The biggest challenge posed by AI training is in moving the massive datasets between the memory and processor.

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