Ultra Low Power Inference at the Very Edge of the Network