Data-Free Quantization of Neural Receivers: When 4-Bit Succeeds, Why 6-Bit Matters for 6G

SaiKrishna S. Yellapragada, Esa Ollila, Mário Costa

Keywords: Neural Network Receivers, Post training quantization, deep learning, efficient machine learning, AI for RAN

TL;DR: 8,6 and 4 bit weight quantization for inference is a perfect fit varying channel and mobility conditions for deep learning receivers.

Abstract: As wireless systems evolve toward 6G, Artificial Intelligence (AI) and Deep Learning (DL) are poised to revolutionize physical-layer processing, offering superior performance over classical methods in throughput and Block Error Rate (BLER). Deploying DL-based receivers in resource-constrained environments requires balancing performance with inference latency, energy consumption, and computational overhead. We study data-free Post-Training Quantization (PTQ) of a neural receiver that processes frequency-domain baseband samples to generate Log-Likelihood Ratios (LLRs) for error-control decoding. Quantization parameters are derived directly from pretrained weights via symmetric per-channel uniform quantization, where each channel’s scale captures the absolute-weight range requiring no calibration data, synthetic data, or activation statistics. We reduce float32 weights to 8-, 6-, and 4-bit and evaluate radio performance across 3GPP Line-of-Sight (LoS)/Non-LoS (NLoS) channels and mobility scenarios. In NLoS, 8- and 6-bit achieve near-float32 BLER, with gains up to 4.9 dB over baseline Least-Squares (LS) under high mobility. In LoS, 4-bit remains robust, surpassing traditional receivers by 1.7–2.6 dB across mobilities, while yielding an 8× smaller model. These findings inform hardware–software co-design for AI native 6G air-interfaces, highlighting low-precision quantization as a key enabler for efficient edge, sensing, and cloud-radio deployments.

Submission Number: 16