Energy-efficient system design for mobile processing platforms

[摘要] Portable electronics has fueled the rich emergence of multimedia applications that have led to the exponential growth in content creation and consumption. New energy-efficient integrated circuits and systems are necessary to enable the increasingly complex augmented-reality applications, such as high-performance multimedia, ;;big-data;; processing and smart healthcare, in real-time on mobile platforms of the future. This thesis presents an energy-efficient system design approach with algorithm, architecture and circuit co-design for multiple application areas. A shared transform engine, capable of supporting multiple video coding standards in real-time with ultra-low power consumption, is developed. The transform engine, implemented using 45 nm CMOS technology, supports Quad Full-HD (4k x 2k) video coding with reconfigurable processing for H.264 and VC-1 standards at 0.5 V and operates down to 0.3 V to maximize energy-efficiency. Algorithmic and architectural optimizations, including matrix factorization, transpose memory elimination and data dependent processing, achieve significant savings in area and power consumption. A reconfigurable processor for computational photography is presented. An efficient implementation of the 3D bilateral grid structure supports a wide range of non-linear filtering applications, including high dynamic range imaging, low-light enhancement and glare reduction. The processor, implemented using 40 nm CMOS technology, enables real-time processing of HD images, while operating down to 0.5 V and achieving 280x higher energy-efficiency compared to software implementations on state-of-the-art mobile processors. A scalable architecture enables 8x energy scalability for the same throughput performance, while trading-off output resolution for energy. Widespread use of medical imaging techniques has been limited by factors such as size, weight, cost and complex user interface. A portable medical imaging platform for accurate objective quantification of skin condition progression, using robust computer vision techniques, is presented. Clinical validation shows 95% accuracy in progression assessment. Algorithmic optimizations, reducing the memory bandwidth and computational complexity by over 80%, pave the way for energy-efficient hardware implementation to enable real-time portable medical imaging.

[发布机构] Massachusetts Institute of Technology