Network-on-Chip Design for Heterogeneous Multiprocessor System-on-Chip

With burgeoning growth of mobile systems, multiprocessor System-on-Chip (MPSoC) connected via Network-on-Chip (NoC) has become ubiquitous. A typical MPSoC in mobile applications consists of multiple CPU cores of varying capabilities, GPU cores, DSP cores, and crypto accelerators and such cores differ widely in their physical size and their bandwidth requirements. Traditional mesh based NoC systems work well for regular structures, but do not map well to heterogeneous MPSoCs. In MPSoC programming model, an application consists of tasks, that represent a unit of work on a core which can be executed asynchronously. The communication between tasks is represented in the form of a directed acyclic graph. The temporal burstness of data which arise from programming model provide opportunity for multiplexing communication between cores, which may be advantageous in reducing network size. Often a task graph needs to meet a real-time deadline. The actual execution time may vary based on the application data. The uncertainty in the execution time may be modeled by a statistical distribution, which further complicates the NoC design. In this paper, we present a synthesis method for hierarchical design of NoC for a given task graph system deadline, that optimizes for router area. A 2-phase design flow is proposed, which consists of topology generation and statistical analysis in an iterative loop. We adopt proportion of Monte-Carlo test cases that meet the deadline as a metric for goodness. The proposed solution is compared against static design approach and simulated annealing (SA) based network generation. On an average, a performance benefit of 10% over SA, 16% over standard mesh and 30% over static design was obtained and a total router area benefit of 59% over SA, 48% over mesh and 55% over static design was observed.