Dynamic Precision Analog Computing for Neural Networks

Analog electronic and optical computing exhibit tremendous advantages over digital for accelerating deep learning when operations are executed at low precision. Although architectures support programmable precision to increase efficiency, analog today only a single, static In this work, we characterize the relationship between effective number of bits (ENOB) processors, which is limited by noise, bit quantized neural networks. We propose extending dynamic levels repeating averaging result, decreasing impact noise. To utilize precision, method each layer pre-trained model without retraining network weights. evaluate on subject shot thermal weight noise find that employing reduces energy consumption up 89% computer vision models such as Resnet50 24% natural language processing BERT. one example, apply shot-noise homodyne simulate inference an 2.7 aJ/MAC 1.6 BERT with ${< }2\%$ accuracy degradation, implying unlikely be dominant cost.