Evaluating the NVIDIA tegra processor as a low-power alternative for sparse GPU computations

Abstract

In the last years, the presence of heterogeneous hardware platforms in the HPC field increased enormously. One of the major reason for this evolution is the necessity to contemplate energy consumption restrictions. As an alternative for reducing the power consumption of large clusters, new systems that include unconventional devices have been proposed. In particular, it is now common to encounter energy-efficient hardware such as GPUs and low-power ARM processors as part of hardware platforms intended for scientific computing. A current line of our work aims to enhance the linear system solvers of ILUPACK by leveraging the combined computational power of GPUs and distributed memory platforms. One drawback of our solution is the limited level of parallelism offered by each sub-problem in the distributed version of ILUPACK, which is insufficient to exploit the conventional GPU architecture. This work is a first step towards exploiting the use of energy efficient hardware to compute the ILUPACK solvers. Specifically, we developed a tuned implementation of the SPD linear system solver of ILUPACK for the NVIDIA Jetson TX1 platform, and evaluated its performance in problems that are unable to fully leverage the capabilities of high end GPUs. The positive results obtained motivate us to move our solution to a cluster composed by this kind of devices in the near future.