Throughput and latency analysis of the Spidergon-Donut interconnection network

The Spidergon-Donut on-chip interconnection network was proposed to interconnect the cores of a 1000+ core chip. In this paper, we derive a queueing model for the Spidergon-Donut, and analyze its throughput and latency in comparison with the commercial Spidergon network on which it is based. Results indicate that the Spidergon-Donut throughput is 3-5 times larger than that of the Spidergon network, for 32-4096 cores and the Spidergon-Donut significantly outperforms the Spidergon in network latency under moderate to large traffic rates. Furthermore, for 1K cores, the average distance of the Spidergon-Donut is one tenth that of the Spidergon.