Hacker News Reader

Comment by FullyFunctional | Hacker News Reader

I figured everyone understood this, but you are right, it bears spelling out:

Any modern [video] compression format is extremely flexible and allow great liberty for the compressor (whereas the decompressor is strictly defined). This means that there is ample room for work on improving the quality and efficiency of the compressor for years to come.

In other words, AV1, the spec, isn't some magical unicorn. It's the sandbox in which the 'corns can be raised.

View on HN

My personal analogy for this is that video encoding standards are basically like toolkits for building houses, while encoding software is like a robot trained to build those houses using the tools in question. New standards bring new tools to the toolkit, but training the robots to use them effectively is going to take its time and is often even worse than the previous robots in the beginning simply because the older robots have gotten so proficient with their respective toolkits. x264 in particular is basically the top robot in this department and the standard which any new entrants should aim to beat both in quality (not necessarily too hard) and in time efficiency (which is the much bigger challenge).

I don't think time efficiency is the primary concern for the big encoders like Netflix and YouTube. They're more interested in lowering the bitrate while maintaining image quality and are willing to throw compute resources at slower encoders which achieve that. Netflix, for example, does multiple encodes per scene in multiple formats to get the best possible quality for each format for each scene.

VP9 outperforms H.264 (libvpx versus x264) for Netflix's use case. Here are some articles from the Netflix TechBlog on their experiences with VP9 and their encoding approach in general:

https://medium.com/netflix-techblog/optimized-shot-based-enc...

https://medium.com/netflix-techblog/more-efficient-mobile-en...

https://medium.com/netflix-techblog/netflix-downloads-on-and...

You're right, time efficiency certainly isn't as big of a concern for companies with effectively infinite resources to throw at encoding, but the rest of us probably want to finish our encodes within this decade :) (the reference AV1 encoder is extremely slow)

Time efficiency is still important for live streaming. Last time I checked (OK that was one year ago) x265 was way too slow for live, even at not so high resolution.

Bitmovin produced an AV1 live stream a year ago using 200 cores: https://bitmovin.com/bitmovin-supports-av1-encoding-vod-live...

And six months ago they had improved on that to do live streaming with 32 cores: https://bitmovin.com/constantly-evolving-video-landscape-dis...

So I'll be interested to read about any live streaming demos they do at NAB in the next couple of weeks.

It's impressive that AV1 allows so much parallelism. That's great news.

It is a concern even for Netflix. The number only starts to make sense assuming a 30% bitrate reduction and 5x speed encoding time. This means the current encoder needs to speed up anywhere from 50x to 500x to get to that point.

Imagine if Netflix spends 1M on encoding, even 5x is 4M more, and the current encoder is anywhere from 250M to 2.5B more. This is not a small sum of money.

I think before they talk about speed up, they just need to iron out all the bugs and bitstream freeze before we speak.

Netflix have said they'll roll it out once it's below 5x slower.

But they can start with only the most popular videos to get the most bang per buck, they'll also likely target geographic areas with low bandwidth but decent spec desktops, which means they're getting a return on investment by increasing reach, not just saving bandwidth.

So basically, there's plenty of niches where it makes sense almost as soon as the spec is frozen and they can expand the roll out as it proves itself and speeds up.

Compressing a video looks like a huge decision tree to me, and newer standards add more interdependencies and more variables at each stage. I'd expect that the resulting optimization problems approximated by encoders are computationally difficult. (Which would explain why they always get better even after many years)

Short answer: yes Long answer: yes

Reply on news.ycombinator.com