Quantization Noise Power Estimation for Floating-Point DSP Circuits

Abstract

In this brief, we present a semianalytical model of the quantization noise power of floating-point DSP circuits, considering heterogeneous precisions. The use of hardware operators with optimized precisions has proven to provide important cost reductions. However, precision optimization is a time-consuming task, and fast and accurate error estimators are required. Moreover, the use of the signal-to-quantization noise ratio (SQNR) as a quality reference is common in DSP design, and there are no proper models to perform fast estimation in the context of floating-point systems with heterogeneous precision. The model presented here accounts for the quantization produced when the size of the mantissa of a floating-point signal is modified along the datapath and it has negligible dependence with the signal statistics. In addition, the model is appropriate for fast SQNR estimations and can be integrated in current quantization optimizers. The results show that the semianalytical model has an estimation error of less than 2% when compared to a simulation-based reference, whereas other approaches introduce estimation errors of up to 52%.