The continuing and expanding popularity of Artificial Intelligence (AI) is driving immense energy demands at a global scale. Sustainability and energy efficiency are therefore crucial to the long-term success of AI. In this talk we will explore a variety of opportunities to tune and optimize energy usage across the AI software stack, focusing on the efficient use of hardware accelerators (e.g. GPUs). Rather than looking to optimize the code itself, we consider further hardware parameters (e.g. power limits) and program parameters (e.g. thread block configurations, batch size) which affect how the code is executed. We discuss a static code analysis tool “FlipFlop” which predicts energy consumption and tunes the choice of thread block configuration and GPU power limit to improve both energy efficiency and performance. Forward-looking, we consider the relationships between model architecture and hyperparameters, execution framework, executing device, and energy consumption. We consider the task of predicting energy consumption for running a given model on a given executing device.

About Alex Brandt

Dr. Alexander Brandt is an Assistant Professor in the Faculty of Computer Science at Dalhousie University and a Principal Scientist in the Ontario Research Centre for Computer Algebra. After receiving a Bachelor of Science in Software Engineering from Memorial University of Newfoundland, he received his PhD in Computer Science from the University of Western Ontario in 2022. Dr. Brandt’s research intersects symbolic computing, scientific computing, software engineering, and high-performance computing. His work looks to design and develop high-performance algorithms and implementations supporting scientific, mathematical, and compute-intensive software through the effective and efficient use of computer hardware. Parallelism and heterogeneous computing are central pillars to his work. Dr. Brandt is also the head coach of the Dalhousie University rowing team.

Este evento se impartirá en inglés