Closed laurentopia closed 8 months ago
Each module polls the server to see if there are jobs available. It's first in, best dressed, since the server isn't pushing, the modules are polling / pulling.
Ok that explains what I'm seeing. I guess that's fine when all devices have roughly the same speed. What happens in the case with multiple Coral TPU? They're cheap again on Mouser...
As a test I spun up a few object detection and it seems that CP-AI uses all of them regardless of past efficiency, meaning that yolo-cpu will be used even though yolo-GPU is available (and 3x faster). Is this a simple idle-first queue? Random? Here we should see 3x more GPU yolo than CPU if scheduling prioritized speed.