3 ms·

I would bet that is limited to low-energy inference which is the most important impediment to deploy deep vision models in mobile devices.