图形处理器低功耗设计技术研究  

Low-power Design Techniques for GPU

作　　者：田泽[1] 张骏[1] 许宏杰[1] 郭亮[1] 黎小玉[1] 

机构地区：[1]中国航空工业西安航空计算技术研究所,西安710077

出　　处：《计算机科学》2013年第06A期210-216,共7页Computer Science

基　　金：2012总装预研基金(9140A08010712HK61095);中国航空工业集团公司创新基金(2010BD63111)资助

摘　　要：图形处理器(GPU)以其强大的图形加速性能以及在通用计算领域的出色表现正在被越来越广泛地应用。但随着芯片规模和集成度的不断提升,单个GPU芯片的功耗已经高达376W,是高端通用处理器的2～3倍。高功耗带来的可靠性、稳定性以及芯片成本问题使"功耗墙"已经成为未来GPU设计过程中需要突破的关键问题之一。立足于体系结构层次,结合图形处理器的渲染流水线的结构特点,从深度测试和消隐、染色器数据通路、纹理映射和压缩、渲染策略、寄存器文件和片上Cache等角度描述了图形处理器的低功耗设计技术,并指出了GPU低功耗设计技术的进一步研究方向。Graphic Processing Unit（GPU） is extensively used to accelerate the graphic processing and general purpose computing. But unfortunately, along with the increasing of chip scale and density, power consumption of single GPU reaches 376W, which is two or three times higher than a typical high-end general purpose CPU. Stability, reliability and cost problems brought by high power consumption had already made the ＂Power Wall＂ be a key obstacle which must be broken through. On architecture level,and combined with the structure characteristics of GPU rendering pipeline, this paper demonstrates the GPU low-power design techniques, such as depth test and elimination, shader data path, texture mapping and compression, rendering strategy, register file and Cache, and then indicates the further research content for GPU low-power design.

关 键 词：图形处理器 低功耗 渲染 CACHE 

分 类 号：TP303[自动化与计算机技术—计算机系统结构]