Deep neural network (DNN) models dominate state-ofthe- art artificial intelligence. Due to their high computational cost, the use of hardware accelerators is key in any system that trains or deploys deep learning models. GPUs are widely used for executing neural networks, but they are too power hungry for many applications, especially in edge computing. In the last years, a wide variety of deep learning-specific accelerators have been proposed that greatly outperform GPUs in terms of energy efficiency, but even those accelerators are far from the desirable efficiency for many applications. In this extended abstract, we present two accelerator architectures we have designed in order to experiment with different energy efficiency maximization ideas.