Energy efficiency is key in many embedded systems that must achieve best performances for a given power budget. Additionally, new neural network-based applications combine multiple processing needs. For such applications, heterogeneous system-on-chips, such as the Nvidia Jetson Orin series, include different computing capabilities to propose new interesting latency and power consumption trade-offs. But, choosing the suitable Jetson module for a given application's need can be confusing since these modules have many operating ranges and several accelerators. In this paper, we evaluate through emulation the embedded performances of popular neural networks to provide a first hands-on insight of all Jetson Orin modules.