JCSE, vol. 14, no. 3, pp.102-111, September, 2020

DOI: http://dx.doi.org/10.5626/JCSE.2020.14.3.102

GPGPU Functional Units Power Gating for Leakage Energy Reduction

Xin Wang and Wei Zhang
Department of Electrical and Computer Engineering, Virginia Commonwealth University, Richmond, VA, USA Department of Computer Science and Engineering, University of Louisville, Louisville, KY, USA

Abstract: The execution units of GPUs (graphics processing units) have been observed to produce many idle cycles that could be a tremendous waste of energy consumption which meanwhile provides a hint to build a more energy-efficient system to operate GPUs if idle cycles can be appropriately taken care of. However, power-gating without foresight can be dangerous since inaccurate decisions on power-gating will introduce unaffordable overhead on both energy consumption and performance. In this paper, we examine the length of execution units' idle cycles for several representative GPGPU applications and evaluate the distribution of the idleness durations. We then propose the energy-saving strategies with focus on discovering potential execution units' power-gating opportunities. The idle durations are recorded in the runtime for various computing units in streaming multiprocessors (SMs) including integer units and floating units in streaming processors (SPs) and special function units (SFUs). By analyzing the observed idleness, we propose to enhance the energy efficiency through two execution units' power-gating policies, the immediate power-gating (IPG) and idle detect power-gating (ID-PG). Furthermore, we examine the policies with various parameter settings to offer insights on possible gains and losses of the power-gating techniques. Besides, by noticing that integer units are the most popular computing units for many applications, we introduce the power-aware SP(s) to increase the throughput of integer instructions. It was observed that the power-aware SP can provide performance enhancement as well as the leakage energy reduction for several applications. The experimental results show that both the policies can result in satisfactory leakage energy saving on execution units. The IPG can reduce the execution unit's leakage energy by 84.3% when the break-even time is set to 5 cycles. Even if the break-even time goes up to 20 cycles, the ID-PG can save 67.1% of the total execution units' leakage energy. Moreover, involving power-aware SP(s) can improve the performance by up to 14.4% and 2.7% on average.

Keyword: GPGPUs; Energy-efficiency; Execution units; Power-gating

Full Paper:   136 Downloads, 1015 View