MINDS Seminar Series | Oh-Kyoung Kwon (KISTI) - INTRODUCTION OF GPU COMPUTING : Definition, Evolution, Architecture, and Programming

MINDS SEMINAR

period : 2024-11-05 ~ 2024-11-05

time : 15:30:00 ~ 16:15:00

개최 장소 : Math Bldg 208 & Online streaming (Zoom)

Topic : INTRODUCTION OF GPU COMPUTING : Definition, Evolution, Architecture, and Programming

개요

Date	2024-11-05 ~ 2024-11-05	Time	15:30:00 ~ 16:15:00
Speaker	Oh-Kyoung Kwon	Affiliation	KISTI
Place	Math Bldg 208 & Online streaming (Zoom)	Streaming link	ID : 688 896 1076 / PW : 54321
Topic	INTRODUCTION OF GPU COMPUTING : Definition, Evolution, Architecture, and Programming
Contents	GPU computing, or Graphics Processing Unit computing, refers to the use of graphics processing units for general-purpose computing tasks, extending their original function of rendering graphics. This presentation explores the multifaceted domain of GPU computing, defining its role and highlighting its recent surge in importance, not only in High-Performance Computing (HPC) but also in Artificial Intelligence (AI) computations. We briefly delve into the historical backdrop of GPU computing and compare GPU architecture to CPU architecture. In the GPU architecture discussion, we emphasize key aspects such as simplified core functions, a multitude of cores, efficient context switching, the importance of registries, and high memory bandwidth. Additionally, we provide concise explanations of popular GPU programming techniques like CUDA[1], OpenACC[2], and HIP[3]. GPU computing traces its roots to the late 1990s when GPUs began to exhibit significant parallel processing capabilities. This evolution culminated in the introduction of CUDA by NVIDIA in 2007, revolutionizing the GPU computing landscape. In GPU architecture, core functionality is streamlined, and the number of cores is substantially higher than in CPUs. This architectural divergence enables efficient parallel processing. Key architectural features are as follows: Firstly, GPUs feature thousands of simpler cores optimized for parallel tasks, facilitating massive parallelism. Secondly, swift context switching is vital for GPU efficiency, and numerous registries support this. Thirdly, GPUs employ a multitude of registers for data storage, enabling high-throughput data processing. Lastly, GPUs depend on high-speed memory access to sustain their voracious computational appetite. By the end of this presentation, you will have gained a comprehensive understanding of GPU computing, its historical context, architectural distinctions compared to CPUs, and insights into popular GPU programming techniques that enable its wide-ranging applications in modern computing.

MinDS · 2024-10-28 16:43 · Views 630

List