One feature I expect we'll get from this sort of research is identifying "hot spots" that are used during inference. Like virtual machines, these could be cache

One feature I expect we'll get from this sort of research is identifying "hot spots" that are used during inference. Like virtual machines, these could be cached in whole or in part and used to both speed up the response time and reduce computation cycles needed.