A Survey of Homogeneous and Heterogeneous System Architectures in High Performance Computing

Abstract

The TOP500 and GREEN500 lists are two major resources to understand and forecast the future architecture design of high performance computing platform. Generally, supercomputer system design can be divided to two parts: single computing node and interconnection. Regardless interconnection, we categorize the systems into two types: homogeneous and heterogeneous, based on single node architecture. While general-purpose graphic processing unit (GPGPU) is introduced to parallel computing system, heterogeneous computing model is getting a more attractive solution to extreme scales system. Now about 25% of Top 100 supercomputers is designed by heterogeneous model. In this comparative study, we will focus on the Top 100 and Top 500 supercomputers and analyze the data to November 2015. Linpack and power efficiency will be the two main measurements to evaluate those systems. We also will highlight the latest progresses and point out the possible trends.