Research

Memory Systems & DRAM Architecture

Recently, to process explosively increasing data, industry companies have introduced large-scale computer systems performing data processing, which leads to a change from the traditional Computing-Centric computer system to the Data-Centric computer system. In other words, computer systems are evolving to optimize and transmission of large-scale data as well as the processing of conventional arithmetic and logical operations. In this situation, to make a better computer system, it is essential to improve the performance and energy efficiency of the memory system where data is stored while the program is running. In particular, in the case of the latest artificial intelligence and big data applications, the amount of data handled is explosively increasing, so the memory bandwidth requirement is also increasing significantly. To satisfy these requirements, our research group conducts research to 1) optimize the overall memory hierarchy (e.g., cache, main memory) based on computer architecture knowledge and 2) improve the structure of DRAM (e.g., DDR5, HBM2), the main memory device.

Accelerator Architecture for AI Applications

Traditionally, most computer systems have been developed for general use. However, since the execution of specific applications or algorithms occupies a significant portion of the total computation in modern data centers, there have been various studies on computer architectures for accelerating these applications. To architect such an accelerator, it is necessary to understand how the application operates in the computer system, which resources it requires, and which resources cause the performance bottleneck. Therefore, our research group conducts profiling of the behavior of various emerging applications (i.e., AI applications) in the latest CPU or GPU. Then, we conduct research to propose a new computer architecture (e.g., Processing-in-Memory (PIM)) that can accelerate the applications.