연구실

We design energy-efficient and high-performance deep learning hardware accelerators. We focus on making neural networks lighter to enable deep learning computation in energy-constrained embedded computing devices. We also invent various circuits with low-power and variation-aware characteristics.

Our main/current interests are

A. Efficient Neural Network
- Model Compression, including pruning and quantization
- Efficient Serving Systems for Generative Models
- Memory/Compute-Efficient Generative Models
- LLM Agent and Multi-Modal Systems
B. Neural Network-Friendly Hardware
- Variable-Precision Neural Network Hardware
- Quantization-Aware Neural Network Hardware
- Efficient Point Cloud Acceleration
- Hardware-Network Co-Design for LLM
C. In Memory Computing
- Digital Computing Logic near/in-memory
- Software/Hardware co-design for DRAM-PIM
- Logic and MAC Operations in Flash Memory
- Analog Compute-in-Memory (ACIM) Design
- Digital Compute-in-Memory (DCIM) Design
D. Technology-Circuit Co-design
- Technology-Circuit Co-Design for Emerging Memory
- Process Architecture Design for Next-Generation Technologies
- Modeling and Circuit Design for Emerging Technologies
- Circuit Design and Optimization for 3D Memory
- Standard Cell Design for Heterogeneous Integration
- Monolithic 3D Memory Design for Processing-in-Memory (PIM)

http://vlsi.snu.ac.kr/publications/