A Review on Statistical Power Modelling for a Graphics Processing Unit (GPU)

The proliferation of portable applications has become a driving force for low power design making it crucial in the development of new processor architectures. Accompanying this is a scaling down of processor nodes which has caused many small channel effects to become more prominent eventually leading to a slow-down in CPU scaling. Smaller nodes in theory could provide better performance at higher frequencies but the apparent slow down has led to a saturation in performance and clock frequencies. This has led to the adoption of heterogeneous computing as a sustainable alternative in computing environments. Graphics Processor Units have gained popularity as a powerful “CPU Co-processor” by reducing the immense workloads on these Processing Units in the computing environments. However, such systems currently lack an effective methodology for power and performance modelling for design optimization. This research study presents a review on the basics of a Graphics Processing unit and the requirement of modelling for power and performance using statistical techniques that help optimize its design to obtain better perf-per-watt results. Methodologies used to obtain prioritized features that affect the power consumed and remove any correlations in data to prevent skewing are also discussed to build an effective Power Model.