Mobile devices supporting a wide variety of multimedia applications under a very stringent energy budget are a key driver of electronic systems in future. The objective of the proposed research is to explore a generic, programmable, reconfigurable, and energy-efficient architectural platform for future mobile devices for real-time multimedia processing. The aim is to exploit the inherent error tolerance in multimedia applications for run-time energy-accuracy trade-off. The proposed research will analyze the interaction of ultra-low-power computing and error characteristics of real-time multimedia processing. This research will pursue a circuit-architecture-algorithm co-design approach to model, analyze, and demonstrate a reconfigurable hardware platform for memories, datapath, and buses to exploit the error characteristics of real-time multimedia processing algorithms for ultra-low power.
This generic architecture for energy-efficient multimedia systems can lay the foundation of future mobile supercomputers performing wide array of applications with minimal energy. The PIs will disseminate the research results through project website, conference and journal publications. The existing interactions with leading microprocessor and mobile handset manufacturers will provide opportunities for technology transfer. The educational goal is to create next generation engineers who understand the effects of energy and accuracy on computations. The PIs will engage in recruiting students from underrepresented groups and mentor students under the Summer Undergraduate Research Experience for minorities (SURE) program at Georgia Tech.
Modern mobile electronics need energy in the form of batteries to operate. Longer battery life by reducing energy consumption of the electronic components is key goals for the designer of cell phones and many other mobile devices. This grant has explored new ways to design low-energy mobile platforms, particularly those that manipulate images, by taking advantage of the fact that small errors in the image may not be visible to the human eye. The grant exploits the fundamental interactions between energy and error that dictates reducing error rates in digital information processing generally requires more energy. We developed new hardware and software techniques that take advantage of the nature of visual perception to reduce energy dissipation. The developed methods operate the electronic circuits at much lower voltage and avoid certain operations or storage that aren't strictly necessary. The techniques help save energy and extend battery life with a graceful degradation in the perceived quality of the image. The techniques are demonstrated through hardware design and measurement and software analysis to ensure resilient system operation at reduced energy. The knowledge developed through the project is disseminated to the broader community including students, academics, and practicing engineers through scientific articles, tutorials in major conferences, and seminars in various computing industry as well as included in electrical and computer engineering courses at Georgia Tech thereby integrating research and education. Our contributions to undergraduate education include engaging multiple undergraduates in research related to this topic and integrating energy-related topics into existing as well as a new course, ECE 3030, Physical Foundations of Computer Engineering. The outcomes will fuel new innovations in algorithms for image/video processing that can take advantage of the ultra-low-energy hardware platform developed in this project. The achievements of this project, coupled with new image processing algorithms, can transform the field for consumer electronics like smart phones where image/video processing is a major task. We envision a new generation of intelligent devices that can dynamically change the operating modes to reduce energy dissipation and extend battery life with graceful degradation in image/video quality.
