Session
Memories Have a Mind of Their Own
Session Chairs
DescriptionComputing-in-memory (CIM) continues to advance energy efficiency for deep learning with breakthroughs in architecture, circuitry, and devices. The first paper introduces a tri-gear heterogeneous digital CIM for flexible data reuse in Diffusion models. The second paper unveils a fine-grained digital CIM with hessian trace-based quantization and approximate computing. The third paper presents a hybrid-domain SRAM CIM macro, harmonizing accuracy and energy efficiency through digital-analog computing synergy. The next two papers harness RRAM and IGZO in CIM applications. The final paper presents a drop-in-replacement design allowing fine-grained interleaving of processing-in-memory accesses and host accesses.
Event TypeResearch Manuscript
TimeWednesday, June 2610:30am - 12:00pm PDT
Location3002, 3rd Floor
Design
In-memory and Near-memory Computing Architectures, Applications and Systems
Presentations