Energy-aware embedded media processing: Customizable memory subsystems and energy management policies

The design of energy-efficient data memory architectures for embedded system platforms has received considerable attention in recent years. In this dissertation we propose a special-purpose data memory subsystem, called Xtream-Fit, targeted to streaming media applications executing on both generic; uniprocessor embedded platforms and powerful SMT-based multi-threading platforms. We empirically demonstrate that Xtream-Fit achieves high energy-delay efficiency across a wide range of media devices, from systems running a single media application to systems concurrently executing multiple media applications under synchronization constraints. Xtream-Fit's energy efficiency, is predicated on a novel task-based execution model that exposes/enhances opportunities for efficient prefetching, and aggressive dynamic energy conservation techniques targeting on-chip and off-chip memory components. A key novelty of Xtream-Fit is that it exposes a single customization parameter, thus enabling a very simple and yet effective design space exploration methodology to find the best memory configuration for the target application(s). Extensive experimental results show that Xtream-Fit reduces energy-delay product substantially---by 32% to 69%---as compared to 'standard' general-purpose memory subsystems enhanced with state of the art cache decay and SDRAM power mode control policies.