Dynamic sampling rate: harnessing frame coherence in graphics applications for energy-efficient GPUs.

Area of Science:

• Computer Graphics
• Hardware Architecture
• Energy Efficiency

Background:

• Real-time rendering relies on GPU fragment shaders, which are computationally intensive and consume significant energy.
• Conventional GPUs sample triangles at a fixed rate per pixel, leading to redundant computations in areas with low visual variation.
• Temporal frame coherence means consecutive frames are similar, offering opportunities for optimization.

Purpose of the Study:

• To introduce Dynamic Sampling Rate (DSR), a novel hardware mechanism for graphics applications.
• To reduce redundancy and enhance energy efficiency in real-time rendering.
• To maintain image quality while optimizing computational load.

Main Methods:

• DSR analyzes spatial frequencies of rendered 3D scenes.
• It utilizes temporal coherence between consecutive frames to determine optimal sampling rates.
• A hardware mechanism dynamically adjusts the sampling rate per screen region to minimize redundant fragment processing.

Main Results:

• DSR effectively removes redundancy in color computations at the fragment level.
• Experimental results on a mobile GPU architecture show average speedups of 1.68x.
• Significant energy savings of 40% were achieved across various applications.

Conclusions:

• DSR offers a hardware-based solution to improve energy efficiency in real-time graphics.
• The dynamic adjustment of sampling rates based on scene content and temporal coherence is effective.
• DSR provides substantial performance and energy benefits without sacrificing visual fidelity.