Comment by Tade0

Only way to have hardware reach this sort of efficiency is to embed the model in hardware.

This exists[0], but the chip in question is physically large and won't fit on a phone.

[0] https://www.anuragk.com/blog/posts/Taalas.html