Two-by-two switch element enables formation of rings and other nocs.
Switch element is registered: hence network can span the chip.
Network needs to carry decoupled requests and response packets.
Local arbitration in each element. Global policies required to avoid deadlock and starvation.
Give priority to traffic already on the ring: LAN-like buffering at source.
Does not carry interrupts or other sideband signals.
Single ring: throughput=2. dual counter-rotating rings: throughput=4.
IBM Cell Broadband Engine uses dual rings. Fat ring better than thin x-bar ?
A two-level heirarchy of bridged rings is sometimes a sweetspot.