基于PCM的GPU存储系统设计与优化  

Evaluating and Optimizing of PCM Based GPU Memory Architecture

作　　者：穆帅[1] 单书畅[2] 邓仰东[1] 王志华[1] 

机构地区：[1]清华大学微电子所,北京100084 [2]中国科学院计算技术研究所,北京100084

出　　处：《计算机科学》2013年第10期29-31,71,共4页Computer Science

基　　金：国家自然科学基金(61272085)资助

摘　　要：以相变存储器(PCM)为代表的新型非易失存储器,具有存储密度高和静态功耗低等传统动态随机存取存储器(DRAM)不具备的优势,但是过长的写操作延时会严重影响访存的性能。设计了基于PCM的图形处理器(GPU)中的存储系统。仿真结果显示,GPU程序中的内存写请求分布极不均匀,对少量的内存地址有非常高的访问频率。面向访存分布不均匀特点的专用缓冲单元设计,能够有效地存储频繁访问的内存数据,从而减少对PCM的访问次数,消除过长的写操作延时对系统性能的负面影响。GPU仿真器上的结果显示,基于缓冲单元的PCM存储系统能够有效地提高GPU的运算性能。Recently, the emerging non-volatile memory（NVM） exemplified by Phase Change Memory（PCM） has been considered to take the place of conventional DRAM in the processors due to their characteristics of large capacity and little static power. However, the overhead of long write latency can cause severe performance degradation. We evaluated the feasibility of PCM based GPU memory architecture. Based on the analysis of unique memory access behaviors cap- tured from GPU benchmarks, a dedicated buffer was designed to alleviate the pressure of frequently accessing PCM. Simulation results prove the efficiency of our proposed dedicated buffer and show its great potential for PCM based GPU architecture.

关 键 词：相变存储器 图形处理器 缓冲单元 存储系统 

分 类 号：TP391.9[自动化与计算机技术—计算机应用技术]