一种高效GPU存储系统体系架构设计  

An Efficient Memory System Structure Design of GPU

作　　者：卢俊[1] 颜哲[1] 田泽[1] 

机构地区：[1]中国航空计算技术研究所,陕西西安710068

出　　处：《计算机技术与发展》2015年第4期6-9,共4页Computer Technology and Development

基　　金：国家"十二五"微电子预研基金项目(51308010601;51308010710;51308010711)

摘　　要：图形处理技术被广泛应用于电影、视频、游戏以及动画的制作,而图形处理系统(GPU)的出现极大地减轻了CPU日益繁重的图形处理任务,使得其能更专注于通用控制。文中阐述了制约GPU性能提升的重要因素,指出提高带宽利用率是应对这一问题的关键措施。通过局部性原理的分析,提出了一种基于层次化架构的高效GPU存储系统的设计。文中介绍了4层结构的存储系统,并逐层说明了各自的功能和架构,评估了基于层次化存储架构的GPU在典型应用中的带宽。文中还描述了Cache以及显存管理等子模块的功能。通过仿真可知,该GPU存储系统能充分利用共享和复用等手段尽量减少外部存储器的访问次数,从而提高了带宽利用率。Graphic processing technique has been widely used in movie, video, game and cartoon making. GPU has greatly reduced the pressure of graphic processing,which was CPU＇ s job in the past. It introduces the key factor which limits improvement of GPU perform- ance, and also indicates that the bandwidth utilization is one of the critical resources to deal with that limits. According to analysis of prin- ciple of locality, attempt to explore an efficient GPU memory system based on layered architecture. This global memory hierarchy struc- ture has four layers of memory, and the function of each layer has been described. In this paper, present the bandwidth utilization ratio of typical GPU running scenarios, also introduce the architecture of the modules such as Cache and MMU. This GPU memory system can re- duce the frequency of accessing SDRAM by sharing cache on chip, so that the utilization ratio of bandwidth has been greatly improved.

关 键 词：图形处理系统 层次化存储 带宽 存储管理模块 

分 类 号：TP31[自动化与计算机技术—计算机软件与理论]