Embedded Computer Vision is a hot field of research that requires trade-offs in order to balance execution time, power consumption and accuracy. In that field, dense optical flow estimation is a major tool used in many applications. Many algorithms have been designed, focusing on accuracy, very few works address trade-offs and implementation on embedded hardware. This paper tackles these trade-offs for embedded GPU through the example of the well-known TV-L 1 algorithm. Thanks to High Level Transforms-operator fusion and pipeline-and taking into account the iterative aspect of these algorithms, we achieve a speedup of ×3.7 versus OpenCV. Moreover, we show that a 16bit half precision implementation has a higher accuracy than the 32-bit precision one for the same frame processing time on NVIDIA Jetson boards. Furthermore, this work can be generalized to any kind of iterative stencil-based algorithms.