摘要
针对深度神经网络(deep neural network,DNN)模型在传统切片与映射方法中存在的资源调度和数据传输瓶颈问题,提出了一种基于片上网络(network on chip,NoC)加速器的高效DNN动态切片与智能映射优化算法。该算法通过动态切片技术灵活划分DNN模型的计算任务,并结合智能映射策略优化NoC架构中的任务分配与数据流管理。实验结果表明,与传统方法相比,该算法在计算吞吐量、NoC传输时延、外部内存访问次数和计算能效等方面均显著提升,尤其在复杂模型上表现突出。
To address the bottlenecks of resource scheduling and data transmission in traditional slicing and mapping methods for deep neural networks(DNN),an efficient dynamic slicing and intelligent mapping optimization algo‐rithm was proposed based on a network on chip(NoC)accelerator.The algorithm was designed to flexibly divide DNN computing tasks through dynamic slicing and optimize task and data flow management in the NoC architecture.Experimental results show that the proposed algorithm significantly outperforms traditional methods in computing throughput,NoC transmission delay,external memory accesses,and energy efficiency,especially for complex models.
作者
齐芸
欧阳一鸣
QI Yun;OUYANG Yiming(Anhui Communications Vocational and Technical College,Hefei 230051,China;School of Computer and Information,Hefei University of Technology,Hefei 230051,China)
出处
《电信科学》
北大核心
2025年第10期151-160,共10页
Telecommunications Science
基金
国家自然科学基金资助项目(No.62374049)
安徽高校自然科学研究项目(No.2024AH050281,No.2024AH040051,No.2024AH050284)。
作者简介
通信作者:齐芸(1984-),女,安徽交通职业技术学院讲师,主要研究方向为片上网络,814993657@qq.com;欧阳一鸣(1963-),男,博士,合肥工业大学教授、博士生导师,主要研究方向为基于片上网络的人工智能应用。
