Explore a 15-minute conference talk from USENIX NSDI '23 that delves into optimizing GPU energy consumption for deep neural network (DNN) training. Learn about Zeus, an innovative optimization framework designed to balance the tradeoff between energy efficiency and performance in recurring DNN training jobs. Discover how common practices aimed at improving training performance can lead to inefficient energy usage, and understand the importance of considering energy consumption in addition to speed. Gain insights into Zeus' online exploration-exploitation approach and just-in-time energy profiling techniques, which eliminate the need for costly offline measurements and adapt to data drifts over time. Examine the impressive results of Zeus, which demonstrate improvements in energy efficiency ranging from 15.3% to 75.8% across diverse workloads. Presented by researchers from the University of Michigan, this talk offers valuable knowledge for those interested in sustainable and efficient deep learning practices.
Zeus - Understanding and Optimizing GPU Energy Consumption of DNN Training
Overview
Syllabus
NSDI '23 - Zeus: Understanding and Optimizing GPU Energy Consumption of DNN Training
Taught by
USENIX
Related Courses
-
ModelKeeper - Accelerating DNN Training via Automated Training Warmup
-
EnvPipe - Performance-Preserving DNN Training Framework for Saving Energy
-
Bamboo - Making Preemptible Instances Resilient for Affordable Training of Large DNNs
-
TopoOpt - Co-optimizing Network Topology and Parallelization Strategy for Distributed Training Jobs
-
BGL - GPU-Efficient GNN Training by Optimizing Graph Data I/O and Preprocessing
-
Accelerating Collective Communication in Data Parallel Training across Deep Learning Frameworks
