high performance computing on graphics processing units: hgpu.org

It is common to exploit the co-processors of modern computer systems to speed up computations which were traditionally done on the CPU. While this is already very common for computer graphical and scientific applications, there is no reason why this cannot be extended to many different kinds of applications. In this paper we study the current state of general-purpose computing using accelerators, with an emphasis on the everyday user. We discuss several aspects of heterogeneous task scheduling, which becomes a concern when you have many different processors to execute a task on. We also show that there are several frameworks in development to support processor heterogeneity, but most of these are still unsuited for mass adoption due to their experimental or low-level nature. Besides conjecture we also did performance measurements on our everyday hardware, in order to find out if the promised performance increase is met. We conclude that this is indeed the case. We also take a look at power consumption and show that the fastest solution may not be the most energy-efficient one when heterogeneity is involved. Finally, we discuss the future work necessary to turn heterogeneous task scheduling into a mainstream programming paradigm.