Bandwidth reduction for video processing in consumer systems

The architecture of the present video processing units in consumer systems is usually based on various forms of processor hardware, communicating with an off-chip SDRAM memory. Examples of these systems are currently available MPEG encoders and decoders, and high-end television systems. Due to the fast increase of required computational power of consumer systems, the data communication to and from the off-chip memory has become the bottleneck in the overall system performance (memory wall problem). This paper presents a strategy for mapping pixels into the memory for video applications such as MPEG processing, thereby minimizing the transfer overhead between memory and the processing. A novelty in our approach is that the proposed communication model considers the statistics of the application-dependent data accesses in memory. With this technique, a 26% reduction of the memory bandwidth was obtained in an MPEG decoding system containing a 64-bit wide memory bus. For double-data-rate SDRAM (DDR SDRAM), the proposed mapping strategy reduces the bandwidth in the system by even 50%. This substantial performance improvement can readily be used for extending the quality or the functionality of the system