SVT-AV1 2.3 Brings More Performance Improvements: AVX-512 & LTO By Default, More Tuning

[phoronix]

Phoronix: SVT-AV1 2.3 Brings More Performance Improvements: AVX-512 & LTO By Default, More Tuning

SVT-AV1 2.3 is now available as the newest feature release to this leading open-source AV1 encoder. With SVT-AV1 2.3 there are yet more performance improvements...

https://www.phoronix.com/news/SVT-AV1-2.3

[galad]

The two tables are about the speed improvements on Arm, not on x86_64. It won't be 30% faster on Intel and AMD.

[decadarkness]

On every release of (non-reference) AV1 encoder I feel sad that "adaptive" keyframe placement (when you place keyframes not at fixed interval, but rather try to guess when scene changes significally) still not implmented. Of course happy to be wrong

[wswartzendruber]

Total fantasy request: An automatic grain encoding switch that:

1. Automatically determines the noise patterns in each part of each frame.
2. Automatically removes this noise.
3. Calculates the grain coefficients for each of these noise samples such that the regenerated grain will be visually similar.
4. Bundles all of this together in a much smaller bitstream that produces a visually similar result on output.

[sophisticles]

The biggest limitation this encoder has is that it does not support anything other than 8-bit 420 and 10-bit 420 output.

VP9 supports all the way to 12-but RGB, it even supports 8-bit YUVA, meaning it supports the alpha channel.

I get that SVT-AV1 is primarily designed to be a delivery format only, but when other AV1 encoders support higher bit depths and greater than 420 chroma subsampling, it's really disappointing.

[Chugworth]

Nice, I've been manually compiling SVT-AV1 because the Arch package didn't have AVX512 enabled. This should save some hassle.

[NekoTrix]

Total fantasy request: An automatic grain encoding switch that:

1. Automatically determines the noise patterns in each part of each frame.
2. Automatically removes this noise.
3. Calculates the grain coefficients for each of these noise samples such that the regenerated grain will be visually similar.
4. Bundles all of this together in a much smaller bitstream that produces a visually similar result on output.

Unfortunately, point 3 is a fantasy because the spec is very limited in terms of what kind of noise patterns can be replicated with AV1 grain synthesis.
Point 1 is doable if you remove the "in each part of each frame", it would simply require some more sophisticated algorithm. Reason why "in each part of each frame" is not possible is again a spec limitation: it doesn't allow for controlling the grain depending on parts of the image.
Point 2 is less evident than you think it is because denoising is ultimately destructive of the image, however sophisticated your denoising solution may be. The lesser destructive solutions are also extremely computationally intensive, so that's not necessarily something you may want. The currently implemented denoiser is very simple and fast, thus it is extremely destructive. It's a matter of priorities and speed is definitely the one of SVT-AV1.

The biggest limitation this encoder has is that it does not support anything other than 8-bit 420 and 10-bit 420 output.

VP9 supports all the way to 12-but RGB, it even supports 8-bit YUVA, meaning it supports the alpha channel.

I get that SVT-AV1 is primarily designed to be a delivery format only, but when other AV1 encoders support higher bit depths and greater than 420 chroma subsampling, it's really disappointing.

Rejoice then because 444 support is apparently at the conceptual stage as we speak. Again, it's a matter of priorities. They have a limited amount of people that can work on the encoder, and I agree completely with their choice to focus on maximizing the efficiency and speed of the encoder over adding arguably niche features. The reference encoder has 444 12-bit support, and it's neither bad nor insufferably slow, so one with such usecase can simply switch to that encoder. Though now that SVT-AV1 has matured to this point, they will likely be focusing on bringing these niche features over soon.

You may know that AVIF supports the alpha channel as well btw.