Layouts for sparse networks supporting throughput computing

We consider underpopulated networks sup- porting high-throughput computing in the NOCs con- text. In dense networks, all nodes can be sources and sinks of communication. Consequently realistic dense constant degree networks have insufficient communi- cation bandwidth, which needs to be compensated either by communication locality or by limiting the communication frequency of nodes. In underpopulated networks, only a portion of nodes are sources and sinks whereas the rest are simple intermediate nodes providing the communication bandwidth. In this paper, we compare scalable underpopulated networks capable to achieve a constant, sustained communication fre- quency, independent of the network size. We compare butterfly and mesh based underpopu- lated networks by presentings layouts for them and cal- culating characteristic cost properties of those layouts. Our study on these properties favor certain regular mesh based underpopulated networks.