CROSS-LAYER DESIGN, OPTIMIZATION AND PROTOTYPING OF NoCs FOR THE NEXT GENERATION OF HOMOGENEOUS MANY-CORE SYSTEMS

This thesis provides a whole set of design methods to enable and manage the runtime heterogeneity of features-rich industry-ready Tile-Based Networkon- Chips at different abstraction layers (Architecture Design, Network Assembling, Testing of NoC, Runtime Operation). The key idea is to maintain the functionalities of the original layers, and to improve the performance of architectures by allowing, joint optimization and layer coordinations. In general purpose systems, we address the microarchitectural challenges by codesigning and co-optimizing feature-rich architectures. In application-specific NoCs, we emphasize the event notification, so that the platform is continuously under control. At the network assembly level, this thesis proposes a Hold Time Robustness technique, to tackle the hold time issue in synchronous NoCs. At the network architectural level, the choice of a suitable synchronization paradigm requires a boost of synthesis flow as well as the coexistence with the DVFS. On one hand this implies the coexistence of mesochronous synchronizers in the network with dual-clock FIFOs at network boundaries. On the other hand, dual-clock FIFOs may be placed across inter-switch links hence removing the need for mesochronous synchronizers. This thesis will study the implications of the above approaches both on the design flow and on the performance and power quality metrics of the network. Once the manycore system is composed together, the issue of testing it arises. This thesis takes on this challenge and engineers various testing infrastructures. At the upper abstraction layer, the thesis addresses the issue of managing the fully operational system and proposes a congestion management technique named HACS. Moreover, some of the ideas of this thesis will undergo an FPGA prototyping. Finally, we provide some features for emerging technology by characterizing the power consumption of Optical NoC Interfaces.