Brandon Tran Preliminary Exam

Title: 3M: Model, Mediate, and Mitigate: Optimizing GPU Energy Consumption

Committee: Matthew Sinclair (Advisor), Shivaram Venkataraman (Advisor), Michael Swift, Ming Liu, Barton Miller, Woong Shin

Abstract:

With the ever-growing demand for high-performance computing systems, it has become the norm for HPC and ML applications to leverage Graphics Processing Units (GPUs). However, as GPU power demands have reached the limits of facility infrastructure, developers are operating within increasingly constrained energy budgets. With traditional hardware and facility scaling reaching their physical limits, application-level optimization has become the primary lever for further gains. While significant strides have been made in infrastructure and system software, the application layer still has substantial efficiency headroom-yet this headroom is inaccessible without transparency. Ultimately, optimizing applications for energy efficiency requires a granular understanding of power consumption. Unfortunately, current vendor-provided measurement profilers are limited to aggregate board-level power and provide no visibility into GPU energy consumption. To bridge the gap between these power readings and actionable application optimizations, developers require a comprehensive framework that provides fine-grained attribution for compute, quantifies the energy costs of distributed communication, and maps these low-level insights directly back to high-level source code.