While general purpose graphics processing units now embark tremendous amount of computing power, their use in real time applications is still a challenge. The COSMIC platform, developed in the context of adaptive optics control for giant astronomical telescopes is a demonstrated solution to perform real time computations using discrete GPUs while maintaining a high level of abstraction and modularity. An implementation on embedded platforms with the goal to reach an acceptable level of time-determinism would enable new real-time use cases in other application domains. In this regard, NVIDIA is offering a broad range of embedded systems on chip delivering great performance and compatible with the CUDA ecosystem. However, specific hardware and software features bring uncertainties regarding real-time performance. The approaches presented in this paper rely on COSMIC recipes to expose part of the underlying unconventional GPU programming model to reach real-time performance. It shows how Jetson Xavier performs on sub-millisecond complex pipelines made of several compute kernels, considering the limitations engendered by missing CUDA features as compared to discrete devices, leveraging unified memory to work around these hurdles and enabling several strategies for implementing real-time workflows on embedded GPU platforms.