STA: A Highly Scalable Low Latency Butterfly Fat Tree Based 3D NoC Design

Since the past decade Network-on-Chip has evolved as the most dominant and efficient solution in on-chip communication paradigm for multi-core systems. With the growing number of on-chip processing cores modern three dimensional NoC design is facing several challenges originating from various network performance parameters like latency, hop count etc. Scalability and network efficiency have generated an important trade off in 3D NoC design, which needs to be balanced, especially for application specific NoC design. Tree based topologies outperform mesh based topologies in terms of network latency and throughput with increasing injection rate of packets/flits. But on the other hand, floor planing becomes much more complex for tree based designs with increasing number of IP blocks compared to mesh due to the hierarchical structure. This paper introduces a novel 3D NoC architecture named Split Tree Architecture (STA), based on butterfly fat tree, which is highly scalable while maintaining low network latency and hop count significantly. There are latency improvements of 51-91%, 84-96%, 55-96%, and 48-96% over mesh, torus, butterfly, and flattened butterfly topologies respectively. Average hop count is improved by 44% and 30% over mesh and torus. Average and minimum acceptance rates are improved by 3-8% and 3-12% over torus and, 4-7% and 4-12% over flattened butterfly. In comparison to the previously reported state of the art 3D BFT based designs, STA achieves performance improvements of 19-78%, 2-42%, 0.2-0.6%, and around 20%, for average latency, average acceptance rate, minimum acceptance rate, and average hop count respectively.