EQSCALE: Energy-quality scalable feature extraction engine for Sub-mW real-time video processing with 0.55 mm2 area in 40nm CMOS

An energy-quality scalable (EQSCALE) feature extraction accelerator for IoT vision applications is presented. Knobs are introduced to dynamically adjust the tradeoff between energy and feature extraction quality, leveraging the intrinsic redundancy in video frames and the robustness of object recognition against missing features. Measurements of a testchip in 40nm show 310pJ/pixel energy at nominal voltage and maximum quality, and 55.6pJ/pixel when tuned for minimum energy, while still allowing correct object recognition. The active area of the accelerator is 0.55mm
²
. EQSCALE enables at least 5.7X energy improvement and 1.8X area reduction over state-of-the-art accelerators. To the best of our knowledge, EQSCALE is the first feature extraction accelerator operating in the sub-mW range (0.51mW) at VGA resolution and 30 fps.