Dynamic Resource Sharing for High-Performance 3-D Networks-on-Chip

Abstract

3D logic-on-logic technology is a promising approach for extending the validity of Moore's law when technology scaling stops. 3D technology can also lead to a paradigm shift in on-chip communication design by providing orders of magnitude higher bandwidth and lower latency for inter-layer communication. To turn the 3D technology bandwidth and latency benefits into network latency reductions and performance improvement, we need networks-on-chip (NoCs) that are specially designed to take advantage of what 3D technology has to offer. While in parallel workloads many packets experience blocking in the network due to losing arbitration for crossbars' input/output ports, we observe that in a considerable fraction of these cases in a 3D NoC, the corresponding input and output ports of the crossbar in the above or below router are idle. Given this observation, we propose FRESH, a router microarchitecture with Fine-grained 3D REsource SHaring capability that leverages the ultra-low latency vertical links of a 3D chip to share crossbars and links at a fine granularity between vertically stacked routers. It enables packets that lose arbitration for crossbars' input/output ports to use idle resources of the above or below routers, and effectively eliminates the unnecessary packet blocking time. We will show that our proposal lowers network latency by up to 21% over the state-of-the-art 3D NoC.