Comment by ColonelPhantom

Great project! How does the performance compare with conventional CPU/GPU based inference? Those devices are usually a lot higher power (and bigger/more expensive), but obviously do not benefit from specialization.