A traffic-aware memory-cube network using bypassing

Three-dimensional stack memory which provides both high-bandwidth access and large capacity is a promising technology for next-generation computer systems. While number of cubes increase the aggregate capacity, communication latency power consumption significantly owing to its low-radix large-diameter packet network. In this context, we propose memory-cube network called Diagonal Memory Network (DMN). A diagonal topology, floor layout, lightweight router were designed low-latency low-voltage memory-read communication. DMN routing efficiently avoids deadlocks packets, although it allows each transmitted processor use bypassing original datapaths. Our evaluation results show that decreases hardware resources by more than 31% compared with conventional virtual channel router. The reduces energy 13% 67% transit along datapath datapath, respectively. Furthermore, using flit-level discrete event simulation, topology achieves high throughput lower existing topologies routers.