Dynamic voltage and frequency scaling (DVFS) is a popular technique to improve the energy-efficiency of high-performance computing systems. It allows placing the devices into lower performance states when the computational demands are lower, opening the possibility for significant power/energy savings. This work presents a GPU power consumption model, used to predict the GPU power consumption of any application at different frequency levels. To obtain this model, an estimation algorithm is proposed, relying on careful benchmarking of the GPU architecture. The model can estimate the contribution of twelve different GPU components (FP32-ADD/MUL/FMA, FP64-ADD/MUL/FMA, INT, SF, CF units, shared memory, L2-cache, and DRAM) to the GPU power consumption. Different model use cases are evaluated (fixed-frequency, DVFS, and scaling-factors), which can obtain both the total or the per-component GPU power consumption. A technique to export models to a distinct GPU from the one it was estimated on is also proposed. These approaches were extensively validated on five different GPUs from the three most recent microarchitectures with a set of 42 standard benchmarks, achieving very accurate predictions. In particular, the scaling-factor power model achieves an average prediction error of 3.5 percent (Titan Xp), 4.6 percent (GTX Titan X), 3.1 percent (GTX 980) and 2.4 percent (Tesla K40c).

Dynamic Voltage and Frequency Scaling (DVFS) on Graphics Processing Units (GPUs) components is one of the most promising power management strategies, due to its potential for significant power and energy savings. However, there is still a lack of simple and reliable models for the estimation of the GPU power consumption under a set of different voltage and frequency levels. Accordingly, a novel GPU power estimation model with both core and memory frequency scaling is herein proposed. This model combines information from both the GPU architecture and the executing GPU application and also takes into account the non-linear changes in the GPU voltage when the core and memory frequencies are scaled. The model parameters are estimated using a collection of 83 microbenchmarks carefully crafted to stress the main GPU components. Based on the hardware performance events gathered during the execution of GPU applications on a single frequency configuration, the proposed model allows to predict the power consumption of the application over a wide range of frequency configurations, as well as to decompose the contribution of different parts of the GPU pipeline to the overall power consumption. Validated on 3 GPU devices from the most recent NVIDIA microarchitectures (Pascal, Maxwell and Kepler), by using a collection of 26 standard benchmarks, the proposed model is able to achieve accurate results (7%, 6% and 12% mean absolute error) for the target GPUs (Titan Xp, GTX Titan X and Tesla K40c).

Prompted by their very high computational capabilities and memory bandwidth, Graphics Processing Units (GPUs) are already widely used to accelerate the execution of many scientific applications. However, programmers are still required to have a very detailed knowledge of the GPU internal architecture when tuning the kernels, in order to improve either performance or energy-efficiency. Moreover, different GPU devices have different characteristics, moving a kernel to a different GPU typically requires re-tuning the kernel execution, in order to efficiently exploit the underlying hardware. The procedure proposed in this work is based on real-time kernel profiling and GPU monitoring and it automatically tunes parameters from several concurrent kernels to maximize the performance or minimize the energy consumption. Experimental results on NVIDIA GPU devices with up to 4 concurrent kernels show that the proposed solution achieves near optimal configurations. Furthermore, significant energy savings can be achieved by using the proposed energy-efficiency auto-tuning procedure.