27 CPUs Benchmarked With AOM AV1, Intel SVT VP9/AV1/HEVC Video Encoders

[tildearrow]

You forgot to add the dav1d AV1 encoder to the test?

That's a decoder.

[johanb]

That's a decoder.

My bad, always mix up dav1d and rav1e, so what about rav1e then?

[milkylainen]

SVT HVEC encoding speed certainly is impressive compared to x264. h265 is usually that more CPU intensive. Almost makes me wish Intel implemented h264 in their sw-encoder setup

[andreano]

None of the tested systems could hit more than 1FPS for AOM-AV1 video encoding with the encoder in its present state.

Have you tried -cpu-used=8, and/or using tiles for parallelization? The quality-to-speed tradeoff is well worth it: Good news: AV1 encoding drops to near-reasonable levels

[ksuite]

You should do a graphics card encoding test with intel vaapi, AMD VCE and NVIDIA NVENC.
Is the video card encoding faster than cpu video encoding?

[Guest]

Thanks for the benchmarks, quite useful! It gives me hope that my older CPUs can have decently fast VP9 decoding/encoding without using the CPU too much.

[andreano]

You should do a graphics card encoding test with intel vaapi, AMD VCE and NVIDIA NVENC.
Is the video card encoding faster than cpu video encoding?

For comparing encoders, I actually disagree. Nothing wrong with comparing numbers between different hardware of course – that's benchmarking – but a proper encoder test is very complicated. I'm talking about objective metrics and BD-rate curves.

You know that encoders produce different output, right? That's why an encoder test must consider quality per bitrate, not just speed – this is a fundamental compromise. Every encoder is a different compromise that evolves over time. Actually many compromises, because of options, and a fair comparison requires tweaking these to bring the test subjects closer. In the case of VP9 and AV1, these options have terrible defaults, and there are no presets like veryfast as for other encoders – very easy to screw up in a comparison. Encoding is an art; decoding is not.

Of course, AV1 encoding speed is interesting in itself (since nobody has shown a reasonably fast encoder), but in order to actually test that, you have to enable the speed options…

Decoders are what makes sense to benchmark like this, because the output is identical.