Intel Releases SVT-AV1 2.0 For Even Faster AV1 Encoding

[phoronix]

Phoronix: Intel Releases SVT-AV1 2.0 For Even Faster AV1 Encoding

Intel has published SVT-AV1 2.0 as the newest major feature release to this leading open-source CPU-based AV1 video encoder. Along with various API changes, SVT-AV1 2.0 has yet more encode performance optimizations...

https://www.phoronix.com/news/Intel-SVT-AV1-2.0

[chuckula]

Definitely interested in some benchmarks on this one!

[avis]

Definitely interested in some benchmarks on this one!

Benchmarks for a video codec? I'm interested in quality (per byte), not "benchmarks".

[pWe00Iri3e7Z9lHOX2Qx]

Benchmarks for a video codec? I'm interested in quality (per byte), not "benchmarks".

I think both are important. If the quality is amazing but you can only squeeze 1 frame per second out of the encoder, nobody is going to use it. I need to play around with this and see how well old Xeons fare.

[avis]

I think both are important. If the quality is amazing but you can only squeeze 1 frame per second out of the encoder, nobody is going to use it. I need to play around with this and see how well old Xeons fare.

For offline encoding performance is not hugely important.

[bug77]

Benchmarks for a video codec? I'm interested in quality (per byte), not "benchmarks".

Not if you're a major website handling thousands of encoding streams at the same time

[sophisticles]

Re: Speed vs quality, I seem to recall over a decade ago one of the lead x264 developers that went by Dark Shikari posting an article on his blog where he talked about testing an experimental encoder that took almost an hour to encode a single frame at the slowest settings.

There is a point where if an encoder is so slow then it's not worth using no matter how high the quality may be.

The flip side is that encoding speed is usually also a good indicator of how demanding decode is.

For instance, ignoring hardware accelerated decoding for the purposes of this discussion, if you encode x264 using one of the faster presets, then software based decode is fairly lightweight.

However if you encode with placebo, that has up to 16 B-frames, 16 reference frames, B=pyramid, 40 frame look-ahead, and a slew of other features for quality, then software decode becomes problematic especially at higher resolutions.

I like the SVT family of encoders, the HEVC and VP9 variants are also pretty good, but honestly, i think most users, including commercial entities, are better served with hardware encoders.

[avis]

Not if you're a major website handling thousands of encoding streams at the same time

Those don't need benchmarks from Phoronix. They have their own encoding settings and requirements, source videos and people to do the job. No way Michael's generic tests with standard settings are of any use for such companies.

And AV1 encoding is still out of reach for the vast majority of people unless you don't care about visually lossless encoding in which case you can use HEVC which is a ton faster to encode.

[Danny3]

Wonderful, many thanks to the developers and the organization behind it!

[Kenta_Hirono]

The flip side is that encoding speed is usually also a good indicator of how demanding decode is.

For instance, ignoring hardware accelerated decoding for the purposes of this discussion, if you encode x264 using one of the faster presets, then software based decode is fairly lightweight.

However if you encode with placebo, that has up to 16 B-frames, 16 reference frames, B=pyramid, 40 frame look-ahead, and a slew of other features for quality, then software decode becomes problematic especially at higher resolutions.

In somewhat yes, but not so much as the decoder algorithm is the same, increasing bframes reference and look-ahead slow down the encode part so much because the encoder have to search between frames and decide if are useful or not. Major decoding issue is the cache memory to store references, usually Level define how many bites a level needs to be decoded.

I like the SVT family of encoders, the HEVC and VP9 variants are also pretty good, but honestly, i think most users, including commercial entities, are better served with hardware encoders.

Hw encoders are tuned for speed and low impact as suited for on the fly compression like streamers or live chat/stream, most don't support multi-pass encoding and with highest quality settings aren't so much faster than sw encoders at the same output quality (not av1 I suspect).