EMT/NANO: Broadcast Optical Interconnects for Global Communication in Many-Core Chip-Multiprocessor

EMT/NANO：多核芯片多处理器中的全球通信广播光互连Alan Mickelson（首席研究员）Dejan Filipovi 'cWon ParkLi ShangManish Vachharajani科罗拉多大学电气和计算机工程摘要：计算机性能扩展对于医学、电信、气候科学、天气预报、工程和设计等广泛领域的进步至关重要。随着多核系统的兴起和处理器核心数量的增加，并行程序性能现在是继续这种性能扩展的关键。这项工作强调延迟减少，因为这将是最重要的性能障碍，随着处理器数量的增加。电互连可以在小距离上支持大的信息带宽，例如相邻处理器之间。分组交换可以在处理器之间共享带宽，但代价是延迟会随着处理器间的间隔而增加。这项研究涉及到使用纳米级光学实现一个光广播网络，可以在一个单一的时钟周期内解决所有的处理器。媒体访问控制将在高带宽的点对点报文和延迟受限的同步报文之间划分信息，研究者从实验和理论两方面解决了混合片上互连网络的问题。大部分的实验工作将涉及制造和演示纳米光子广播网络的操作。该广播网络将由一个二维平板波导区域与光学天线，从纳米通道波导组件馈送和读出寻址。光源将被放置在芯片外，并通过绝缘体上的硅波导输入到芯片。波分复用将通过每个处理器互连的单个收发器提供多个同时的通道操作。将处理器间通信划分为高带宽分组交换通信和延迟限制同步消息的媒体访问控制将使用实际互连网络的实验数据和唯象模型来模拟。

EMT/NANO: Broadcast Optical Interconnects forGlobal Communication in Many-CoreChip-MultiprocessorAlan Mickelson (Principal Investigator)Dejan Filipovi´cWon ParkLi ShangManish VachharajaniElectrical and Computer EngineeringUniversity of ColoradoAbstract:Computer performance scaling is critical to advances in a wide range of fields, includingmedicine, telecommunications, climate science, weather forecasting, engineering, and design.With the rise of multicore systems and increasing processor-core count as the hardwaremechanism for increased compute power, parallel program performance is now critical tocontinue this performance scaling. This work emphasizes latency reduction as this will bethe most important performance impediment as the number of processors increases. Electricalinterconnections can support the large information bandwidth over small distance, forexample, between neighboring processors. Packet switching can share bandwidth betweenprocessors at a cost of latency that increases with inter-processor spacing. This researchinvolves the use of nano scale optics to implement an optical broadcast network that canaddress all processors during a single clock cycle. Media access control will partition informationbetween the high bandwidth point to point messages to be packet switched andlatency bound synchronization messages to be optically broadcast.The investigators address the problem of a hybrid on-chip interconnection network bothexperimentally and theoretically. The bulk of the experimental work will involve fabricatingand demonstrating the operation of nanophotonic broadcast network. The broadcastnetwork will consist of a two dimensional slab waveguide region addressed with optical antennasthat are fed and read out from nanophonic channel waveguide components. Opticalsources will be placed off-chip and input to the netwrk through silicon on insulator waveguides.Wavelength division multiplexing will provide multi simultaneous channel operationthrough the single transceiver per processor interconnect. The media access control thatpartitions inter processor communication into high bandwidth packet switched communicationsand latency bound synchronization messages will be simulated using experimental dataand phenomenological models of the actual interconnection network.