If you're going to be doing ML at the edge, NPUs still seem like the most efficient use of die space to me.
If you're going to be doing ML at the edge, NPUs still seem like the most efficient use of die space to me.