Abstract :
In order to work efficiently, clusters for high performance computing require a balance between the compute, memory, inter-node communication, and I/O. Fast communications among one thousand multicore nodes requires short wire paths and power-efficient CPUs tightly integrated with memory, communication, and I/O controllers. The tutorial describes the characteristics of a six thousand core cluster that puts all of these elements on a single chip, dramatically reducing cost and power consumption while increasing reliability and performance compared to commodity clusters.