Hearing aid and Extreme Edge IoT Acceleration

Abstract

As commented on previously [1], IoT processing directly in edge devices is becoming increasingly necessary and advantageous, providing a number of distinct advantages over cloud based computation. Provided the edge device has sufficient resources, computation is not dependent on external (cloud) resources. Depending on the application or deployment, these external resources might be non-existent, scarce, unreliable, or overly power intensive for ongoing communication with the edge device for farming out part of the processing. Independent, isolated computation can also be beneficial to mitigate security concerns. Edge computing is local and scaled to the recognition effort required, yielding a much more efficient and responsive system. Local processing eliminates transmission power, facilitates accurate and quick environment sensing and assessment enabling advanced algorithms to take corrective action quickly. The remaining challenge is, of course, fitting the recognition system within the constraints of the given edge device. Further progress in this field has yielded preliminary results of a tiny accelerator for extreme edge devices. The procedure and experiment using a new standardized benchmark – EEMBC will be described in this paper and compared to the general computation approach.