Ecole Polytechnique Federale de Lausanne
In this paper, the authors advocate the placement of Architecturally Visible Communication (AVC) buffers between adjacent cores in MPSoCs to provide high-throughput communication for streaming applications. Producer/consumer relationships map poorly onto cache-based MPSoCs. Instead, they instantiate application specific AVC buffers on top of a distributed consistent and coherent cache-based system with shared main memory to provide the desired functionality. Using JPEG compression as a case study, they show that the use of AVC buffers in conjunction with parallel execution via heterogeneous software pipelining provides a speedup of as much as 4.2x compared to a baseline single processor system, with an increase in estimated memory energy consumption of only 1.6x.