摘要
随着无线通信技术的发展,实现多输入多输出(MIMO)系统检测性能与复杂度之间的最优权衡日益困难,深度学习DL为此提供了新方向。文中提出基于片上网络(NoC)的多核动态可重构架构MCDBP,以提高基于DL的MIMO检测算法的性能,并增强架构的可编程性和扩展性。MCDBP通过集成轻量级计算内核及片上网络互连,并行处理矢量-矩阵乘法、常数-矢量乘法、矢量点积、矢量加法等大多数深度展开网络的基本运算,有效提高复杂MIMO检测性能。架构的创新在于可重构的处理元件PE设计,可以依据DL驱动的MIMO检测需求动态调整。该设计对基于DL的MIMO检测算法共性进行深入分析,支持多种基本运算模式,展现极高灵活性。实验结果显示,MCDBP在执行基于DL的MIMO检测算法时,与通用CPU相比,可以实现12.66~22.98的加速比,算法性能有所提高,可以适应不同应用场景。
With the advancement of wireless communication technologies,achieving the optimal balance between the detec-tion performance and complexity of multiple-input multiple-output(MIMO)systems is increasingly challenging.Deep learning(DL)offers a new direction for this.This paper presents a multi-core dynamic reconfigurable architecture based on network on chip(NoC).This architecture,termed MCDBP(multi-core architecture for dynamic baseband processing),strives to enhance the performance for DL-based MIMO detection algorithms and the architecture's programmability and scalability.The MCDBP leverages integrated lightweight computing cores and NoC interconnects to process the fundamental operations of deep unfolded networks in parallel,such as vector-matrix multiplication(VMM),constant-vector multiplication(CVM),vector dot product(VDP),and vector addition(VA),so as to improve the performance of complex MIMO detection significantly.The innovation of the architecture lies in the reconfigurable design of the processing elements(PEs),and the architecture can be adjusted according to different DL-based MIMO detection algorithms dynamically.This design is grounded in a thorough analysis of the commonalities of DL-based MIMO detection algorithms,showcasing extreme flexibility in supporting multiple fundamental operational modes.Experimental results indicate that,in comparison with the general-purpose CPU,MCDBP can achieve an acceleration ratio of 12.66~22.98 when implementing DL-based MIMO detection algorithms.It can be seen that the performance of the algorithm is improved,so the algorithm can adapt to different application scenarios.
作者
范文杰
周牧也
朱凌晓
李世平
陈铠
邓松峰
何国强
冯书谊
宋文清
李丽
傅玉祥
FAN Wenjie;ZHOU Muye;ZHU Lingxiao;LI Shiping;CHEN Kai;DENG Songfeng;HE Guoqiang;FENG Shuyi;SONG Wenqing;LI Li;FU Yuxiang(School of Electronic Science and Engineering,Nanjing University,Nanjing 210023,China;School of Integrated Circuits,Nanjing University,Suzhou 215163,China;Jiangsu Huachuang Microsystems Co.,Ltd.,Nanjing 211899,China;Shanghai Aerospace Electronic Technology Research Institute,Shanghai 201100,China)
出处
《现代电子技术》
北大核心
2024年第21期1-6,共6页
Modern Electronics Technique
基金
国家自然科学基金青年科学基金项目(62104098)
国家自然科学基金企业创新发展联合基金重点项目(U21B2032)
江苏省基础研究专项资金(自然科学基金)青年基金项目(BK20210178)
科技部重点研发计划(2021YFB3600104,2023YFB2806802)。
关键词
无线通信
MIMO检测
深度学习
数据驱动网络
模型驱动网络
NOC
可重构
多核架构
wireless communication
MIMO detection
deep learning
data-driven network
model-driven network
NoC
re-configurable
multi-core architecture
