high performance computing on graphics processing units: hgpu.org

In the last few years, Graphics Processing Units (GPUs) have become a great tool for massively parallel computing. GPUs are specifically designed for throughput and face several design challenges, specially what is known as the Power and Memory Walls. In these devices, available resources should be used to enhance performance and throughput, as the performance per watt is really high. For massively parallel applications or kernels, using the available silicon resources for power management was unproductive, as the main objective of the unit was to execute the kernel as fast as possible. However, not all the applications that are being currently ported to GPUs can make use of all the available resources, either due to data dependencies, bandwidth requirements, legacy software on new hardware, etc, reducing the performance per watt. This new scenario requires new designs and optimizations to make these GPGPU’s more energy efficient. But first comes first, we should begin by analyzing the applications we are running on these processors looking for bottlenecks and opportunities to optimize for energy efficiency. In this paper we analyze some kernels taken from the CUDA SDK in order to discover resource underutilization. Results show that this underutilization is present, and resource optimization can increase the energy efficiency of GPU-based computation. We then discuss different strategies and proposals to increase energy efficiency in future GPU designs.