Benchmarks of the current Dav1d (AV1 software decoder) are out.

Summary:
Dav1d scales very well with more cores and threads. They tried out a 32-core Epyc processor and it was still showing signs of increased performance for each additional thread when decoding a 1080p clip.

Performance is essentially cut in half going from 8 bit color depth to 10 or 12 bit color depth.

Firefox does not have the necessary components for using the assembly code and instead relies 100% on dav1d written in C. The C version is roughly 50% slower than the assembly version.

A quad core Haswell processor is getting around 30-50 FPS decoding 4K video. For 1080p the same processor gets well over 100 FPS (in some cases, close to 200 FPS) doing decoding fully in software.

Dav1d seems to be around 2 times as fast as the reference aomdec (AOMedia's decoder) in general, although in some clips it is up to 5 times as fast.

On mobile, the Snapdragon 835 gets around 50 FPS playing a 1080p clip.

The Apple A12X gets something like 120 FPS.

Their next focus is improving 10 bit support, as well as SSSE3 processors and ARM64 processors.