us find areas of improvement for the different codecs."
I just want websites to use the HTML5 video player as opposed to Flash. x265 is not very important except for 4K content and mobile phones. It will, though, eventually become the standard.
There seem to be a lot of these lately. Why is it that as companies get bigger, they get less competent, despite having hired more of the competent people?
Alternative Right.
As long as folks don't just re-encode h.264 content. Record things in with an x265 encoder, at the beginning, and have more actual bits in the B and P frames.
I thought VP10 was supposed to be the real competitor to HEVC.
This is not the case of video codecs.
Video is a huge bandwidth hog and small improvements can really make a difference. This is why we still use JPEG and MP3 but video benefits from the latest technologies. It is also why Vorbis (audio) is much more successful than Theora (video). Both are patent-free formats from xiph.org.
Additionally, with hardware improving, more advanced compression algorithms can be used. For example entropy coders are mostly a performance/ratio tradeoff and newer standards tend to use more advanced schemes. Not that these couldn't have been used before, but the hardware requirement were too steep at that time. And you can't really put "insert preferred encoding scheme here" in your standard to make it future proof, because it wouldn't be a standard. All parts have to work together. You don't want to specify a super duper filter that melts CPUs and botch the job with a crappy entropy coder, you have to balance each part to get the best of your target hardware.
In my totally subjective and anecdotal experience the perceived quality of different mediums and resolutions goes like this:
4K Blu-ray > Upscaled HD Blu-Ray > 4K Netflix/Amazon streaming = Native resolution HD Blu-ray > Upscaled HD streaming > Native resolution HD streaming > Anything DVD
You must be referring to an article on Toms Hardware. The journalist was referring to the same Netflix study, but he was confused by the results, and he jumped the gun. He changed his title the next day. In the Netflix study they used x264 and x265 settings that are optimized for visual quality, and they also did test encodes with settings optimized for PSNR measurement. But then Netflix actually published results of PSNR measurements using the visual quality optimized encodes, which are not meaningful results. On these results, for 1080P content, VP9 appeared to be slightly better than x265. What matters most with video encoders is the actual visual quality they can produce at any target bit rate. Objective measurements like PSNR and SSIM don't correlate strongly with human subjective visual quality testing. Netflix sponsored development of a new visual quality tool, VMAF, which correlates with human subjective testing much more strongly. With VMAF, Netflix showed that x265 produces identical quality to VP9 at ~ 20% lower bit rates.
I have noticed that x265 requires much more CPU for encoding AND decoding than x264. For example, my slightly aged laptop will not handle playing my 1080p x265 streams.
This is still true:
"Quote from Netflix on their blog regarding this:
Here’s a snapshot: x265 and libvpx demonstrate superior compression performance compared to x264, with bitrate savings reaching up to 50% especially at the higher resolutions. x265 outperforms libvpx for almost all resolutions and quality metrics, but the performance gap narrows (or even reverses) at 1080p."
https://entertainment.slashdot...
New things are always on the horizon
All these codecs allow you to choose the bitrate, so efficiency is meaningless without a common basis for comparison. In this case it turns out they mean efficiency at the same video quality. But video quality is a completely subjective thing - how can you compare it in a reproducible manner? So I dug into how Netflix is deterministically measuring something subjective. That in itself is a pretty fascinating read.
tl;dr - they took subjective test results from showing video samples to people, then used machine learning to develop an algorithm which produced similar results.
"The 20% is just over VP9."
So, basically, VP9 offers little to no advantage over 264, while even 265 is somewhat limited.
That strikes me as somewhat sad, considering all the verbiage wasted by Google on the licensing issue.
> Just count the bits. Done.
That is naive. Perception is *not* uniform, that is non-linear, and I'm not even talking about gamma correction.
Can you tell the difference in quality between:
* These 2 grays? 0xFFFFFF and 0xFEFEFE ?
* What about 0xFFFFFF and 0xFDFDFD
* Or between 0xFFFFFF and 0xFCFCFC
It is possible to encode these in the same number of bit but yet one will be more accurate then the other(s).
The only results that showed VP9 matching x265 were with PSNR measurements of Visual Quality optimized encodes. These are not valid measurements (if you want to measure PSNR of x264 or x265, you need to use --tune psnr, which Netflix did for a different set of encodes). And here's a quote from the article above... "I asked Netflix which set of results they felt was most significant. Their response was, “We believe that VMAF results will have the best correlation to user perception of quality. We use this metric, and sanity-check against other metrics (PSNR, SSIM, VIF, etc.) internally.” In other words, according to Netflix, the results shown in Figure 1 were the most relevant of the three."
Even with JPEG and non-video compression, there's a lot of optimising that can be done that often isn't. I've seen and managed tons of sites that seem to think that every has big internet pipes, so it's perfectly OK to use a 2MB un-optimized image as the main backdrop for your site (or worse, a 2MB 4k resolution image that is reality scaled down when actually used). Throw that together with ever-increasing web libraries/toolkits and sometimes you're several megs in just to load a single page. Yes, there's often a somewhat optimized "mobile" version but often it doesn't load automatically on all devices, looks like ass, or is missing functionality. Enabling an optimised compression profile does help this somewhat.
Seriously though, web-devs, you shouldn't need a sysadmin to tell you that your 2MB banners and several megabytes of javascript libraries, plus a brick-ton of 3rd-party libs is *not* making for a good web experience.
Speaking of which (HTML5 Video and Netflix):
The IETF has a working group to produce a new gen video codec "NetVC" (Designed to be easy for wide adoptions, as the previous efforts of the IETF like Opus for audio).
The main candidate is by a group called "AOMedia" (association for openmedia), working on AV1 (AOMedia's Video codec 1).
The association includes:
- Google (of Youtube fame) : They are using their current development as a base for AV1 (what would have become VP10 if there wasn't this whole NetVC story).
- Xiph (of Vorbis and Opus fame, with also contributions toward Flac, Speex, etc.) : They are developing a very interesting project called Daala, and they ended up also contributing the innovation done for Daala into AV1.
- Cisco : They gave what they have developed for their Thor codec also into AV1.
Netflix has also joined the AOMedia and they are investing resources into it.
Same with several browser makers (including Mozilla).
With all the people involved:
- you know there's some interesting performance coming (given the brains involved here, given past successes like Opus, and given the promising results of research projects like Daala).
- given that 2 top content providers like Google (Youtube) and Netflix are on board, there's a high chance of seeing deployment of the new codec.
- given that browser makers like Google (chrome), Mozilla, and Microsoft (Edge) are on board, there's a high chance of seeing browser support for the new codec.
- given that hardware chip maker like ARM, Intel, AMD, Nvidia, etc. are on board there's going to be hardware decoding support.
(Adobe is on board too, so browser support is guaranteed for the Widevine DRM plug-in required by Netflix' licensors. Not that it matters that much, because that part of HTML5 Video is already defined and deployed everywhere, except maybe with Firefox on Linux which is a bit delayed)
But you know that this looks promision,
and maybe same time next year, we'd be reading summaries along the lines of "Netflix and Google find AV1 20% more efficient than HEVC/H.265" "And also cheaper, royalty-free and widely supported"
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
This is why we still use JPEG and MP3 but video benefits from the latest technologies.
They also remain because of the same reason ipv4 remains: noone wants to move first. There are better codecs for still images (webp) and for audio (AAC or opus), but nobody uses them as the original formats weren't available. Heck, its hard enough for us to get rid of gifs for animated images.
It is also why Vorbis (audio) is much more successful than Theora (video). Both are patent-free formats from xiph.org.
That's 1 of the reason.
But other reasons are in play:
- Vorbis was released back at a time when music decoding was still an important task for low-power embed devices.
And MP3-hardware decoding cheap where available for cheap.
By the time most PDA/Smartphones/Tablet had enough power to decode audio on CPU without any difficulties, other better competitor were starting to appear: things like AAC (which are not free and patent encumbered).
Meaning Vorbis wasn't competitive anymore.
Vorbis still saw quite some use as a audio/soundtrack codec for video games. (It just simply did attract much attention the role it played into this position).
- Vorbis itself has been superseeded.
Opus is the successor, partly developped by the same Xiph guys, partly developped based on technology donated by Skype.
The result is a codec that beats the crap out of anything else (except for some very special corner cases at a tiny bandwidth that isn't relevant on the internet).
Including out of AAC.
It's royalty free, not patent encumbered with available opensource code.
It's established as an official standard for web audio by IETF.
This it is widely use by tons of modern applications: Skype (obviously), WhatsApp, etc.
So the main reason that Vorbis isn't successful these days, is because Opus is.
Now keep in mind that IETF has a video codec work group called NetVC,
and that AOMedia (Alliance for Openmedia) is working to provide them with codec AV1
ant that alliance includes the same Xiph guys, but also other codec makes (Google [VP10], Cisco [Thor]), browser makers (Google, Mozilla, Microsoft) content providers (Google [Youtube], Netflix) and hardware makers (ARM, Intel, AMD, Nvidia)
There's a high chance that the same will happen in the video world, with AV1 taking everything just like Opus for the audio.
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
VP9 and H264 have hardware decoders already. You only have to buy recent hardware :). Even H265 has, in the iphone 6.
As all graphics chip makers (Intel, AMD, nvidia) are sitting on the AV1 table, hardware support for it will be almost guaranteed.
The only bigger company of significance which is _not_ member of AOM is apple, who suprise suprise sits at the MPEG table and makes bucks with their H.265 patents.
Which also explain why they insist of using AAC for everything audio,
despite the tremendous success and performance of Opus - the previous similar success story of company collaboration (Xiph and Skype) to provide a standard to the IETF.
(Opus is currently used by Skype, WhatsApp, etc. - basically if it's on the web today, it probably uses Opus as an audio codec).
(I really wish IETF and AOMedia similar succes for their NetVC initiative/AV1 codec.
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]
Throw away proprietary FPGA implementations, Google makes the real stuff Free and Open Source, software/hardware. I would never ever attempt to use x265 proprietary cruft. It is time every one make VP9 available in hardware by default.
The tyranny of backwards compatibility dictates we can never use anything new because some nitwit is using 10-year old software, of course they will never bother to upgrade unless things start breaking so....
Instead we get a chunk of silicon that can't be used for anything except x265
That is not at all the case, for instance all of Apples advanced math libraries (under the umbrella "Accelerate") will make use of a GPU if one is present and faster than at the CPU. They handle things like linear algebra stuff (linpack) and all sorts of vector operations, including FFT...
"There is more worth loving than we have strength to love." - Brian Jay Stanley
Assuming the video encode is a one time thing... wouldn't there be an energy cost if one codec takes a lot more computation on the decode end? I assume there would be an energy savings on the datacenter and bandwidth end... That would be an interesting benchmark.
I noticed that my favorite bluray ripper started supporting h.265, previously I had been using h.264, so I decided to give it a try. I took about twice a long to encode, and was a bit smaller in size, but 2 of my home computers could not playback the video. H.265 requires quite a bit more CPU/GPU processing power for both encoding and decoding. I personally will stick with h.254 because the cost of upgrading those machines is more than the cost of hard drive space.
If netflix switches they may find that a lot of home users can no longer use the service.
It's also a tyranny of "good enough". Seriously when a 40mpxl image comes down as a 10MB JPEG and look identical to any better format it really doesn't make much of a difference. Doubly so that most images are not 40mpxl.
Video on the other hand is not yet good enough.
A small nitpick, but sad to see a common but serious maths error in a technical article.
20% fewer bits is not equivalent to 20% more efficient, but 25% more .
Efficiency would be the reciprocal of the bitrate. A ratio of 4:5 becomes 5:4 when looked at the other way around.
If you were to halve the bitrate, it would be twice as efficient, not 50% more.
Or to put it in simple money terms, its like if two items are $100, one gets a 20% discount to $80, the other is now 25% more expensive.
VP9 is free, as in beer. There's something to be said for that.
Or, do you want to keep sticking your head in nooses?
https://www.youtube.com/c/BrendaEM
I didn't know about this, so researched and found this page that explains "10-Bit H.264".
I offer it here so that someone can complain that I am karma whoring.
I come here for the love
not a chance.. There's no way 720x480 has a shot at conveying the same information as HD, upscaled and edge sharpened or not..
..and sometimes the old stuff is actually better, eg: pre 2005 desktop interfaces vs post 2005..
Your connection to them is probably choking and down-grading the resolution.
Strangely enough your own comment supports the idea that Netflix is inferior to the original media.
If DVD represents the native resolution of the original, Netflix isn't going to be any better. Plus you won't have to worry about Netflix doing anything stupid with the original like cropping or zooming it.
A Pirate and a Puritan look the same on a balance sheet.
It all depends on how bad the "HD" encode is. Some of them can be quite horrible. It's quite easy to be stingy on the bitrate and end up with a disaster.
A Pirate and a Puritan look the same on a balance sheet.
This is true..
The incredibly arbitrary nature of your claim should be some indication that it is factually incorrect.
The title is misleading.
We sampled 5000 12-second clips from our catalog, covering a wide range of genres and signal characteristics. With 3 codecs, 2 configurations, 3 resolutions (480p, 720p and 1080p) and 8 quality levels per configuration-resolution pair
and then
x265 and libvpx demonstrate superior compression performance compared to x264, with bitrate savings reaching up to 50% especially at the higher resolutions. x265 outperforms libvpx for almost all resolutions and quality metrics, but the performance gap narrows (or even reverses) at 1080p.
So the highest resolution they tested was 1080p and performance between the 2 codecs was very close with libvpx beating out x265 in some cases. As far as bandwidth goes, saving at 1080p and above is more valuable than saving at 480p. Practically everything we watch at home is streamed 1080p. I don't see that x265 is the winner here. And where are the 4k tests?
Hey, that's great Netflix. Nice to see progress on the horizon in video encoding tech. Now would you please add an option to buffer the start of shows so they don't look like pixelated crap for the first 30 seconds or more on my HDTV? Maybe a checkbox somewhere? Even my wife notices, and she's not usually picky about these things. Thanks.
my, your, his/her/its, our, your, their
I'm, you're, he's/she's/it's, we're, you're, they're
Google owns Widevine, not Adobe. Did you mean Primetime?
Ooops. My mistake. I confused the DRM plug-ins.
Thank you for rectifying
(Tells you how often I use DRM in my day-to-day life)
"Sufficiently advanced satire is indistinguishable from reality." - [Tips: 1DrYakQDKCQ6y52z6QbnkxHXAocMZJE61o ]