Approaching the theoretical limits of a mesh NoC with a 16-node chip prototype in 45nm SOI

Abstract

In this paper, we present a case study of our chip prototype of a 16-node 4x4 mesh NoC fabricated in 45nm SOI CMOS that aims to simultaneously optimize energy-latency-throughput for unicasts, multicasts and broadcasts. We first define and analyze the theoretical limits of a mesh NoC in latency, throughput and energy, then describe how we approach these limits through a combination of microarchitecture and circuit techniques. Our 1.1V 1GHz NoC chip achieves 1-cycle router-and-link latency at each hop and energy-efficient router-level multicast support, delivering 892Gb/s (87.1% of the theoretical bandwidth limit) at 531.4mW for a mixed traffic of unicasts and broadcasts. Through this fabrication, we derive insights that help guide our research, and we believe, will also be useful to the NoC and multicore research community.

An Efficient 2D Router Architecture for Extending the Performance of Inhomogeneous 3D NoC-Based Multi-Core Architectures

“…Existing work reduces latency of NoC routers by either enhancing the classic router [16] or developing simpler micro-architectures. Approaches such as lookaheads [17] add wiring and logic complexity to routers, and increase NoC's area overhead and power consumption. Speculation [7] does not reduce the worst case pipeline delay.…”

SCORPIO: A 36-core research chip demonstrating snoopy coherence on a scalable mesh NoC with in-network ordering

“…To reduce the network latency and buffer read/write power, we implement looka- head (LA) bypassing [19,27]; a lookahead containing control information for a flit is sent to the next router during that flit's ST stage. At the next router, the lookahead performs route-computation and tries to pre-allocate the crossbar for the approaching flit.…”

“…To alleviate the overhead imposed by the coherence broadcast requests, routers are equipped with single-cycle multicast support [27]. Instead of sending the same requests for each node one by one into the main network, we allow requests to fork through multiple router output ports in the same cycle, thus providing efficient hardware broadcast support.…”

“…Swizzle acknowledged scalability issues and proposed stopping at 64-port crossbars, and leveraging these as high-radix routers within NoCs. There are several other stand-alone NoC prototypes that also explored practical implementations with timing, power and area consideration, such as the 1 GHz Broadcast NoC [27] that optimizes for energy, latency and throughput using virtual bypassing and low-swing signaling for unicast, multicast, and broadcast traffic. Virtual bypassing is leveraged in the SCORPIO NoC.…”

“…The Achilles heel of buses lie in limited bandwidth, while that of rings is delay, and for crossbars, it is area. Higher-dimension NoCs such as meshes provide scalable bandwidth and is the subject of a plethora of research on low-power and low-latency routers, including several chip prototypes [16,17,27,30]. However, meshes are unordered and cannot natively support snoopy protocols.…”