Fine grained adaptive simulation with application to NoCs

We present a new adaptive simulation approach, where the simulation precision can be changed between RTL and gate level for individual submodules online during simulation. As gate level precision is rarely required for the total simulated duration, the low simulation performance that comes along with full gate level simulations calls for a methodology to simulate at highest precision only when actually needed. To fulfill this need, we introduce adaptive SystemC models encapsulating an RTL model and a synthesized gate level counterpart. Adaptive models multiplex between fixed abstraction models and efficiently perform the necessary state transfer, which is enabled in an automated way by giving limited access to the simulation kernel. Evaluations highlight the applicability of adaptive simulations whenever gate level simulations are only required part time, as speedups of 150 were achieved compared to full gate level simulation.