Shubham Lonkar Defends Master’s Thesis on Cache Compression

As machine learning models grow, moving and storing data efficiently is becoming just as important as performing the computation itself. Shubham Lonkar explored this challenge in his master’s thesis, “Dynamic Cache Compression: Revisiting Compression for Machine Learning Workloads.”

His research asks whether processors can make better use of the cache capacity they already have instead of relying on larger physical caches. The key idea is to compress data stored in the cache, allowing more information to fit within the same hardware footprint.

Rather than relying on a single compression technique, Shubham developed an approach that dynamically selects a compression scheme for each cache line based on the data it contains. His work also introduces a hybrid compression method designed to handle mixed data patterns, including cases where conventional techniques struggle with pointer values.

Through this work, Shubham shows how adaptive compression can improve effective cache capacity for machine learning workloads while reducing pressure on the broader memory system. The thesis points toward more memory-efficient processor designs that can better support increasingly data-intensive applications.

Congratulations to Shubham on successfully defending his master’s thesis and completing this important milestone!