Session

DescriptionProcessing-in-memory (PIM) greatly improves performance and efficiency, but only when coupled with thorough considerations from the perspective of the entire system. The first two papers suggest solutions to accelerate recommendation models using PIM hardware. The third paper presents a near-data computing method to accelerate Mixture-of-Experts LLM inference, followed by the fourth paper which introduces a framework for efficiently realizing bulk bitwise operations in NVMs. The last two papers discuss SRAM-NVM hybrid designs, each aimed at accelerating transformer models and on-device learning.