Learning-Based Design Space Exploration of Emerging 3D NoC Architectures

In the absence of Moore’s Law, three-dimensiona a promising direction to satisfy the demand from big-data applications. However, due to the growing heterogeneity, system size, and issues stemming from emerging interconnecl (3D) manycore systems that incorporate both CPUs and GPUs presentsts such as monolithic 3D (M3D), it is difficult to satisfactorily explore the design space in reasonable amounts of time. In this paper, we reinforce the need for multi-objective formulations (optimizing for power, performance, thermals, etc. simultaneously) and present a solution that combines machine learning and local searches to quickly and intelligently explore the design space. In addition, we look at two specific problems (TSV-based 3D heterogeneous manycore systems and M3D-based NoC) and what considerations need to be made at design-time.