Implementation and Evaluation of Large Interconnection Routers for Future Many-core Networks on Chip
Amir H.M. Zaytoun, Hossam A. H. Fahmy, Khaled Elsayed · 2012
As the number of processing elements in the future Networks on Chip (NoC) increases from multi-cores to many-cores, the role of the interconnection communications becomes more critical. The number of cores on a System on Chip (SoC) will reach thousands in the near future as predicted by the International Technology Roadmap for Semiconductors (ITRS). Currently, NoC interconnections are mostly implemented with m×n 2-D mesh topology connecting small size routers. This will represent the bottleneck to the communication latency for the increasing number of cores where the average number of hops the data have to pass will increase. In this paper, we propose an alternative NoC interconnecting scheme by using large routers interconnecting large number of cores in star topology. This interconnection scheme can be scaled up by using hierarchical-star or fat-tree topologies. We present the implementation and performance evaluation of three large router architectures and compare their efficiency to the small 5×5 router used in the mesh topology. We develop a simulating environment that resembles the real NoC conditions to test the routers throughput and average latency on different buffer sizes and under different traffic loads. We also synthesize them to estimate the area and power consumption. Then, routers efficiencies are calculated with respect to the area and power consumption.