基于预采样的模块级热分析方法

摘要

对片上多核系统（MPSoC）进行高效结构级热分析是进行温敏布图规划与实时功耗温度管理研究的关键．由于需要预先使用 HotSpot 提取布图规划中功能模块之间的相关热阻参数，最新的结构级热分析算法 BloTAM 在温敏布图规划中的热分析效率并不理想．对于散热条件与内核尺寸确定的温敏布图规划，该文提出了一种预采样的结构级热分析算法 PS-BloTAM，它先使用 HotSpot 为采样模块阵列建立相关热阻预采样矩阵S（即相关热阻库），温敏布图规划产生一个方案后，PS-BloTAM 可以根据方案中模块的大小与位置直接使用S 计算出模块之间的相关热阻矩阵R ，然后使用 R 计算出不同工作模式下的模块温度．采用传统的设计库思想，PS-BloTAM 能够快速计算出不同布图规划方案所产生的热点温度．实验数据表明：（1）与 HotSpot 相比，PS-BloTAM 的平均误差与最大误差分别小于1．65％和6．64％，可以获得43倍的加速效果；（2）与 BloTAM 相比，PS-BloTAM 可以获得3．4倍的加速效果；（3）在温敏布图规划的应用中，PS-BloTAM 以小于1．24％误差的代价、获得了比 HotSpot 快近10倍的加速效果．%Efficient thermal analysis (TA)plays a key role in the temperature-aware floorplanning design and Dynamic Power and Temperature Management (DPTM)for Multi-Processor System-on-Chip (MPSoC).Due to the prerequisite of using HotSpot,a commercial TA tool,to extract the thermal resistance matrix,the latest architecture-level TA method BloTAM is inefficient to analyze temperature distributions for numerous floorplanning solutions.For temperature-aware floorplanning with defined die geometric size and unchanged heat dissipation system,this paper proposes a novel pre-sampling BloTAM,e.g.PS-BloTAM.This method first adopts HotSpot to buildS,the sampling thermal resistance matrix for sampling block arrays.Then according to size and position of modules in a given floorplanning solution,PS-BloTAM analytically computes the thermal resistance matrix R withS.Finally with R,the method can analytically compute temperature distributions for modules in different working modes.Adopting traditional idea of design library, PS-BloTAM is able to quickly give user temperature distributions of numerous floorplans. Experiments show that:(1)Comparing to HotSpot,PS-BloTAM can achieve 43X speedup with average error and maximum error less than 1.65% and 6.64%,what’s more,it can provide speedup as high as 43 times;(2 )Comparing to BloTAM,PS-BloTAM can provide speedup as high as 3.4 times;(3)In temperature floorplanning,PS-BloTAM can provide nearly ten times speedup over HotSpot with the slight accuracy penalty of less than 1.24% errors.