TL;DR
This paper presents a comprehensive measurement study on how Dynamic Voltage and Frequency Scaling (DVFS) impacts the energy consumption of GPUs. The authors found that GPU DVFS is an effective energy-saving technique, particularly by scaling down core voltage, which can significantly reduce system power with minimal performance loss.
What Problem Does This Paper Address?
Modern computing systems, especially those employing GPUs for high-performance tasks, consume substantial amounts of power. While DVFS is a well-established method for conserving energy in CPUs, its effectiveness and nuances on GPUs were not thoroughly understood. This paper aims to fill that gap by investigating how adjusting GPU core voltage, core frequency, and memory frequency affects overall system energy consumption across a wide range of applications. The motivation is to understand how to make these powerful computing units more energy-efficient without sacrificing too much performance.
Methods
The researchers conducted extensive experiments on a real GPU platform (MSI N560GTX-Ti Hawk) using 37 benchmark applications. They employed a combination of software tools (NVIDIA Inspector, Afterburner) to adjust GPU core voltage, core frequency, and memory frequency within their allowable ranges. System power consumption was measured using a commercial power meter, and execution times were recorded. They analyzed the impact of these adjustments on both system idle power and runtime energy consumption. To quantify energy savings, they used metrics like $R$ (energy saved compared to default) and $R_{max}$ (maximum energy saving capability). The study systematically varied one or two parameters at a time to isolate their effects.
Datasets
The study utilized 37 benchmark applications. These were drawn from the NVIDIA CUDA SDK 4.1 and the Rodinia benchmark suite. The authors modified the source code of these applications to ensure they ran for a sufficient duration (over half a minute) to allow for accurate measurements, particularly considering the "RLC effect" (likely referring to runtime-dependent characteristics).
Key Insights and Ideas
- GPU DVFS is a viable energy-saving strategy: Contrary to potential assumptions that GPUs are too performance-critical for aggressive DVFS, the study demonstrates significant energy savings are achievable.
- Core voltage scaling is king for energy savings: Lowering the GPU core voltage proved to be the most effective method for reducing both idle and runtime system energy consumption.
- Frequency scaling is application-dependent: While core voltage scaling generally led to energy savings, the impact of core and memory frequency scaling varied significantly based on the application's characteristics (e.g., computation-intensive vs. memory-intensive).
- Performance trade-offs are manageable: The study highlights that substantial energy savings can be achieved with only a minor performance decrease, suggesting a favorable energy-performance trade-off.
- Idle power is not negligible: Even at idle, scaling down GPU core voltage contributed to a noticeable reduction in overall system power consumption.
Key Results
- Average Energy Reduction: The study achieved an average of 19.28% system energy reduction compared to the default settings, with an average performance decrease of only 4%.
- Core Voltage Impact: Scaling down the core voltage was highly effective. At the lowest tested core voltage (0.849 V) and a reduced core frequency (880 MHz), an average of 18.91% energy was saved with a 3.45% performance decrease.
- Memory Frequency Impact: Memory frequency scaling offered an average energy saving of 10.20% ($R_{max}$), but its effectiveness was highly application-specific. Computation-intensive kernels benefited from lower memory frequencies, while memory-intensive ones saw benefits from higher frequencies.
- Best Case Savings: In the best-case scenario (CFD application), up to 32.62% of energy could be reduced through core voltage scaling. For memory frequency scaling, SobolQRNG showed up to 28.65% energy reduction.
- Idle Power Savings: Scaling down the idle GPU core voltage from 1.049 V to 0.849 V reduced system idle power by approximately 8%.
Limitations and Future Work
- Limitations: The study focused on a single GPU model (NVIDIA GeForce GTX 560 Ti) and did not explore CPU DVFS in conjunction with GPU DVFS. The optimal settings for memory frequency were found to be highly application-specific, indicating that a one-size-fits-all approach is not feasible.
- Future Work: The authors propose developing a statistical power model that incorporates GPU DVFS. This model, combined with an execution time prediction model, could enable analytical determination of optimal GPU DVFS settings to minimize system energy consumption.
Summary updated