An Efficient Partitioning Algorithm Based on Hypergraph for 3D Network-On-Chip Architecture Floorplanning

Abstract

Network-on-Chip (NoC) supplies a scalable and fast interconnect for the communication between the different IP cores in the System-on-Chip (SoC). With the growing complexity in consumer embedded systems, the emerging SoC architectures integrate more and more components for the different signal processing tasks. Two dimensional Network-on-Chip (2D NoC) becomes a bottleneck for the development of the SoC architecture because of its limitation on the area of chip and the long latency. In this case, SoC research is forcing on the exploration of three dimensions (3D) technology for developing the next generation of large SoC which integrates three dimensional Network-on-Chip (3D NoC) for the communication architecture. 3D design technology resolves the vertical inter-layer connection issue by Through-Silicon Vias (TSVs). However, TSVs occupy significant silicon area which limits the inter-layer links of the 3D NoC. Therefore, the task partitioning on 3D NoC must be judicious in large SoC design. In this paper, we propose an efficient layer-aware partitioning algorithm based on hypergraph (named ELAP-NoC) for the task partitioning with TSV minimization for 3D NoC architecture floorplanning. ELAP-NoC contains divergence stage and convergence stage. ELAP-NoC supplies firstly a multi-way min-cut partitioning to gradually divide a given design layer by layer in the divergence stage in order to get an initial solution, then this solution is refined in convergence stage. The experiments show that ELAP-NoC performs a better capacity in the partitioning of the different numbers of cores which supplies the first step for the 3D NoC floorplanning.