Mixed-Signal Computing for Deep Neural Network Inference
Boris Murmann
Abstract
Modern deep neural networks (DNNs) require billions of multiply-accumulate operations per inference. Given that these computations demand relatively low precision, it is feasible to consider analog computing, which can be more efficient than digital in the low-SNR regime. This overview article investigates the potential of mixed analog/digital computing approaches in the context of modern DNN processor architectures, which are typically limited by memory access. We discuss how memory-like and in-memory compute fabrics may help alleviate this bottleneck and derive asymptotic efficiency limits at the processing array level. It is shown that single-digit fJ/op energy efficiencies are feasible for 4-bit mixed-signal arithmetic. In this analysis, special consideration is given to the SNR and amortization requirements of the analog-digital interfaces. In addition, we consider the pros and cons for a variety of implementation styles and highlight the challenge of retaining high compute efficiency for a complete DNN accelerator design.