Computational science has benefited in the last years from emerging accelerators that increase the performance of scientific simulations, but using these devices hinders the programming task. This paper presents AMA: a set of optimization techniques to efficiently manage multi-accelerator systems. AMA maximizes the overlap of computation and communication in a blocking-free way. Then, we can use such spare time to do other work while waiting for device operations. Implemented on top of a task-based framework, the experimental evaluation of AMA on a quad-GPU node shows that we reach the performance of a hand-tuned native CUDA code, with the advantage of fully hiding the device management. In addition, we obtain up to more than 2x performance speed-up with respect to the original framework implementation.