On-chip interconnection architecture optimization using a multicommodity flow approach

Recent technology advent makes the efficient on-chip interconnection architecture critical in modern IC design. First, on-chip communication requirements are dramatically increasing due to the continuously shrinking of the device feature size and integration of billions of transistors on a single chip. Second, the interconnects, rather than devices, become the dominant factors in deciding performance and power consumption of VLSI systems. As a result, we are urged to optimize interconnection architecture to improve the overall system performance. In this dissertation, we propose a methodology to optimize the power consumption or communication latency of two promising on-chip interconnection architectures, Network-on-Chip (NoC) and FPGA global routing architecture. Our methodology adopts two optimization schemes, topology optimization and wire style optimization, and uses multicommodity flow (MCF) models to perform and evaluate these optimizations. We implement and optimize the MCF solver using polynomial approximation algorithms, which are significantly faster than the commercial linear programming solver CPLEX. For NoC optimization, our objective is to search for most power efficient NoC topologies and their wire assignments, which at the same time satisfy all communication latency and bandwidth requirements. We use MCF formulations to model this problem, and build NoC topology library and its power and delay library as inputs to MCF formulations. The solution of MCF formulations indicates the estimated NoC power consumption under a certain NoC topology and wire style assignments, from which we can observe the best NoC optimization. The experimental results show that our methodology can effectively improve the NoC power consumption or communication latency. The results also indicate the importance of power and latency co-optimization in NoC design. For FPGA global routing architecture optimization, we study segmentation distribution, flexible track assignment and wire style optimization. The objectives are power consumption and switch area density. The design methodology is also based on MCF formulations. We examine our FGPA routing architecture optimization by a set of standard MCNC [57] benchmark circuits. The experimental results show that our optimized FPGA routing architectures achieve average up to 10% to 15% power savings and up to 20% switch area savings when compared to traditional FPGA architecture.