Streaming Factor CPU vs GPU

[alexmoini]

Does anyone have any insight on the audio speed to segmentation output speed ratio on cpus and/or gpus? I'm sure its model dependent but would be curious what the base model being used performs like.

[juanmc2005]

Hi @alexmoini,

As stated in the paper (Section 4, paragraph "real time"), it takes 165ms to process a single input with a CPU Intel Cascade Lake 6248, and 50ms with a GPU Nvidia Tesla V100 SXM2, which gives real time factors of 0.33 and 0.1 respectively.

I haven't had the time to measure this on other hardware but it's fairly easy to do with diart.stream, which prints the average time taken to process each audio buffer (then divide by the step: 500ms by default).
You can also set the --cpu flag to ignore any available GPU.

I would be very interested in reporting the real time factors in a table somewhere in the existing documentation.
If somebody has done any measurements or is willing to help, we could start adding this information.