SVT-AV1 Already Seeing Nice Performance Improvements Since Open-Sourcing

[phoronix]

Phoronix: SVT-AV1 Already Seeing Nice Performance Improvements Since Open-Sourcing

It was just a few weeks ago that Intel open-sourced the SVT-AV1 project as a CPU-based AV1 video encoder. In the short time since publishing it, there's already been some significant performance improvements...

http://www.phoronix.com/scan.php?pag...Speed-Progress

[jacob]

Actually those figures are not bad at all. They seem to suggest that by 2020, it will become possible to encode 1080 lines video at around 4-5 fps or, in other words, a feature length HD movie will be encoded in under 24 hours.

Formany users that is kind of the limit of usability for personal purposes, such as rip a BRay movie and copy it on a laptop to watch on a plane during a long flight, etc. If the compression ratio is better than HEVC for the same quality, then AV1 will suddenly become very appealing.

[tildearrow]

dav1d results:

-4K-

EPYC: 91.12fps
Ryzen: 100.19fps
9900K: 131.45fps
7980XE: 135.6fps

-1080p-

EPYC: 243.28fps
7980XE: 262.23fps
Ryzen: 286.41fps
9900K: 437.79fps

[smitty3268]

Actually those figures are not bad at all. They seem to suggest that by 2020, it will become possible to encode 1080 lines video at around 4-5 fps or, in other words, a feature length HD movie will be encoded in under 24 hours.

Huh? Michael's tests show that it's already possible to get 8.5 fps now, on a fast cpu. Did you just look at the first chart and skip the rest?

[jacob]

Huh? Michael's tests show that it's already possible to get 8.5 fps now, on a fast cpu. Did you just look at the first chart and skip the rest?

Ah, sorry I misread the chart, I thought the 2nd one was HEVC.

[ms178]

I hope AMD can improve their AVX2 performance significantly with Zen 2 and maybe even support for AVX512. As shown in these benchmarks, Intel has a notable advantage where it is used extensively.

[Brisse]

I wonder how the quality differs. It's difficult to assess lossy codecs in just frames per second when the end result can vary wildly in quality depending on the encoder and what settings are used.

Anyway, looks like the most recent SVT-AV1 results are almost up to 50% of the speed I get when encoding with libvpx-vp9 so that's not too bad.

[andreano]

SVT-AV1 has added multi-threaded CDEF search

That's what I'm saying, search is another avenue for parallelism in encoders.

[edwaleni]

For more detail on the value of CDEF, see this:

https://hacks.mozilla.org/2018/06/av...cement-filter/

[_Alex_]

And it seems there are some commits pending, from SSE to AVX2... so more gains ahead...