In this paper, we propose a technique for lowering the latency of the communication in a NoC （network on chip）. The technique, which can support two qualities of service （QoS）, i.e., the guaranteed throughput （GT...In this paper, we propose a technique for lowering the latency of the communication in a NoC （network on chip）. The technique, which can support two qualities of service （QoS）, i.e., the guaranteed throughput （GT） and best effort （BE）, is based on splitting a wider link into narrower links to increase throughput and decrease latency in the NoC. In addition, to ease the synchronization and reduce the crosstalk, we use the l-of-4 encoding for the smaller buses. The use of the encoding in the proposed NoC architecture considerably lowers the latency for both BE and GT packets. In addition, the bandwidth is increased while the power consumption of the links is reduced.展开更多
This paper presents the result of experiments conducted in mesh networks on different routing algorithms, traffic generation schemes and switching schemes. A new network on chip (NoC) topology based on partial interco...This paper presents the result of experiments conducted in mesh networks on different routing algorithms, traffic generation schemes and switching schemes. A new network on chip (NoC) topology based on partial interconnection of mesh network is proposed and a routing algorithm supporting the proposed architecture is developed. The proposed architecture is similar to standard mesh networks, where four extra bidirectional channels are added which remove the congestion and hotspots compared to standard mesh networks with fewer channels. Significant improvement in delay (60% reduction) and throughput (60% increase) was observed using the proposed network and routing when compared with the ideal mesh networks. An increase in number of channels makes the switches expensive and could increase the area and power consumption. However, the proposed network can be useful in high speed applications with some compromise on area and power.展开更多
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...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。展开更多
三维片上网络(three-dimensional network on chip,3D No C)是在三维集成电路(three-dimensional integrated circuit,3D IC)、片上系统(system on chip,So C)和二维片上网络(two-dimensional network on chip,2D No C)的基础上发展起来...三维片上网络(three-dimensional network on chip,3D No C)是在三维集成电路(three-dimensional integrated circuit,3D IC)、片上系统(system on chip,So C)和二维片上网络(two-dimensional network on chip,2D No C)的基础上发展起来的,主要解决高集成度芯片通信瓶颈等问题,已引起国内外学术界和产业界的高度重视。3D No C拓扑结构体现了通信节点在芯片中的布局与连接,对三维芯片性能起决定性作用。简介了2D No C、2D No C到3D No C的演变、3D No C的优点与存在的问题以及3D No C解决的关键技术问题,分析了3D No C总体发展状况。三维拓扑结构是3D No C设计中的关键问题之一,重点研究了3D No C拓扑结构的分类方法,从通信角度将3D No C拓扑结构分成9大类,分类论述了3D No C拓扑结构,并分析比较了现有63种拓扑结构各自的特点,最后指出了3D No C拓扑结构的未来研究方向。展开更多
基金Project supported by the Iranian National Science Foundation
文摘In this paper, we propose a technique for lowering the latency of the communication in a NoC （network on chip）. The technique, which can support two qualities of service （QoS）, i.e., the guaranteed throughput （GT） and best effort （BE）, is based on splitting a wider link into narrower links to increase throughput and decrease latency in the NoC. In addition, to ease the synchronization and reduce the crosstalk, we use the l-of-4 encoding for the smaller buses. The use of the encoding in the proposed NoC architecture considerably lowers the latency for both BE and GT packets. In addition, the bandwidth is increased while the power consumption of the links is reduced.
文摘This paper presents the result of experiments conducted in mesh networks on different routing algorithms, traffic generation schemes and switching schemes. A new network on chip (NoC) topology based on partial interconnection of mesh network is proposed and a routing algorithm supporting the proposed architecture is developed. The proposed architecture is similar to standard mesh networks, where four extra bidirectional channels are added which remove the congestion and hotspots compared to standard mesh networks with fewer channels. Significant improvement in delay (60% reduction) and throughput (60% increase) was observed using the proposed network and routing when compared with the ideal mesh networks. An increase in number of channels makes the switches expensive and could increase the area and power consumption. However, the proposed network can be useful in high speed applications with some compromise on area and power.
文摘针对深度神经网络(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。
文摘三维片上网络(three-dimensional network on chip,3D No C)是在三维集成电路(three-dimensional integrated circuit,3D IC)、片上系统(system on chip,So C)和二维片上网络(two-dimensional network on chip,2D No C)的基础上发展起来的,主要解决高集成度芯片通信瓶颈等问题,已引起国内外学术界和产业界的高度重视。3D No C拓扑结构体现了通信节点在芯片中的布局与连接,对三维芯片性能起决定性作用。简介了2D No C、2D No C到3D No C的演变、3D No C的优点与存在的问题以及3D No C解决的关键技术问题,分析了3D No C总体发展状况。三维拓扑结构是3D No C设计中的关键问题之一,重点研究了3D No C拓扑结构的分类方法,从通信角度将3D No C拓扑结构分成9大类,分类论述了3D No C拓扑结构,并分析比较了现有63种拓扑结构各自的特点,最后指出了3D No C拓扑结构的未来研究方向。
