GPPT: A Power Prediction Tool for CUDA Applications

Abstract

Graphics Processing Unit (GPU) is no longer a specialised equipment for visual processing and is now a day-to-day commodity for general-purpose computing. Due to this transition, it has become crucial to understand GPU's con-tribution to power consumption. If application developers are assisted with a tool which understands the power consumption of CUDA code and which does not involve executing the code; it can be an asset to make GPU a energy-aware computing alternative. We present here GPU Power Prediction Tool (GPPT), an eclipse plugin for assessing the power of CUDA applications based on static analysis of PTX code. GPPT utilizes a machine learning model which utilizes application features generated by dissecting PTX code with the help of hardware attributes and user inputs. GPPT is an architecture-agnostic tool which is tested for three architecture: Tesla, Maxwell, Volta. R2 score for GPPT using XGBoost technique is 0.93. Thus, we have developed an end-to-end fully automated architecture agnostic tool for power prediction of CUDA kernel with reasonable precision.