high performance computing on graphics processing units: hgpu.org

As the clock frequency of silicon chips is leveling off, the computer architecture community is looking for different solutions to continue application performance scaling. One such solution is the multicore approach, i.e., using multiple simple cores that enable higher performance than wide superscalar processors, provided that the workload can exploit the parallelism. Another emerging alternative is the use of customized designs (accelerators) at different levels within the system. These are specialized functional units integrated with the core, specialized cores, attached processors, or attached appliances. The design tradeoff is quite compelling because current processor chips have billions of transistors, but they cannot all be activated or switched at the same time at high frequencies. Specialized designs provide increased power efficiency but cannot be used as general-purpose compute engines. Therefore, architects trade area for power efficiency by placing in the design additional units that are known to be active at different times. The resulting system is a heterogeneous architecture, with the potential of specialized execution that accelerates different workloads. While designing and building such hardware systems is attractive, writing and porting software to a heterogeneous platform is even more challenging than parallelism for homogeneous multicore systems. In this paper, we propose a taxonomy that allows us to define classes of accelerators, with the goal of focusing on a small set of programming models for accelerators. We discuss several types of currently popular accelerators and identify challenges to exploiting such accelerators in current software stacks. This paper serves as a guide for both hardware designers by providing them with a view on how software best exploits specialization and software programmers by focusing research efforts to address parallelism and heterogeneity.