3 ms·

This kind of analysis is interesting but must be considered rapidly outdated since we won't know what's changing behind the scenes at YouTube. Have they ever re

This kind of analysis is interesting but must be considered rapidly outdated since we won't know what's changing behind the scenes at YouTube. Have they ever revealed whether their VCU handles the audio?

cf100clunk 2y ago
From 2022, so a similar analysis today would be revealing if substantial differences were seen/heard.
- jeffbee 2y ago
  Sure. I think it's hard to control for every relevant factor though. We won't really know what stream they choose to serve under given circumstances to various clients, for example.
supertrope 2y ago
Video takes up almost all of the bitrate/CPU time/ASIC budget, audio is a rounding error. Server CPUs can encode audio at hundreds of times faster than realtime so there’s no need for hardware acceleration. Back in the iPod era DSPs were used to decode MP3/AAC but now only the most CPU or battery constrained devices like AirPods need hardware acceleration.
- jeffbee 2y ago
  Do you have information on this or are you just supposing? The VCU accelerator is supposedly 100x faster than software for VP9 and they have up to 20 accelerators per host. At those ratios it's not clear that a machine would be able to keep up with the audio processing, or that it would be easy to manage the division of labor between accelerators and host CPUs. Also there definitely are broadcast transcoders on the market that do the audio in hardware.
  - gruez 2y ago
    see: https://openbenchmarking.org/test/pts/encode-opus https://openbenchmarking.org/test/pts/encode-opus Modern CPUs seem to be able encode opus at around 266x speed. In other words they can encode 266 seconds of audio in 1 second. The tests results also don't scale with core count, so the program itself is probably single threaded and therefore you could encode even faster if you have concurrent streams. It's highly unlikely that even with VCUs, a server can encode video streams at several thousand times faster than playback speed.

cf100clunk 2y ago

From 2022, so a similar analysis today would be revealing if substantial differences were seen/heard.

jeffbee 2y ago
Sure. I think it's hard to control for every relevant factor though. We won't really know what stream they choose to serve under given circumstances to various clients, for example.

supertrope 2y ago

Video takes up almost all of the bitrate/CPU time/ASIC budget, audio is a rounding error. Server CPUs can encode audio at hundreds of times faster than realtime so there’s no need for hardware acceleration. Back in the iPod era DSPs were used to decode MP3/AAC but now only the most CPU or battery constrained devices like AirPods need hardware acceleration.

jeffbee 2y ago
Do you have information on this or are you just supposing? The VCU accelerator is supposedly 100x faster than software for VP9 and they have up to 20 accelerators per host. At those ratios it's not clear that a machine would be able to keep up with the audio processing, or that it would be easy to manage the division of labor between accelerators and host CPUs. Also there definitely are broadcast transcoders on the market that do the audio in hardware.
- gruez 2y ago
  see: https://openbenchmarking.org/test/pts/encode-opus https://openbenchmarking.org/test/pts/encode-opus Modern CPUs seem to be able encode opus at around 266x speed. In other words they can encode 266 seconds of audio in 1 second. The tests results also don't scale with core count, so the program itself is probably single threaded and therefore you could encode even faster if you have concurrent streams. It's highly unlikely that even with VCUs, a server can encode video streams at several thousand times faster than playback speed.