Основные понятия
This work introduces mixed-precision neural networks that employ quantized low-precision fixed-point parameters to reduce the computational complexity and memory footprint of digital predistortion (DPD) for wideband power amplifiers, thereby lowering power consumption without compromising linearization efficacy.
Аннотация
This article presents a novel approach to implementing mixed-precision (MP) arithmetic operations and model parameters in a gated Recurrent Neural Network (RNN)-based Digital Pre-distortion (DPD) model for wideband power amplifiers (PAs). The proposed method aims to curtail the DPD model inference power consumption by substituting most high-precision floating-point operations with low-precision fixed-point operations through quantizing neural network weights (W) and activations (A).
The key highlights and insights are:
The substantial computational complexity and memory requirements of machine learning-based DPD systems, especially those using deep neural networks (DNNs), pose significant obstacles to their efficient deployment in wideband transmitters, particularly in the context of future 5.5G/6G base stations or Wi-Fi 7 routers, where limited power resources constrain real-time DPD model computation.
The authors adopt a mixed-precision strategy utilizing low-precision fixed-point integer arithmetic for inference to enhance the energy efficiency of DPD models. This method involves a quantization scheme that converts the model's weights and activations, including other intermediate variables, to lower precision while retaining full-precision operations for feature extraction from the I/Q signal.
Experimental results show that the proposed W16A16-GRU DPD model achieves no performance loss against 32-bit floating-point precision DPDs, while attaining -43.75 (L)/-45.27 (R) dBc in Adjacent Channel Power Ratio (ACPR) and -38.72 dB in Error Vector Magnitude (EVM). A 16-bit fixed-point-precision MP-DPD enables a 2.8× reduction in estimated inference power consumption compared to the full-precision baseline.
The authors also demonstrate that lower-precision configurations, such as W8A8 and W12A12, can achieve up to 4.5× and 3.8× power reduction, respectively, at the expense of some linearization performance.
The proposed MP-DPD approach is compatible with existing strategies, allowing for further power savings when combined, making it a promising solution for energy-efficient wideband transmitters in power-sensitive environments.
Статистика
The DPD input I/Q data sample rate is 640 MHz.
The PAPR of the test signal is 10.38 dB.
The DPA outputs at 13.75 dBm.
Цитаты
"As bandwidths in future radio systems expand, the energy demands of DPD computation intensify."
"Utilizing 32-bit floating-point (FP32) arithmetic, while beneficial for accuracy, can increase model size, negatively impacting energy efficiency."
"The energy consumption of on-chip Static Random Access Memory (SRAM) is up to 12.2× higher than that of a MAC operation. Moreover, the energy costs for off-chip memory access are roughly three orders of magnitude greater than for arithmetic operations."