Network-on-Chip or Network-on-a-Chip (NoC or NOC) is an approach to designing the communication subsystem between IP cores in a System-on-a-Chip (SoC). NoCs can span synchronous and asynchronous clock domains or use unclocked asynchronous logic. NoC applies networking theory and methods to on-chip communication and brings notable improvements over conventional bus and crossbar interconnections. NoC improves the scalability of SoCs, and the power efficiency of complex SoCs compared to other designs. Research has been done on integrated optical waveguides and devices comprising an Optical Network-on-Chip (ONoC).[1][2]
Contents |
Network-on-Chip (NoC) is an emerging paradigm for communications within large VLSI systems implemented on a single silicon chip. Sgroi et al. call "the layered-stack approach to the design of the on-chip intercore communications the Network-on-Chip (NOC) methodology." In a NoC system, modules such as processor cores, memories and specialized IP blocks exchange data using a network as a "public transportation" sub-system for the information traffic. A NoC is constructed from multiple point-to-point data links interconnected by switches (a.k.a. routers), such that messages can be relayed from any source module to any destination module over several links, by making routing decisions at the switches. A NoC is similar to a modern telecommunications network, using digital bit-packet switching over multiplexed links. Although packet-switching is sometimes claimed as necessity for a NoC, there are several NoC proposals utilizing circuit-switching techniques. This definition based on routers is usually interpreted so that a single shared bus, a single crossbar switch or a point-to-point network are not NoCs but practically all other topologies are. This is somewhat confusing since all above mentioned are networks (they enable communication between two or more devices) but they are not considered as network-on-chips. Note that some articles erroneously use NoC as a synonym for mesh topology although NoC paradigm does not dictate the topology. Likewise, the regularity of topology is sometimes considered as a requirement which is, obviously, not the case in research concentrating on "application-specific NoC topology synthesis".
The wires in the links of the NoC are shared by many signals. A high level of parallelism is achieved, because all links in the NoC can operate simultaneously on different data packets. Therefore, as the complexity of integrated systems keeps growing, a NoC provides enhanced performance (such as throughput) and scalability in comparison with previous communication architectures (e.g., dedicated point-to-point signal wires, shared buses, or segmented buses with bridges). Of course, the algorithms must be designed in such a way that they offer large parallelism and can hence utilize the potential of NoC.
Traditionally, ICs have been designed with dedicated point-to-point connections, with one wire dedicated to each signal. For large designs, in particular, this has several limitations from a physical design viewpoint. The wires occupy much of the area of the chip, and in nanometer CMOS technology, interconnects dominate both performance and dynamic power dissipation, as signal propagation in wires across the chip requires multiple clock cycles. (See Rent's rule for a discussion of wiring requirements for point-to-point connections).
NoC links can reduce the complexity of designing wires for predictable speed, power, noise, reliability, etc., thanks to their regular, well controlled structure. From a system design viewpoint, with the advent of multi-core processor systems, a network is a natural architectural choice. A NoC can provide separation between computation and communication, support modularity and IP reuse via standard interfaces, handle synchronization issues, serve as a platform for system test, and, hence, increase engineering productivity.
Although NoCs can borrow concepts and techniques from the well-established domain of computer networking, it is impractical to blindly reuse features of "classical" computer networks and symmetric multiprocessors. In particular, NoC switches should be small, energy-efficient, and fast. Neglecting these aspects along with proper, quantitative comparison was typical for early NoC research but nowadays they are considered in more detail. The routing algorithms should be implemented by simple logic, and the number of data buffers should be minimal. Network topology and properties may be application-specific.
Some researchers think that NoCs need to support quality of service (QoS), namely achieve the various requirements in terms of throughput, end-to-end delays and deadlines. Real-time computation, including audio and video playback, is one reason for providing QOS support. However, current system implementations like VxWorks, RTLinux or QNX are able to achieve sub-millisecond real-time computing without special hardware. This may indicate that for many real-time applications the service quality of existing on-chip interconnect infrastructure is sufficient, and dedicated hardware logic would be necessary to achieve microsecond precision, a degree that is rarely needed in practice for end users (sound or video jitter need only tenth of milliseconds latency guarantee). Another motivation for NOC-level quality-of-service is to support multiple concurrent users sharing resources of a single chip multiprocessor in a public cloud computing infrastructure. In such instances, hardware QOS logic enables the service provider to make contractual guarantees on the level of service that a user receives, a feature that may be deemed desirable by some corporate or government clients.
To date, several prototype NoCs have been designed and analyzed in both industry and academia but only few have been implemented on silicon. However, many challenging research problems remain to be solved at all levels, from the physical link level through the network level, and all the way up to the system architecture and application software. The first dedicated research symposium on Networks on Chip was held at Princeton University, in May 2007.[3] The second IEEE International Symposium on Networks-on-Chip was held in April 2008 at Newcastle University.
NoC development and studies require comparing different proposals and options. And NoC benchmarks are developed to help such evaluations. Existing NoC benchmarks include NoCBench and MCSL NoC Benchmark Suite[4].
Adapted from Avinoam Kolodny's's column in the ACM SIGDA e-newsletter by Igor Markov
The original text can be found at http://www.sigda.org/newsletter/2006/060415.txt