SVT-AV1 Performance Continues Speeding Ahead, Xeon/EPYC Video Encode Benchmarks

[phoronix]

Phoronix: SVT-AV1 Performance Continues Speeding Ahead, Xeon/EPYC Video Encode Benchmarks

The recently open-sourced Intel video encoders for VP9, AV1, and HEVC under the "Streaming Video Technology" (SVT) umbrella continue looking very positive especially for the newer VP9/AV1 video formats...

http://www.phoronix.com/scan.php?pag...-AV1-Xeon-EPYC

[FastCode]

Is it single threaded?
Why is EPYC higher than 2x EPYC?

[Spooktra]

Have you actually check the resulting video file to see if anything was produced? The reason I am asking is because yesterday I built the latest SVT HEVC encoder from git on a Ubuntu system and followed the sample command line to try and encode a wmv source using ffmpeg to pipe the file to SVT but even though the encode seemed to go fine, and fast, nothing was produced, there was no output file.

Can you share the command line your software is using to run these encoding tests?

[oliw]

Is everything else equal? Output size, movement tracking quality, introduction of artefacts, etc...

You can't just compare encoders by how long they take. That's worthless unless they're generating the same quality content.

[darkbasic]

An output quality comparison would be nice.

[andreano]

I agree with everyone here pointing out that quality also matters. Except that for AV1, speed is all that matters right now – 18 fps on 2 × Xeon Gold is obviously no cigar.

But let me point out that a lot of speed can potentially be had by flipping some switches in the encoder. Here is the speed/quality curve for the AV1 reference encoder (as measured by streamingmedia.com) – as shown, you can drop encoding time down to 5% and keep 99% of the quality:

The sweet spot is pretty much the fastest setting. Guess what the default is…