Latency is a major issue in the design and validation of a Network-on-Chip (NoC). Various techniques for establishing latency bounds exist. Formal and mathematical methods, such as network calculus, can be used to analyze an NoC model. Simulation-based methods can be used to estimate latency bounds by exploring reachable states. Both have their advantages and disadvantages. This paper presents an approach that finds a middle ground between these two worlds. Our approach is based on simulation of high-level formal models. In contrast to traditional formal methods for worst-case latency, we do not require error-prone manual computation or the absence of cycles. In contrast to traditional simulation-based methods, we leverage the high level of abstraction to explore up to billions of states within a couple of hours. We apply our approach on an 8 core case study where a simple cache protocol runs on top of a ring-based Spidergon architecture. We show that deadlocks or starvations are easily found, and that for live networks a worst-case bound estimation can be produced within reasonable time.