Do we want to model every contention point and queuing detail ?
Use a high-level model: Treat the NoC just as a square array corresponding to the floor plan of the chip and in each entry we hold a running average local utilisation.
Or model even more abstractly: uses Rent's Rule estimate of average communication length assuming a `sensible' layout.
Rent's rules of thumb predict number of connections per sub-system and reflect the expected reduction in average wire length arising from good design and layout compared to a random collection of components.
Problems: