Caching Techniques for Multi-Processor Streaming Architectures

In the world of complex SoCs for consumer applications, multiprocessor architectures usually deploy caching techniques to alleviate the cost of data communication between processing elements. In this application domain, the characteristics of streaming applications play a dominant role in the design of the multiprocessor architectures. These characteristics not only influence the design at SoC level, but also permeate the design of lower level blocks, such as caches. This paper proposes three novel techniques for data caching in multiprocessor streaming architectures. These techniques exploit the combination of mechanisms from the domains of stream caching and cache coherency, optimizing the effectiveness of data caches in a streaming multiprocessor context. Our first implementation targets cache systems in a multiprocessor architecture for high-definition MPEG decoding, where six function-specific processors communicate through shared embedded SRAM. Simulation results demonstrate the effectiveness of the presented techniques, even though the caches are of very small dimensions, i.e., 0.12 mm2 in 0.12µ technology.