Enabling Approximate Storage through Lossy Media Data Compression

Memory capacity, bandwidth, and energy all continue to present hurdles in the quest for efficient, high-speed computing. Recognition, mining, and synthesis (RMS) applications in particular are limited by the efficiency of the memory subsystem due to their large datasets and need to frequently access memory. RMS applications, such as those in machine learning, deliver intelligent analysis and decision making through their ability to learn, identify, and create complex data models. To meet growing demand for RMS application deployment in battery constrained devices, such as mobile and Internet-of-Things, designers will need novel techniques to improve system energy consumption and performance. Fortunately, many RMS applications demonstrate inherent error resilience, a property that allows them to produce acceptable outputs even when data used in computation contain errors. Approximate storage techniques across circuits, architectures, and algorithms exploit this property to improve the energy consumption and performance of the memory subsystem through quality-energy scaling. This thesis reviews state of the art techniques in approximate storage and presents our own contribution that uses lossy compression to reduce the storage cost of media data.