[Paper Review] A Survey and Measurement Study of GPU DVFS on Energy Conservation

Xinxin Mei, Qiang Wang+1 more|arXiv (Cornell University)|Oct 6, 2016

Parallel Computing and Optimization Techniques48 references4 citations

TL;DR

This paper investigates GPU dynamic voltage and frequency scaling (DVFS) for energy conservation, conducting real-world experiments on NVIDIA Fermi and Maxwell GPUs. It finds that core voltage/frequency scaling significantly reduces energy consumption—averaging 18.91% savings on Fermi—and that memory frequency scaling has platform-dependent effects, with Maxwell benefiting from higher memory frequencies while Fermi shows mixed results, highlighting the need for application-specific DVFS tuning.

ABSTRACT

Energy efficiency has become one of the top design criteria for current computing systems. The dynamic voltage and frequency scaling (DVFS) has been widely adopted by laptop computers, servers, and mobile devices to conserve energy, while the GPU DVFS is still at a certain early age. This paper aims at exploring the impact of GPU DVFS on the application performance and power consumption, and furthermore, on energy conservation. We survey the state-of-the-art GPU DVFS characterizations, and then summarize recent research works on GPU power and performance models. We also conduct real GPU DVFS experiments on NVIDIA Fermi and Maxwell GPUs. According to our experimental results, GPU DVFS has significant potential for energy saving. The effect of scaling core voltage/frequency and memory voltage/frequency depends on not only the GPU architectures, but also the characteristic of GPU applications.

Motivation & Objective

• To analyze the impact of GPU DVFS on performance, power consumption, and energy efficiency in modern GPU architectures.
• To survey state-of-the-art GPU DVFS characterization and modeling techniques, focusing on recent NVIDIA Fermi and Maxwell platforms.
• To conduct empirical measurements of core and memory voltage/frequency scaling on real GPU hardware to establish benchmarks.
• To identify optimal voltage/frequency settings that minimize energy consumption while maintaining acceptable performance.
• To highlight the challenges in modeling GPU DVFS due to architectural differences and application-specific behaviors.

Proposed method

• Conducted real-world DVFS experiments on NVIDIA GTX 980 (Maxwell) and Tesla K20X (Fermi) GPUs to measure performance and power consumption across various voltage and frequency settings.
• Measured energy consumption and execution time across 34 kernels from standard benchmarks to evaluate DVFS impact on different workloads.
• Analyzed the relationship between frequency scaling and energy consumption, identifying non-linear behaviors such as 'pacing' and 'racing' effects.
• Evaluated both core and memory frequency scaling independently, comparing results across Fermi and Maxwell architectures.
• Used empirical metrics such as maximum energy increase (Rmax) and average energy change (R̂) to quantify scaling inefficiencies.
• Compared results against vendor default settings to determine optimal configurations for energy efficiency.

Experimental results

Research questions

• RQ1 How does dynamic voltage and frequency scaling (DVFS) affect energy consumption and performance on modern GPU architectures?
• RQ2 What are the differences in DVFS behavior between the Fermi and Maxwell GPU generations?
• RQ3 How do application characteristics influence the effectiveness of core and memory frequency scaling for energy conservation?
• RQ4 Is there a universal optimal frequency setting for all GPU applications, or does it vary by workload?
• RQ5 To what extent does scaling down memory frequency lead to energy waste due to increased execution time?

Key findings

• On the Fermi GPU, scaling down core voltage and frequency achieved an average energy saving of 18.91%, with a maximum of 24.4%.
• For the Maxwell GPU, core frequency scaling showed significant energy savings, with the lowest voltage/frequency settings yielding minimal energy consumption across all kernels.
• Memory frequency scaling on the Maxwell platform showed a 'racing' effect: higher memory frequencies consistently reduced energy consumption, with average Rmax at 35.87% and average R̂ at 0.72%.
• On the Fermi platform, memory frequency scaling had a less consistent effect, with average Rmax at 10.2% and average R̂ at 3.5%, indicating that default memory frequency often remains optimal.
• For 24 kernels on the Maxwell GPU, scaling down memory frequency by 30% caused energy increases exceeding 30%, with the worst-case energy waste reaching 53.9% (bfs kernel).
• The default memory frequency setting was optimal for 34 out of 34 kernels on the Maxwell platform, suggesting it is well-tuned for general workloads.