Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Compression of Site-Specific Deep Neural Networks for Massive MIMO Precoding

Ghazal Kasalaee, Ali Hasanzadeh Karkan, Jean-François Frigon, François Leduc-Primeau

TL;DR

The paper tackles the high energy and memory demands of DL-based precoding for massive MIMO by introducing a mixed-precision quantization framework and neural-architecture search (NAS) guided by site-specific ray-tracing data. The proposed approach searches over per-layer bit-widths and architectural sizes to produce compressed DL precoders that achieve comparable sum-rate to traditional methods with substantially lower energy consumption, up to 35x over WMMSE in challenging environments. Self-supervised training with a sum-rate objective and a detailed energy model for both quantized networks and baselines underpins the quantitative results. Overall, the work provides a practical benchmark and demonstrates that site-aware, quantization-aware DL precoding can deliver significant energy efficiency gains for real-time mMIMO beamforming without sacrificing performance.

Abstract

The deployment of deep learning (DL) models for precoding in massive multiple-input multiple-output (mMIMO) systems is often constrained by high computational demands and energy consumption. In this paper, we investigate the compute energy efficiency of mMIMO precoders using DL-based approaches, comparing them to conventional methods such as zero forcing and weighted minimum mean square error (WMMSE). Our energy consumption model accounts for both memory access and calculation energy within DL accelerators. We propose a framework that incorporates mixed-precision quantization-aware training and neural architecture search to reduce energy usage without compromising accuracy. Using a ray-tracing dataset covering various base station sites, we analyze how site-specific conditions affect the energy efficiency of compressed models. Our results show that deep neural network compression generates precoders with up to 35 times higher energy efficiency than WMMSE at equal performance, depending on the scenario and the desired rate. These results establish a foundation and a benchmark for the development of energy-efficient DL-based mMIMO precoders.

Compression of Site-Specific Deep Neural Networks for Massive MIMO Precoding

TL;DR

The paper tackles the high energy and memory demands of DL-based precoding for massive MIMO by introducing a mixed-precision quantization framework and neural-architecture search (NAS) guided by site-specific ray-tracing data. The proposed approach searches over per-layer bit-widths and architectural sizes to produce compressed DL precoders that achieve comparable sum-rate to traditional methods with substantially lower energy consumption, up to 35x over WMMSE in challenging environments. Self-supervised training with a sum-rate objective and a detailed energy model for both quantized networks and baselines underpins the quantitative results. Overall, the work provides a practical benchmark and demonstrates that site-aware, quantization-aware DL precoding can deliver significant energy efficiency gains for real-time mMIMO beamforming without sacrificing performance.

Abstract

The deployment of deep learning (DL) models for precoding in massive multiple-input multiple-output (mMIMO) systems is often constrained by high computational demands and energy consumption. In this paper, we investigate the compute energy efficiency of mMIMO precoders using DL-based approaches, comparing them to conventional methods such as zero forcing and weighted minimum mean square error (WMMSE). Our energy consumption model accounts for both memory access and calculation energy within DL accelerators. We propose a framework that incorporates mixed-precision quantization-aware training and neural architecture search to reduce energy usage without compromising accuracy. Using a ray-tracing dataset covering various base station sites, we analyze how site-specific conditions affect the energy efficiency of compressed models. Our results show that deep neural network compression generates precoders with up to 35 times higher energy efficiency than WMMSE at equal performance, depending on the scenario and the desired rate. These results establish a foundation and a benchmark for the development of energy-efficient DL-based mMIMO precoders.

Paper Structure

This paper contains 16 sections, 13 equations, 3 figures, 2 tables.

Table of Contents

  1. Introduction
  2. System model
  3. Problem Definition
  4. Methodology
  5. Model Architecture and NAS
  6. Quantization Methods
  7. with
  8. Training Method
  9. Energy Model for Quantized Neural Networks
  10. Energy Model for Baseline Methods
  11. Numerical Results
  12. Dataset Definition
  13. Energy Consumption Examples
  14. NAS Results
  15. Impact of Deployment Environment and Energy Efficiency Comparison
  16. ...and 1 more sections

Figures (3)

  • Figure 1: DNN architecture template.
  • Figure 2: Trade-off between energy efficiency and sum rate for with varying $C_{\text{out}}$, $D_{\text{FCL}}$, and MPQ bit widths, on the "UdeM-LOS" scenario ($N_{\sf{T}}$$=$$64$, $N_{\sf{U}}$$=$$4$, average SNR $=$ 29 dB). All the model configurations that were evaluated are shown, while the curves provide the Pareto front associated with each architecture configuration.
  • Figure 3: Comparison of energy efficiency (bits/s/Hz/$\mu$J) and sum rate (bits/s/Hz) across two environments: UdeM-NLOS (average SNR $=$ 15 dB) and Okapark-LOS (average SNR $=$ 28 dB). The proposed method achieves a superior balance of energy efficiency and sum rate performance compared to WMMSE. Results are derived for models with varying ($C_{out}$) and ($D_{FCL}$).