Methodologies and Toolflows for the Predictable Design of Reliable and Low-Power NoCs

There is today the unmistakable need to evolve design methodologies and tool ows for Network-on-Chip based embedded systems. In particular, the quest for low-power requirements is nowadays a more-than-ever urgent dilemma. Modern circuits feature billion of transistors, and neither power management techniques nor batteries capacity are able to endure the increasingly higher integration capability of digital devices. Besides, power concerns come together with modern nanoscale silicon technology design issues. On one hand, system failure rates are expected to increase exponentially at every technology node when integrated circuit wear-out failure mechanisms are not compensated for. However, error detection and/or correction mechanisms have a non-negligible impact on the network power. On the other hand, to meet the stringent time-to-market deadlines, the design cycle of such a distributed and heterogeneous architecture must not be prolonged by unnecessary design iterations. Overall, there is a clear need to better discriminate reliability strategies and interconnect topology solutions upfront, by ranking designs based on power metric. In this thesis, we tackle this challenge by proposing power-aware design technologies. Finally, we take into account the most aggressive and disruptive methodology for embedded systems with ultra-low power constraints, by migrating NoC basic building blocks to asynchronous (or clockless) design style. We deal with this challenge delivering a standard cell design methodology and mainstream CAD tool ows, in this way partially relaxing the requirement of using asynchronous blocks only as hard macros.