Reducing Quantization Errors for Inner-Product Operations in Embedded Digital Signal Processing Systems [Tips&Tricks]

摘要

Inner-product operations are used extensively in embedded digital signal processing (DSP) systems. Their applications range from signal processing (filtering/convolution) to inference (classification). In embedded systems, resources (energy, area, etc.) are typically highly constrained, making tradeoffs with computational precision a fundamental concern. Indeed, with increasing requirements on algorithmic performance, many systems are trending toward higher computational precision to ensure accuracy of results. This article describes an approach to significantly enhance accuracy for inner-product operations at very low bit precisions, significantly improving the energy/area-versus-accuracy tradeoffs traditionally incurred.