Empirical analysis of operand usage and transport in multimedia applications

As transistor feature sizes decrease exponentially, the critical problem in massively parallel architectures for multimedia systems becomes the wire delay latency imposed by transporting operands. In this paper, we first characterize the locality properties of operands in multimedia applications to motivate the development of alternate low-cost and low-latency operand communication mechanisms. Based on the locality properties, we then present the results of a simulation study that estimates the impact of varying sized local storage on the transport complexity of the operands. A small amount of local storage (eight-entry) per functional unit with approximate information on operand dynamic lifetimes is sufficient to suppress 79.4% of operand writes in MediaBench applications. The empirical study shows that local storage alone reduces operand read traffic by up to 35.5%.