Hacker News Reader

Comment by rossmohax | Hacker News Reader

AV1 have other costs, AFAIK it's so slow to encode that encoding farm cost become much more significant.

For now. When h.264 was new, I remember my dual G5 Mac taking 2-3 times the duration of a 1080p video to encode. Today, my phone can encode video of roughly equal resolution in near real time. It would be unsurprising to see the next generations of processors include optimizations for AV1.

Could that be in part because newer iPhones perform H264 decoding on the GPU?

IIRC there's usually a dedicated piece of hardware for h264 encoding in most phones now (not the gpu per se)

Just realized they were taking about encoding anyway. I’m not sure if that’s usually performed on the same chipset. I don’t know if iPhones have a dedicated chipset for H264, I would be a bit surprised. I could probably look it up later.

> I don’t know if iPhones have a dedicated chipset for H264,

Yes it does, both H.264 ( AVC ) and H.265 ( HEVC ).

But that doesn't really matter because your ( initial ) point still stands, real time H.264 encoding isn't taxing on an A14 anymore. Remember H.264's work started before 2000 and finalised in 2003. That is like playing Doom 3 on your iPhone. No longer technically taxing by any means even if you are playing it at native resolution.

However the industry is well aware of the slow down [1] of Moore's law. The cost increase as well as TDP headroom. So thinking we could continue to throw hardware at the problem simply doesn't work.

[1] Moore's Law actually includes a time frame within it, so it would not be wrong to say the law is "dead" considering it has slowed down way beyond its time limit within margin of error. But modern day's media and computer PR tends to use Moore's law as word for transistor improvement.

Okay, thanks. Just got a chance to look it up and indeed, beginning with the Apple A9 processor, there are H.264 and H.265 hardware decoders.

Its going to see adoption is widely distributed content first - its why Google and Netflix are so gung-ho on it, they see the same file downloaded millions of times and in those circumstances bandwidth savings are huge over the encode times.

Eventually the ubiquity of hardware acceleration plus Moore's Law should make even lackluster hardware encoders generally available so average joes can use AV1 without needing to wait a day for a half hour of footage.

Netflix decided AV1 was worth it:

https://netflixtechblog.com/netflix-now-streaming-av1-on-and...

https://netflixtechblog.com/svt-av1-an-open-source-av1-encod...

Netflix surely spends less compute power encoding videos than sites like YouTube though, right? Netflix skews more towards fewer videos, higher view counts, more money per video.

YouTube rolled out AV1 support quite some time ago. Many YouTube videos are AV1 encoded. Here's one example:

https://www.youtube.com/watch?v=k3Yk6FrjMkQ

You can tell if it's playing back in AV1 in your browser by right-clicking on the video and selecting "Stats for Nerds" to find out the video and the audio codecs. On my system in Firefox it's AV1 for video and Opus for audio.

For that video the file size savings are a bit hit and miss. AV1 saves a lot versus H.264 at 1080p, saves a little bit at 720p and 480p, and saves a bit more at 360p.

The 1080p results for that video are:

- AV1 50.28MiB

- VP9 69.12MiB

- H.264 106.68MiB

So a clear win at 1080p for AV1 versus VP9 and H.264.

Sometimes YouTube's AV1 encodes can actually end up being bigger than the H.264 versions due to the vagaries of automated transcoding.

When AV1 support was newly added to Firefox I used this 1080p60 video as a benchmark for my system:

https://www.youtube.com/watch?v=Fmdb-KmlzD8

Over the course of the development of Firefox's AV1 support and the switch to the dav1d decoder (https://code.videolan.org/videolan/dav1d) that video has gone from being unwatchable due to too many dropped frames to now playing back smoothly.

My system still struggles with AV1 above 1080p in Firefox though.

I’m interested in the relative CPU costs of encoding, so this doesn’t really address that.

When you say “many” videos are AV1-encoded, it makes me wonder why other videos aren’t AV1-encoded, if supported. Presumably there is some cost/benefit analysis at Google to decide which formats to use, and it is decided that a particular video should not be encoded using AV1 due to the resulting impact on disk/CPU/network usage (more disk, more CPU, less network egress).

I also suppose YouTube has a massive pool of low-priority batch services for transcoding and can do last-minute encoding for unpopular videos / unpopular formats, so YouTube’s tradeoff will not be the same as your tradeoff. I just wonder what the tradeoff is.

> When you say “many” videos are AV1-encoded, it makes me wonder why other videos aren’t AV1-encoded

They will be eventually. The priority goes to the videos with the highest trending view counts because that's where the bandwidth savings are going to be the most significant. This video is five days old and has an AV1 encode:

https://www.youtube.com/watch?v=j8oD9jlCFJw

Whereas this video which has been there for nearly 7 years does not yet have an AV1 encode:

https://www.youtube.com/watch?v=Dqf1BdsV5WE

With only 7,777 views it doesn't matter that much if it has an AV1 encode (or even a VP9 encode really). The bandwidth savings won't be that significant for the number of views it gets.

Also consider that all video have a H264 encoding to play on low end devices.

Any additional encoding, in VP9 or AV1, is an additional cost as the H264 encoding cannot be avoided.

Even VP9 is not used for all video. Only if the video is 4K or 2K7, or has a high number of view.

> Even VP9 is not used for all video. Only if the video is 4K or 2K7, or has a high number of view.

No, the low view count video example I provided above has VP9 encodes for all resolutions, 144p through to 1080p.

Lots of low view count videos are encoded in VP9. But when VP9 was new it followed the same approach AV1 is following now: high trending view count videos get the priority.

Those comparisons are meaningless without also factoring in some measure of quality such as VMAF. The advantages of AV1 become clear in that case.

Or you can just look at them with your eyes. The end goal of video is for it to be looked at, after all.

The quality of these different encodes is about the same. With different codecs at about the same quality it becomes a case of picking your errors and artefacts. A given codec will be weak in some areas and strong in others.

Reply on news.ycombinator.com