Quantitative design space exploration of routing-switches for Network-on-Chip

Abstract. Future Systems-on-Chip (SoC) will consist of many embedded functional units like e.g. embedded processor cores, memories or FPGA like structures. These SoCs will have huge communication demands, which can not be fulfilled by bus-based communication systems. Possible solutions to this problem are so called Networks-on-Chip (NoC).

These NoCs basically consist of network-interfaces which integrate functional units into the NoC and routing-switches which connect the network-interfaces. Here, VLSI-based routing-switch implementations are presented. The characteristics of these NoCs like performance and costs (e.g. silicon area respectively logic elements, power dissipation) depend on a variety of parameters. As a routing-switch is a key component of a NoC, the costs and performance of routing-switches are compared for different parameter combinations. Evaluated parameters are for example data word length, architecture of the routing-switch (parallel vs. centralized implementation) and routing-algorithm.

The performance and costs of routing-switches were evaluated using an FPGA-based NoC-emulator. In addition different routing-switches were implemented using a 90 nm standard-cell library to determine the maximum clock frequency, power-dissipation and area of a VLSI-implementation. The power consumption was determined by simulating the extracted layout of the routing-switches. Finally, these results are benchmarked to other routing-switch implementations like Aetheral and xpipes.