面向机器学习的高性能SIMT处理器cache的设计与实现
|
摘要
为了满足机器学习中大数据、并行计算及降低处理器与主存之间的差距等要求,设计基于自主研发的SIMT处理器的流水线cache结构。依据局部性原理与LRU替换算法相结合设计专用的伪LRU替换算法,与通用的轮询、LFU、LRU替换算法共同完成cache替换算法的可配置要求,实现处理器与主存之间的快速交互。采用Xilinx公司virtex ultrascale系列的xcvu440-flga2892-2-e FPGA芯片对设计进行综合。结果表明该结构指令cache最大时延为2.923 ns,数据cache最大时延为3.258 ns,满足SIMT处理器性能要求。
In order to meet the requirements of big data, parallel computing and reduce the gap between processor and main memory in machine learning, we designed a pipeline cache structure based on self-developed SIMT processor. It was designed according to the principle of locality and LRU replacement algorithm. The dedicated pseudo LRU replacement algorithm, together with the general Round Robin, LFU, and LRU replacement algorithms, completed the configurable requirements of the cache replacement algorithm to achieve fast interaction between the processor and the main memory. The design was integrated with the xcvu440-flga2892-2-e FPGA chip of Xilinx virtex ultrascale series. The results show that the maximum latency of the structure instruction cache is 2.923 ns, and the maximum delay of the data cache is 3.258 ns, which satisfies the performance requirements of SIMT processor.
In order to meet the requirements of big data, parallel computing and reduce the gap between processor and main memory in machine learning, we designed a pipeline cache structure based on self-developed SIMT processor. It was designed according to the principle of locality and LRU replacement algorithm. The dedicated pseudo LRU replacement algorithm, together with the general Round Robin, LFU, and LRU replacement algorithms, completed the configurable requirements of the cache replacement algorithm to achieve fast interaction between the processor and the main memory. The design was integrated with the xcvu440-flga2892-2-e FPGA chip of Xilinx virtex ultrascale series. The results show that the maximum latency of the structure instruction cache is 2.923 ns, and the maximum delay of the data cache is 3.258 ns, which satisfies the performance requirements of SIMT processor.
引文
[1] 石弘一.机器学习综述[J].通讯世界,2018(10):253-254.
[2] Lindholm E,Nickolls J,Oberman S,et al.NVIDIA Tesla:A Unified Graphics and Computing Architecture[J].IEEE Micro,2008,28(2):39-55.
[3] Kyung G,Jung C,Lee K.An implementation of a SIMT architecture-based stream processor[C]//Tencon IEEE Region 10 Conference.IEEE,2015.
[4] 韩峰.可重构及SIMT处理器系统架构存储映射方法研究[D].南京:南京大学,2017.
[5] 吕华智.高性价比DSP中指令Cache的研究与设计[D].湘潭:湘潭大学,2017.
[6] 袁潇.DSP处理器中数据Cache的设计和验证[D].西安:西安电子科技大学,2013.
[7] 张多利,张宇,宋宇鲲,等.一种多核SoC中基于Cache机制的存储结构设计[J].微电子学与计算机,2017,34(10):26-31.
[2] Lindholm E,Nickolls J,Oberman S,et al.NVIDIA Tesla:A Unified Graphics and Computing Architecture[J].IEEE Micro,2008,28(2):39-55.
[3] Kyung G,Jung C,Lee K.An implementation of a SIMT architecture-based stream processor[C]//Tencon IEEE Region 10 Conference.IEEE,2015.
[4] 韩峰.可重构及SIMT处理器系统架构存储映射方法研究[D].南京:南京大学,2017.
[5] 吕华智.高性价比DSP中指令Cache的研究与设计[D].湘潭:湘潭大学,2017.
[6] 袁潇.DSP处理器中数据Cache的设计和验证[D].西安:西安电子科技大学,2013.
[7] 张多利,张宇,宋宇鲲,等.一种多核SoC中基于Cache机制的存储结构设计[J].微电子学与计算机,2017,34(10):26-31.