Memory-Centric Computing

Moving computation closer to where the data lives

Memory-Centric Computing

Overview

As the cost of data movement continues to increase relative to computation, data transfer has become a major performance and energy bottleneck in modern computing systems. We investigate memory-centric architectures that bring computation to where data resides, spanning Processing-in-Memory (PIM), near-data processing, and in-storage processing. Our research covers computation placement, memory–compute cooperation, and data movement optimization, with a particular focus on graph and sparse workloads whose irregular access patterns make them especially sensitive to memory bandwidth. We explore these designs through system-level architectural evaluation of data-intensive workloads.

Selected Publications

ISOCC
Reducing Data Transfer Overhead with CXL-Based Near-Data Processing for LLM Inference
Woo Hyun Kim, Han Sol Kim, Tae Hee Han
2025 22nd International SoC Design Conference (ISOCC), Oct. 15, 2025
#Memory-Centric #AI Accel #NDP #CXL #LLM Abstract
ISOCC
Efficient Graph Index Updates for ANNS with Near-Data Processing
Sung Yong Yoo, Hyeong Jin Kim, Tae Hee Han
2025 22nd International SoC Design Conference (ISOCC), Oct. 15, 2025
#Memory-Centric #NDP #Graph Abstract
IEEE Access
On-the-fly Lowering Engine: Offloading Data Layout Conversion for Convolutional Neural Networks
Min Gu Kang, Sang Min Hyun, Tae Hee Han, Seok In Hong, and Jung Rae Kim
IEEE Access, Vol. 10, pp. 79370-79746, Jul. 2022
#Memory-Centric #AI Accel #CNN #Dataflow #Offloading Abstract
IEEE Access
Task Parallelism-Aware Deep Neural Network Scheduling on Multiple Hybrid Memory Cube-based Processing-in-Memory
Young Sik Lee, Tae Hee Han
IEEE Access, Vol. 9, pp. 68561-68572, May 2021.
#Memory-Centric #AI Accel #PIM #DNN Scheduling Abstract