•This assumes that most computations access data from local DRAM memory, with little or no cache re-use (typical of many applications). •If substantial long-distance communication is required, the concurrency requirement may be even higher!