Energy-efficient computing for mobile signal processing

Mobile devices have rapidly proliferated, and deployment of handheld devices continues to increase at a spectacular rate. As today's devices not only support advanced signal processing of wireless communication data but also provide rich sets of applications, contemporary mobile computing requires both demanding computation and efficiency. Most mobile processors combine general-purpose processors, digital signal processors, and hardwired application-specific integrated circuits to satisfy their high-performance and low-power requirements. However, such a heterogeneous platform is inefficient in area, power and programmability. Improving the efficiency of programmable mobile systems is a critical challenge and an active area of computer systems research.;SIMD (single instruction multiple data) architectures are very effective for data-level-parallelism intense algorithms in mobile signal processing. However, new characteristics of advanced wireless/multimedia algorithms require architectural re-evaluation to achieve better energy efficiency. Therefore, fourth generation wireless protocol and high definition mobile video algorithms are analyzed to enhance a wide-SIMD architecture. The key enhancements include 1) programmable crossbar to support complex data alignment, 2) SIMD partitioning to support fine-grain SIMD computation, and 3) fused operation to support accelerating frequently used instruction pairs.;Near-threshold computation has been attractive in low-power architecture research because it balances performance and power. To further improve energy efficiency in mobile computing, near-threshold computation is applied to a wide SIMD architecture. This proposed near-threshold wide SIMD architecture---Diet SODA---presents interesting architectural design decisions such as 1) very wide SIMD datapath to compensate for degraded performance induced by near-threshold computation and 2) scatter-gather data prefetcher to exploit large latency gap between memory and the SIMD datapath. Although near-threshold computation provides excellent energy efficiency, it suffers from increased delay variations. A systematic study of delay variations in near-threshold computing is performed and simple techniques---structural duplication and voltage/frequency margining---are explored to tolerate and mitigate the delay variations in near-threshold wide SIMD architectures.;This dissertation analyzes representative wireless/multimedia mobile signal processing algorithms, proposes an energy-efficient programmable platform, and evaluates performance and power. A main theme of this dissertation is that the performance and efficiency of programmable embedded systems can be significantly improved with a combination of parallel SIMD and near-threshold computations.