Comparative Evaluation of FPGA and ASIC Implementations of Bufferless and Buffered Routing Algorithms for On-Chip Networks

Most existing packet-based on-chip networks assume routers have buffers to buffer packets at times of contention. Recently, deflection-based bufferless routing algorithms have been proposed as an alternative design to reduce the area, power, and complexity disadvantages associated with buffering in routers. While bufferless routing shows significant promise at an algorithmic level, these algorithms have not been shown to be efficiently implementable in practice. Neither were they extensively compared to existing buffered routing algorithms in realistic designs. This paper presents our comparative evaluation of and experiences with realistic FPGA and ASIC designs of state-of-the-art (1) virtual-channel buffered, (2) deflection-based bufferless, and (3) deflection-based buffered routing algorithms using two different network topologies and network sizes. We show that bufferless routing algorithms are implementable without significant complexity, and compare their performance, area, frequency, and power consumption to their buffered counterparts. Our results indicate that bufferless routing can lead to significant area (38%), power consumption (30%), and router cycle time (8%) reductions over the best buffered router implementation on 65nm ASIC design, while operating at higher frequency.