An exploration of heterogeneous systems

Abstract

Heterogeneous computing represents a trendy way to achieve further scalability in the high-performance computing area. It aims to join different processing units in a networked-based system such that each task is preferably executed by the unit which is able to efficiently perform that task. Memory hierarchy, instruction set, control logic, and other properties may differ in processing units so as to be specialized for different variety of problems. However, it will be more time-consuming for computer engineers to understand, design, and program on these systems. On the other hand, proper problems running on well-chosen heterogeneous systems present higher performance and superior energy efficiency. Such balance of attributes seldom makes a heterogeneous system useful for other fields than embedded computing or high-performance computing. Among them, embedded computing is more area and energy efficient while high-performance computing obtains more performance. GPUs, FPGAs or the new Xeon Phi are example of common computational units that, along with CPUs, can compose heterogeneous systems aiming to accelerate the execution of programs. In this paper, we have explored these architectures in terms of energy efficiency, performance, and productivity.