Designing embedded systems includes two main phases: (i) HW/SW Partitioning performed from high-level functional and architecture models, and (ii) Software Design performed with significantly more detailed models. Partitioning decisions are made according to performance assumptions that should be validated on the more refined software models. In this paper, we focus on one such metric: latencies between operations. We show how they can be modeled at different abstraction levels (partitioning, SW design) and how they can help determine accuracy of the computational complexity estimates made during HW/SW Partitioning.