ESE & CIS Seminar: “Toward Power-Efficient Computing with Applied Inference”

Abstract: We require computers that extract value from exponential data growth even as exponential transistor scaling falters. With limited scaling, power and thermal density threaten performance. Architects must pursue power efficiency in design and management. My research group addresses these challenges by integrating systems architecture with statistical inference. First, we design systems by tailoring hardware to software. We construct statistical surrogates for expensive design tools, solving previously intractable problems in heterogeneous, adaptive, and specialized processing. Second, we manage systems by allocating datacenter power. We construct multi-agent games and discover statistical equilibria in which users draw power strategically and avoid oversubscribing shared power. Third, we diagnose system anomalies by inferring causal performance models and extracting themes from these models with semantic clustering. Finally, we will discuss future directions for efficient computing in datacenters and beyond.