The point is offloading ML workloads to hardware that is energy efficient, not necessarily "fast" hardware. You want to minimize the real and energy costs at t

The point is offloading ML workloads to hardware that is energy efficient, not necessarily "fast" hardware. You want to minimize the real and energy costs at the expense of time. Assuming NPUs don't get pulled from consumer hardware altogether, theoretically the time/efficiency trade-off gap will become smaller and smaller as time goes on.