Due to increasing complexity of modern real-time image processing applications, classical hardware development at register transfer level becomes more and more the bottleneck of technological progress. Modeling those applications by help of multi-dimensional data flow and providing efficient means for their synthesis in hardware is one possibility to alleviate the situation. The key element of such descriptions is a multi-dimensional FIFO whose hardware synthesis shall be investigated in this paper. In particular, it considers the occurring out-of-order communication and proposes an architecture which is able to handle both address generation and flow control in an efficient manner. The resulting implementation allows reading and writing one pixel per clock cycle with an operation frequency of up to 300 MHz. This is even sufficient to process very huge images occurring in the domain of digital cinema in real-time.