Bilkent University
Department of Computer Engineering




Pouria Hasani
MS Student
(Supervisor: Prof. Dr. Özcan Öztürk)
Computer Engineering Department
Bilkent University

By nature, machine learning applications run the same instructions on huge amounts of data, which makes GPUs a suitable hardware infrastructure to be run on. However, the performance per watt metrics has become a critical evaluation concern than speed up. One of the most effective ways to reduce GPUs' power consumption is dynamic voltage and frequency scaling (DVFS). The power consumption is directly related to the voltage and frequency applied to the hardware. Using robust tools to adjust voltage and frequency based on current workload dynamically can drastically save energy usage. Although this technique has been successfully applied on CPUs, its impact on GPUs is still under exploring. This is mainly due to the complicated memory system and control flow. In this work, we are trying to use machine learning models to find the most optimal voltage and frequency adjustment while running massive parallel machine learning applications. For this purpose, we are using the GPGPU-SIM simulator to find the hardware performance counters and get more control over GPUs' architectures. The input of the model will be hardware performance counters which are collected while running the application. To collect such data, we will be using nvprof profiling tool. The profiling's overhead will be neglectable because the running applications are machine learning models that run the same instruction over thousands of epochs. So, the controlling models will collect the data, find the best adjustment, and apply it over each epoch.


DATE: 05 April 2021, Monday @ 16:00