The past decade has witnessed a 300,000x increase in the amount of compute for AI. The rapid growth of AI computing comes with significant costs. First, the model size and computation requirement scaling outpace the AI system performance improvement. Second, the energy footprint of AI has massive impact in our environment and profound implications in our industry’s Green and Sustainability strategies. AI accelerators (training and inference) are excellent first steps to bend the growing use cases and system resource demand of AI. However, majority of efficiency opportunities in the AI domain lie above the hardware layers, in model and data efficiency. Advances in AI are currently driven by AI research that seeks to improve accuracy (or related measures) through the use of massive computational power while disregarding the cost. We need to develop AI research while considering their computational (and thus carbon) cost, encouraging a reduction in resources spent. This talk will discuss the research and deployment opportunities to advance sustainability adoption and describe ML system development strategies to scale AI computing sustainably for the next decades to come with the goal of reducing the rising environmental footprint of AI.

Ramya Raghavendra received her MS & PhD in Computer Science at the University of California, Santa Barbara. She spent the next decade working as a research scientist at IBM Research, during which time she worked on a range of problems that spanned networking, network analytics and large-scale systems for Machine Learning. She has co-authored over 25 papers in top venues including Sigcomm, Infocom, ASPLOS, ICDE, BigData and JMLR, granted over 40 patents and earned the title of ‘Master Inventor’ at IBM. Ramya currently works at Facebook AI where she works on problems at the intersection of systems and AI, leading the Green AI effort aimed at developing and deploying AI that is efficient-by-design.