ALPHA: A Hybrid Topology for Memory-centric Network

Abstract

Memory system, as the basic constituent of the computing system, has a inevitable impact on the system performance. Traditional bus-based memory access system cannot fulfill the performance requirement of the future high-performance computing system due to its limitation on energy-efficiency. Memory-centric network (MCN) is considered as a promising candidate for future system interconnect. To improve the energy-efficiency of memory access system by utilizing MCN, the topology of MCN needs to be carefully designed, as interconnection links take a huge part of power consumption. In this letter, we propose a new memory network topology named ALPHA. ALPHA is a 2-D topology. As ALPHA employs different types of interconnection links in the X-dimension and the Y-dimension, respectively, it can maintain high throughput while using fewer links. Besides we redesign the switch of HMC for its implementation in ALPHA. We make a comparison between ALPHA and four popular topologies. The simulation results show that ALPHA greatly increases throughput and decreases latency.