The growing popularity of powerful mobile devices such as smart phones and tablet devices has resulted in the exponential growth of demand for video applications. However, due to the intensive computation of the video coding and display process, mobile video applications require frequent memory access, which consumes a large amount of power and limits battery life. Various low-power memory techniques have been explored by researchers. Unfortunately, existing solutions suffer from shortcomings, including high design complexity and significant implementation overhead. 

This project explores and characterizes the behaviors of video data and develops new power-efficient mobile video memory designs that utilize the identified video data characteristics. For example, we have applied association rule mining techniques and identified interesting data patterns existing in video data. Based on the obtained data patterns, we implemented a low-cost Data Driven power-efficient Adaptable SRAM Hardware (D-DASH) design with dynamic power-quality tradeoff for mobile video applications. D-DASH enables three levels of power-quality management (up to 43.7% power savings) with negligible area overhead (0.06%). This work was presented at International Symposium on Low Power Electronics and Design (ISLPED'16) and received a best paper nomination in the conference. Furthermore, by utilizing the discovered data correlation and association relationship in video chroma data, we presented another data-informed video memory with self-recovery ability in IEEE International Conference on Computer Design (ICCD’16). This technique successfully delivers good video quality for minimum-sized SRAM at near-threshold voltage (0.5V). Based on it, we extended the data-correlation/association-enabled recovery to video luma data, further enabling power savings. We developed a two-dimensional data pattern approach to explore horizontal data association and vertical data correlation characteristics. Such data relationship discovery and pattern identification enable a new dimension for the hardware design space and bring self-recovery ability to memories in the presence of bitcell failures. This work was published in IEEE Trans. on Big Data. To summarize, we have published a total of 14 peer-reviewed papers, including 4 IEEE journal papers and 10 conference papers.

Intellectual Merit: The exploration of these intelligent low power techniques through the interaction of both hardware and software viewpoints enables a new dimension for power savings. By introducing data mining to the hardware design process, the discovered data knowledge, combined with novel data-driven hardware design techniques, enable more intelligent video memory with a better trade-off among energy efficiency and cost, thereby relieving the current data storage burden for today’s video systems. Also, the research outcomes have great potential to extend to other data-intensive applications such as deep learning systems. More broadly, our cross-disciplinary research in integrating data-mining and hardware design can lead to significant advances in software/hardware co-design theory and practice.

Broad Impacts: The memory techniques developed by this project have increased the energy efficiency of mobile video systems, accelerating the creation of low-cost mobile devices with a longer battery life. This project serves as an excellent educational platform to improve the understanding of green computing amongst future computer scientists and computer engineers. The PIs have jointly developed a graduate-level course focusing on software/hardware co-design for mobile devices. The project also had an enormous impact of student engagement from different disciplines, particularly from computer science and computer engineering fields. The PIs continued their efforts to recruit underrepresented students, such as females and minorities, to work on this project. Special activities that demonstrate efficient hardware design have been conducted during the organized K-12 programs to girls and Native American high-school students

Last Modified: 07/28/2018
Modified by: Na Gong