Exploiting the Approximate Computing Paradigm with DNN Hardware Accelerators

The implementation of hardware accelerators for DNNs may benefit in terms of energy and performance from the application of Approximate Computing (AC) techniques. For example, the adoption of approximate multipliers with their inexact but energy-efficiently computed results can positively impact the figures of hardware accelerators, especially given the tremendous amount of multiply-and-accumulate (MAC) operations required by DNNs, which by their nature and to some extent, are resilient to errors. This paper presents an analysis of the application of AC techniques to all the subsystems of a hardware accelerator, namely, the computation, communication, and memory subsystems, evaluating the trade-offs in terms of energy vs. accuracy obtained in the inference phase.