Dual Data Rate Network-on-Chip Architectures

Doctoral thesis, 2021

Networks-on-Chip (NoCs) are becoming increasing important for the performance of modern multi-core systems-on-chip. The performance of current NoCs is limited, among others, by two factors: their limited clock frequency and long router pipeline. The clock frequency of a network defines the limits of its saturation throughput. However, for high throughput routers, clock is constrained by the control logic (for virtual channel and switch allocation) whereas the datapath (crossbar switch and links) possesses significant slack. This slack in the datapath wastes network throughput potential. Secondly, routers require flits to go through a large number of pipeline stages increasing packet latency at low traffic loads. These stages include router resource allocation, switch traversal (ST) and link traversal (LT). The allocation stages are used to manage contention among flits attempting to simultaneously access switch and links, and the ST stage is needed to change the routing dimension. In some cases, these stages are not needed and then requiring flits to go through them increases packet latency. The aim of this thesis is to improve NoC performance, in terms of network throughput, by removing the slack in the router datapath, and in terms of average packet latency, by enabling incoming flits to bypass, when possible, allocation and ST stages. More precisely, this thesis introduces Dual Data-Rate (DDR) NoC architectures which exploit the slack present in the NoC datapath to operate it at DDR. This requires a clock with period twice the datapath delay and removes the control logic from the critical path. DDR datapaths enable throughput higher than existing single data-rate (SDR) networks where the clock period is defined by the control logic. Additionally, this thesis supplements DDR NoC architectures with varying levels of pipeline stage bypassing capabilities to reduce low-load packet latency. In order to avoid complex logic required for bypassing from all inputs to all outputs, this thesis implements and evaluates a simplified bypassing approach. DDR NoC routers support bypassing of the allocation stage for flits propagating an in-network straight hop (i.e. East to West, North to South and vice versa) and when entering or exiting the network. Disabling bypassing during XY-turns limits its benefits, but, for most routing algorithms under low traffic loads, flits encounter at most one XY-turn on their way to the destination. Bypassing allocation stage enables incoming flits to directly initiate ST, when required conditions are met, and propagate at one cycle per hop. Furthermore, DDR NoC routers allow flits to bypass the ST stage when propagating a straight hop from the head of a specific input VC. Restricting ST bypassing from a specific VC further simplifies check logic to have clock period defined by datapath delays. Bypassing ST requires dedicated bypass paths from non-local input ports to opposite output ports. It enables flits to propagate half a cycle per hop. This thesis shows that compared to current state-of-the-art SDR NoCs, operating router’s datapath at DDR improves throughput by up to 20%. Adding to a DDR NoC an allocation bypassing mechanism for straight hops reduces its packet latency by up to 45%, while maintaining the DDR throughput advantage. Enhancing allocation bypassing to include flits entering and exiting the network further reduces latency by another 24%. Finally, adding ST bypassing further reduces latency by another 32%. Overall, DDR NoCs offer up to 40% lower latency and about 20% higher throughput compared to the SDR networks.