云基础设施中的GPU服务器性能评估  

Performance Evaluation of GPU Server in Cloud Infrastructure

作　　者：邱红飞[1] 李先绪[1] 黄春光[1] QIU Hongfei;LI Xianxu;HUANG Chunguang(Guangzhou Research Institute of China Telecom Co.,Ltd.,Guangzhou 510630,China)

机构地区：[1]中国电信股份有限公司广州研究院,广东广州510630

出　　处：《移动通信》2018年第7期1-6,共6页Mobile Communications

摘　　要：基于GPU的基础设施云是未来人工智能应用与开放智能云服务的重要基础架构,是未来企业基础架构的技术高点和发展方向。因此对GPU相关技术、应用场景和NVIDIA GPU产品进行介绍,分析了GPU服务器性能衡量指标中存在的问题,提出了GPU服务器性能的测试方法,并验证了GPU服务器在深度学习领域的性能,从而解决了不同GPU服务器之间的性能对比难题。The infrastructure cloud based on GPU is the important infrastructure for future artifi cial intelligence applications and open intelligent cloud services,as well as the technical point and development direction of future enterprise infrastructure.The GPU related technology,application scenario and NVIDIA GPU product are summarized.The problems in the benchmark of GPU server performance is analyzed.The testing method of GPU server performance is presented.The performance of GPU server in deep learning is verifi ed to deal with the diffi culty in performance comparison for different GPU servers.

关 键 词：GPU 人工智能 深度学习 带宽 服务器性能 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]