Input-Adaptive and Quality-Configurable Approximate Computing: A Full-System Perspective

Approximate computing is an emerging design paradigm that leverages the intrinsic resilience of applications to execute computations approximately, and more efficiently, leading to improvements in energy consumption or performance. A key challenge in approximate computing is that the extent to which computations can be approximated varies significantly from application to application, as well as across inputs for a single application. This makes quality-configurability (modulation of the energy vs. quality trade-off of applications at runtime) and input-adaptivity (tuning the degree of approximation based on individual inputs) essential for obtaining significant energy savings while retaining output quality at acceptable levels. The first part of this research explores input-adaptive and quality-configurable approximate computing at various layers of design abstraction. Most prior work in approximate computing deals with approximate computation only. However, in a typical embedded system, computation is only one component of the entire system. In addition to the computation subsystem, we also have other subsystems such as memory, sensing, and communication. To maximize the energy benefits, it is essential to focus not only on approximate computation but also investigate approximations in these other subsystems. Towards this goal, the second part of this research first explores approximate memory by proposing a systematic methodology to construct a quality configurable approximate DRAM. Finally, this thesis proposes a full-system approximation methodology that synergistically tunes the different subsystem-level approximation knobs to maximize system-level energy savings for a specified target quality constraint. As an embodiment of this principle, a case-study of an approximate smart-camera system is presented and demonstrated