이진호 연구실

With no doubt the most popular accelerator for AI nowadays is GPU. However the world is heading towards the next step: AI-specific accelerators. There is much room to improve in terms of accelerator designs. For example, optimizing dataflow, utilizing sparse network structure, or processing-in-memory techniques.

To utilize multiple devices (i.e., GPUs) for high-speed DNN training, it's common to employ distributed learning. There are still many ways to improve current distributed learning methods: Devising a new communication algorithm, smartly pipelining the jobs, or changing the ways that devices synchronize.

Efficient deployment and training of deep neural networks require reducing both computation and memory overhead. Our research focuses on advanced quantization and compression techniques to accelerate both inference and training across a wide range of models, including large language models (LLMs), diffusion-based image generation, graph neural networks, and spiking neural networks.

We focus on optimizing and accelerating diverse programs and applications across emerging commercial hardwares including GPUs, CXL, and PIM (Processing-In-Memory). Our research spans multiple domains: communication libraries, graph processing algorithms, vector databases, and scientific applications.