| As digital systems continue to grow in complexity, the design of conventional synchronous systems is facing unprecedented challenges. The number of transistors on individual chips is already in the multi-billion range, and a greatly increasing number of components are being integrated onto a single chip. As a consequence, modern digital designs are under strong time-to-market pressure, and there is a critical need for composable design approaches for large complex systems.;In the past two decades, networks-on-chip (NoC's) have been a highly active research area. In a NoC-based system, functional blocks are first designed individually and may run at different clock rates. These modules are then connected through a structured network for on-chip global communication. However, due to the rigidity of centrally-clocked NoC's, there have been bottlenecks of system scalability, energy and performance, which cannot be easily solved with synchronous approaches. As a result, there has been significant recent interest in combing the notion of asynchrony with NoC designs. Since the NoC approach inherently separates the communication infrastructure, and its timing, from computational elements, it is a natural match for an asynchronous paradigm. Asynchronous NoC's, therefore, enable a modular and extensible system composition for an 'object-orient' design style.;The thesis aims to significantly advance the state-of-art and viability of asynchronous and globally-asynchronous locally-synchronous (GALS) networks-on-chip, to enable high-performance and low-energy systems. The proposed asynchronous NoC's are nearly entirely based on standard cells, which eases their integration into industrial design flows. The contributions are instantiated in three different directions.;First, practical acceleration techniques are proposed for optimizing the system latency, in order to break through the latency bottleneck in the memory interfaces of many on-chip parallel processors. Novel asynchronous network protocols are proposed, along with concrete NoC designs. A new concept, called 'monitoring network', is introduced. Monitoring networks are lightweight shadow networks used for fast-forwarding anticipated traffic information, ahead of the actual packet traffic. The routers are therefore allowed to initiate and perform arbitration and channel allocation in advance. The technique is successfully applied to two topologies which belong to two different categories -- a variant mesh-of-trees (MoT) structure and a 2D-mesh topology. Considerable and stable latency improvements are observed across a wide range of traffic patterns, along with moderate throughput gains.;Second, for the first time, a high-performance and low-power asynchronous NoC router is compared directly to a leading commercial synchronous counterpart in an advanced industrial technology. The asynchronous router design shows significant performance improvements, as well as area and power savings. The proposed asynchronous router integrates several advanced techniques, including a low-latency circular FIFO for buffer design, and a novel end-to-end credit-based virtual channel (VC) flow control. In addition, a semi-automated design flow is created, which uses portions of a standard synchronous tool flow.;Finally, a high-performance multi-resource asynchronous arbiter design is developed. This small but important component can be directly used in existing asynchronous NoC's for performance optimization. In addition, this standalone design promises use in opening up new NoC directions, as well as for general use in parallel systems. In the proposed arbiter design, the allocation of a resource to a client is divided into several steps. Multiple successive client-resource pairs can be selected rapidly in pipelined sequence, and the completion of the assignments can overlap in parallel.;In sum, the thesis provides a set of advanced design solutions for performance optimization of asynchronous and GALS networks-on-chip. These solutions are at different levels, from network protocols, down to router- and component-level optimizations, which can be directly applied to existing basic asynchronous NoC designs to provide a leap in performance improvement. |
