Sensitivity-Guided Mixed-Precision Floating-Point Hardware for Digital Predistortion

Paper i proceeding, 2026

Short floating-point formats can reduce arithmetic and storage cost in digital predistortion (DPD) hardware. However, using a uniform precision across the datapath ignores the different error sensitivities of individual stages. This work develops a first-order rounding sensitivity analysis based on the actual RTL rounding points and demonstrates it on a fixed 12-basis datapath using basis-propagating selection (BAPS). Numerical profiling provides a reduced-precision anchor, while the sensitivity analysis guides the relative mantissa allocation across the datapath. Based on this analysis, we implement a sensitivity-guided mixed-precision (SG-MP) design. Real power-amplifier measurements give an average normalized mean-square error (NMSE) of -36.466 dB, with a 0.119-dB gap to the 32-bit single-precision reference and the best average NMSE among the evaluated low-precision configurations. At iso-performance, SGMP reduces synthesized area by around 11% and 78% relative to the uniform low-precision baseline and the 32-bit single-precision reference, respectively