Energy efficiency vs. programmability trade-off: architectures and design principles

Performance achievements on programmable architectures due to process technology are reaching their limits, since designs are becoming wire- and power-limited rather than device limited. Likewise, traditional exploitation of instruction level parallelism saturates as the conventional approach for designing wider issue machines leads to very expensive interconnections, big instruction memory footprint and high register file pressure. New architectural concepts targeted to the application domain of media processing are needed in order to push current state-of-the-art limitations. To this end, we regard media applications as a collection of tasks which consume and produce chunks of data. The exploitation of task level parallelism as well as more traditional forms of parallelism is a key issue for achieving the required amount of MOPS/Watt and MOPS/mm² for media applications. Tasks comprise data transfers and number crunching algorithm kernels, which are very computing-intensive yet highly predictable. Moreover, most of the data manipulated by a task is of a local nature. Granularity and characteristics of these tasks will lead us in this paper to draw conclusions about memory hierarchy, task scheduling strategies and efficient low-overhead programmable architectures for highly predictable kernel computations