3 ms·

desktop CPUs running inference on a single background thread would be the ideal case for what I'm considering.