Comment by frontsideair

I'm interested in this, my impression was that the newer chips have unified memory and high memory bandwidth. Do you do inference on the CPU or the external GPU?