Computer Architecture Seminar (9/9 17:00 P210)
작성일 2019.09.06 11:19:52
Computer Architecture Seminar
Title: Resolving Critical GPU Performance Bottlenecks
Speaker: Dr. Yunho Oh (EPFL)
Date: 9/9 Monday
관심있는 학생들의 많은 참여 바랍니다.
Graphics processing units (GPUs) have become the architectural choice to achieve high throughput in general-purpose computing. Thread-level parallelism (TLP) in GPUs is implemented by concurrently executing a large number of threads. However, GPUs cannot often achieve the theoretical peak performance. We found that the critical performance bottlenecks on GPUs are 1) limited memory system performance and 2) limited thread scheduling resources and register file. In this talk, we will introduce the GPU execution model and two above performance bottlenecks on GPUs in detail. Then, we will introduce two solutions addressing these challenges. First solution is a new GPU architecture, called Linebacker, that overcomes the limited memory system performance by improving cache efficiency on GPUs. Second solution is another work, called FineReg, that schedules threads over the limits of scheduling resources and register file on GPUs.
Yunho Oh is currently a postdoctoral researcher at École Polytechnique Fédérale de Lausanne (EPFL). He completed his Ph.D. in the School of Electrical and Electronic Engineering at Yonsei University in August 2018. His research interests include high-performance GPU architecture, in-storage processing architecture, energy-efficient datacenter architectures, and database processing systems. From 2016 to 2017, Yunho worked as a visiting graduate scholar at the University of Southern California. From 2011 to 2014, he worked as a software engineer at Mobile Communications Business, Samsung Electronics.