Approximate computing: Enabling voltage over-scaling in multimedia applications

The demand for power efficient design has increased exponentially with the advent of smartphones capable of supporting multimedia services in mobile communications. Power hungry applications such as web browsing, playing video and listening to music reduce the battery lifespan of these devices significantly. Recently, supply voltage over-scaling has emerged as one of the most effective techniques for reducing power in digital designs due to the super linear dependence of both dynamic and leakage power on supply voltage. Voltage over-scaling differs from traditional voltage scaling in that we do not scale the clock frequency, thereby intentionally causing the critical paths to violate the clock period. Furthermore, in addition to delay failures in logic sections, aggressive voltage over-scaling significantly degrades the stability of memory arrays in presence of process parameter variations in scaled technologies. To combat the detrimental effects of voltage over-scaling we resort to what we term as the Approximate Computing design methodology. The fundamental premise of our approach lies in the fact that all datapath computations and bits of stored data are not equally significant in shaping the output response of multimedia systems. Approximate computing leverages this key observation to advocate a priority based design approach for logic and memory blocks based on a variable latency elastic clocking technique and unequal error protection based hybrid memory, respectively. Under aggressive voltage over-scaling in presence of parametric variations, the elastic clocking technique prevents delay failures in important datapath computations while the hybrid memory approach protects the important bits of stored data in a preferential manner. With the aid of these two techniques, approximate computing enables voltage over-scaling in multimedia hardware, obtaining substantial power savings by allowing graceful degradation in output quality since the important computations and data bits remain unaffected. Approximate computing applied to motion estimation in MPEG encoder shows large power benefits at reasonable video quality while tolerating errors induced by voltage over-scaling and process variations. In addition, this work also explores approximate computing in the context of emerging recognition and data mining applications which manifest significantly higher degree of inherent error resiliency, demonstrating its viability in the future