摘要
当前图形处理器的通用计算取得长足发展,为适应通用计算图形处理器在硬件体系结构和软件支持方面完成相应调整和改变,面对各种应用领域中数据规模增大的趋势,多GPU系统和GPU集群的研究应用日趋增多.以流处理器及图形处理器硬件体系为依据,介绍学术和工业领域中流处理器及图形处理器体系变化趋势.从软件编程环境、硬件计算与通信等方面展开讨论,阐述通用计算中图形处理器的关键问题,包括编程模型及语言的发展和方向,存储模型的量化研究、访存模式和行为的优化以及分布式存储管理的热点问题,典型通信原型系统的对比及通信难点的分析,GPU片内和片间的负载均衡,可靠性和容错计算,GPU功耗评测及低功耗优化的研究进展.综述在海量数据处理、智能计算、复杂网络、集群应用领域中图形处理器的研究进展及成果.总结在通用计算发展中存在的技术问题和未来挑战.
The General-purpose computing on graphics processing unit has been developing rapidly in recent years.To further improve the General-purpose computing capacity,the graphics processing units have evolved both in the hardware architecture and software support.Aiming at the trend of large-scale data processing emergence in the various application fields,the research about Multi-GPUs system or GPU Clusters also become an urgent research problem.The trend of stream processor and GPUs architecture in the academic and industrial fields is introduced based on the architecture design.The state of the art of key issues in GPGPU are summarized from the programming environment,computing and communication perspectives.It includes the development and the trend of the programming model and the programming languages,the memory model,the accessing patterns and behaviors analysis,the hot issues in the distributed memory management,the comparison and analysis of the existing communication prototype systems,the workload balance on the chip and out of the chip,the reliability model and the fault tolerance,the power consumption measurement and optimization.The currently development and research results of the applications of GPUs are discussed.These application fields are the large-scale data processing,intelligent computing,complex networks and GPUs clusters.Finally,the survey proposes the difficult problems in GPGPU and the new challenges in future.
出处
《计算机学报》
EI
CSCD
北大核心
2013年第4期757-772,共16页
Chinese Journal of Computers
基金
国家自然科学基金(60970012)
上海市科委重点攻关项目(09511501000
09220502800)
上海市重点学科建设项目(XTKX2012)资助~~
关键词
图形处理器
通用计算
可编程性
GPU集群
graphics processing unit
general-purpose computing
programmability
GPU clusters
