Design and Validation of Network-on-Chip Architectures for the Next Generation of Multi-synchronous, Reliable, and Reconfigurable Embedded Systems.

Abstract

NETWORK-ON-CHIP (NoC) design is today at a crossroad. On one hand, the design principles to efficiently implement interconnection networks in the resource-constrained on-chip setting have stabilized. On the other hand, the requirements on embedded system design are far from stabilizing. Embedded systems are composed by assembling together heterogeneous components featuring differentiated operating speeds and ad-hoc counter measures must be adopted to bridge frequency domains. Moreover, an unmistakable trend toward enhanced reconfigurability is clearly underway due to the increasing complexity of applications. At the same time, the technology effect is manyfold since it provides unprecedented levels of system integration but it also brings new severe constraints to the forefront: power budget restrictions, overheating concerns, circuit delay and power variability, permanent fault, increased probability of transient faults. Supporting different degrees of reconfigurability and flexibility in the parallel hardware platform cannot be however achieved with the incremental evolution of current design techniques, but requires a disruptive approach and a major increase in complexity. In addition, new reliability challenges cannot be solved by using traditional fault tolerance techniques alone but the reliability approach must be also part of the overall reconfiguration methodology. In this thesis we take on the challenge of engineering a NoC architectures for the next generation systems and we provide design methods able to overcome the conventional way of implementing multi-synchronous, reliable and reconfigurable NoC. Our analysis is not only limited to research novel approaches to the specific challenges of the NoC architecture but we also co-design the solutions in a single integrated framework. Interdependencies between different NoC features are detected ahead of time and we finally avoid the engineering of highly optimized solutions to specific problems that however coexist inefficiently together in the final NoC architecture. To conclude, a silicon implementation by means of a testchip tape-out and a prototype on a FPGA board validate the feasibility and effectiveness