Jingwen's Homepage

Safe Limits on Voltage Reduction Efficiency in GPUs: a Direct Measurement Approach

Jingwen Leng, Alper Buyuktosunoglu, Ramon Bertran, Pradip Bose, Vijay Janapa Reddi

In Proc. of International Symposium on Microarchitecture (MICRO), 2015.

Energy efficiency of GPU architectures has emerged as an important aspect of computer system design. In this paper, we explore the energy benefits of reducing the GPU chip's voltage to the Vmin point at a fixed frequency, using a set of commercially available off-the-shelf GPU cards. In contrast to prior GPU-centric work, we adopt a direct measurement-based characterization methodology. The goal of this work is to build insightful observations about program-specific voltage guardband. We also explore voltage guardband reduction opportunities without compromising program correctness. We find that as much as 23% of voltage guardband can be reduced across the studied GPUs. We measure energy improvements to be as high as 25% without incurring any performance loss. The exact improvement opportunity depends on the available voltage guardband associated with the program. We quantify process, temperature and voltage variation (noise) contribution to the voltage guardband. We infer that the voltage noise has the highest impact on the guardband. We design a series of experiments and show that the intra-kernel activity within the program's execution causes the largest level of voltage noise. We characterize how program characteristics impact the Vmin behavior and show that we can use microarchitectural performance counters to predict Vmin value accurately. We demonstrate that we can achieve a large energy saving by operating the GPU at the predicted Vmin point.