AI computing reaches for the edge
A computing platform called “NorthPole” is described that facilitates high inference speed and prediction accuracy but with a moderate energy requirement, a promising step toward chip designs that support low-power edge AI inference.
Abstract
A chip design integrates computation and memory to efficiently process data at low energy cost Artificial intelligence (AI)—the ability of computers to perform human cognitive functions in real-world scenarios—requires substantial computation power, energy, and vast datasets. Once trained, AI models are deployed to make predictions (inferences) for new situations. In the traditional paradigm, inference is supported by centralized, high-performance computational platforms and high-bandwidth network connections. However, in real-time mission-critical applications such as facial recognition, object detection tracking, and behavior monitoring, an “edge” computing system is desirable, for fast and accurate inference and, hence, fast response times. Edge computing requires moving the large AI model from a centralized location to a position closer to the source of data (hence, working at the edge). On page 329 of this issue, Modha et al. (1) describe a computing platform called “NorthPole” that facilitates high inference speed and prediction accuracy but with a moderate energy requirement. This is a promising step toward chip designs that support low-power edge AI inference.