Gemel - Model Merging for Memory-Efficient, Real-Time Video Analytics at the Edge

Explore a conference talk on Gemel, an innovative system for model merging in edge-based video analytics. Learn how this technique addresses the growing resource constraints in edge deployments by efficiently sharing layers between vision models, reducing memory usage and swapping delays. Discover how Gemel integrates into existing pipelines, leveraging observations about memory usage and inter-layer dependencies to identify optimal merging configurations. Understand the system's approach to altering edge inference schedules for maximizing benefits. Gain insights into experimental results demonstrating significant memory reduction and accuracy improvements compared to traditional time or space sharing methods.

NSDI '23 - Gemel: Model Merging for Memory-Efficient, Real-Time Video Analytics at the Edge