Researchers propose low-latency topologies and processing-in-network as memory and interconnect bottlenecks threaten inference economic viability ...