Article Preview
TopIntroduction
Network-on-Chips (NoCs) provide a general purpose communications fabric for multi-core architectures that allow functional blocks to communicate with each other. These networks are intended to provide a scalable, power and space efficient means to support growth in component density and many-core designs (Dally & Towles, 2001). Modern Chip Multi-Processors (CMPs) exploit NoCs to efficiently connect many processing elements (PEs) and other devices. System performance is dependent on low-latency, high-bandwidth communications between cores. As the number of cores integrated in a single chip increases, greater demands are placed on the communication network. Thus, the design of the inter-connection fabric and not the PE is now the bottleneck in computation. Power and area are central considerations in the design of the interconnect. This has led to a wide range of literature, with topics including power analysis (Banerjee, Mullins, & Moore, 2007), router design (Michelogiannakis & Dally, 2009) and routing algorithms such as Opt-y (Schwiebert & Jayasimha, 1993) and IVAL (Towles, Dally, & Boyd, 2003). However, the main issue today is the delivery of energy-efficient and scalable interconnects: the focus of this paper.
NoC Design
An ideal NoC provides an area-efficient low-power network that exhibits good latency characteristics across a range of traffic patterns. A NoC can be designed for a specific application, affecting the choice of topology, routing algorithm and router design. These specialised networks can offer a performance advantage over a regular architecture, but are inflexible and costly to design. They require static analysis of the target application’s inter-process communications behaviour and offer poor performance for general purpose computation.
In Asanovic et al. (2006) Dwarfs were introduced, where applications are classified in terms of their computation and communication patterns. For example, some applications have fixed, predictable communication between specific cores, while others have sporadic communication between all cores. In Asanovic et al. (2006) the shortcomings of fixed-topology on-chip interconnects are discussed and it is suggested that many applications have communication patterns for which an optimal topology can be configured. Therefore, the design of a flexible, common NoC topology is of paramount importance in the design of efficient next-generation systems. For example, Shalf et al. (2005) consider reconfigurable hybrid interconnects for Ultra-scale applications. They conclude that many Ultra-scale applications will benefit from a reconfigurable interconnection fabric. In this paper, we demonstrate how this can be achieved in a general sense, using the mesh topology as a basis. We will introduce dynamically-reconfigurable Skip-links to enable on-the-fly adaptation of the network topology to traffic behavior.