Spatial Parallelism in the Routers of Asynchronous On-Chip Networks

Wei Song

Abstract

State-of-the-art multi-processor systems-on-chip use on-chip networks as their communication fabric. Although most on-chip networks are implemented synchronously, asynchronous on-chip networks have several advantages over their synchronous counterparts. Timing division multiplexing (TDM) flow control methods have been utilized in asynchronous on-chip networks extensively. The synchronization required by TDM leads to significant speed penalties. Compared with using TDM methods, spatial parallelism methods, such as the spatial division multiplexing (SDM) flow control method, achieve better network throughput with less area overhead. This thesis proposes several techniques to increase spatial parallelism in the routers of asynchronous on-chip networks. Channel slicing is a new pipeline structure that alleviates the speed penalty by removing the synchronization among bit-level data pipelines. It is also found out that the lookahead pipeline using early evaluated acknowledgement can be used in routers to further improve speed. SDM is a new flow control method proposed for asynchronous on-chip networks. It improves network throughput without introducing synchronization among buffers of different frames, which is required by TDM methods. It is also found that the area overhead of SDM is smaller than the virtual channel (VC) flow control method -- the most used TDM method. The major design problem of SDM is the area consuming crossbars. A novel 2-stage Clos switch structure is proposed to replace the crossbar in SDM routers, which significantly reduces the area overhead. This Clos switch is dynamically reconfigured by a new asynchronous Clos scheduler. Several asynchronous SDM routers are implemented using these new techniques. An asynchronous VC router is also reproduced for comparison. Performance analyses show that the SDM routers outperform the VC router in throughput, area overhead and energy efficiency.

The thesis is available as PDF (2.3MB).