Bio: Kirk W. Cameron is the Boeing Distinguished Professor of Computer Science and Managing Director of the Institute for Advanced Computing at Virginia Tech. He is an IEEE Fellow and ACM Distinguished Scientist. From 2012–2022, he directed the stack@cs Center for Computer Systems, which was ranked #26 nationally by U.S. News in March 2022. Cameron pioneered research in green computing, developing power measurement and management techniques that have influenced the design and operation of computers, supercomputers, and data centers. He is a co-founder of the Green500 List and a founding contributor to SPECpower, which is used in the ENERGY STAR program for servers. He also founded an energy measurement and management startup whose software was used by more than 500,000 people in over 160 countries. Cameron’s work has been featured in The New York Times, The Guardian, Time, Nature, Newsweek, and other outlets. His research and educational artifacts have been exhibited nationally and internationally, including at the Consumer Electronics Show, South by Southwest, and the Smithsonian’s National Museum of American History. He serves as Associate Editor-in-Chief of IEEE Transactions on Parallel and Distributed Systems.
Title: Power-Aware Computing in the Age of AI
Abstract: Some consider Artificial Intelligence (AI) to be the most disruptive technology in history. While the Internet took about 15 years to reach a billion users, AI surpassed a billion users in about 3 years. The city of San Francisco uses about 18.6 terawatt-hours (TWh) or about 18.6 billion kilowatt-hours of energy annually. U.S. datacenters that support AI and e-commerce use more than 300 TWh of energy — more than 5% of all U.S. electricity — and that share is growing. Driven by anticipated AI growth, datacenters will likely exceed 10% of all U.S. electricity use by 2030.
Although the current projections are roughly two orders of magnitude larger than those we considered in 2001, the trajectory closely tracks the hypotheses we formulated for HPC data centers at the time. We created the PowerPack Toolkit to measure, analyze, and correlate energy use by device and code function to identify opportunities for energy savings while maintaining performance. Additionally, we identified fundamental concepts (e.g., power-aware speedup) and created predictors (e.g., the Computer-Overlap-Stall Model) to intelligently manage computing resources at scale. In this talk, we will discuss these technologies and our early attempts to apply them to understand and ultimately improve the efficiency of present and future AI applications.