Ina-ocmc 2011

MPSoCs are today frequently designed as the composition of multiple voltage/frequency islands, thus calling for a GALS clocking style. In this context, the on-chip interconnection network can be either inferred as a single and independent clock domain or it can be distributed among core’s domains. This paper targets the former scenario, since it results in the homogeneous speed of the NoC switching elements. From a physical design viewpoint, the main issues lie however in the chip-wide extension of the network domain and in the growing uncertainties affecting nanoscale silicon technologies. This paper proves that partitioning the network into mesochronous domains and merging synchronizers with NoC building blocks, two main advantages can be achieved. First, it is possible to evolve synchronous networks to mesochronous ones with marginal performance and area overhead. Second, the mesochronous NoC exposes more degrees of freedom for power optimization.