Source-linked AI summary
Floating-Point Microformat Quantization and Pruning for Efficient MU-MIMO Neural Receivers
SaiKrishna Saketh Yellapragada, Esa Ollila, Mário Costa, Yawei Li
TL;DR
Neural receivers’ compute and memory demands hinder real-time deployment, and evidence is limited on whether 4-bit formats preserve performance in standard-compliant MU-MIMO reception. The paper compares quantization-aware weight and activation training with separate 50% pruning, finding FP4 more effective than INT4 and projecting lower inference costs for pruned 4-bit weights.
Problem
Whether floating-point microformats preserve 4-bit performance in a standard-compliant MU-MIMO receiver with INT8 activations or pruning remains untested, amid neural receivers’ compute and memory demands.
Source: §1 · p. 1Method
The study applies weight and activation quantization-aware training across INT8, INT4, FP8, and FP4 formats, deriving INT8-activation and 50%-pruned models separately from weight-only checkpoints.
Source: §1 · p. 1, §4 · p. 2, §3.3 · p. 2Results
FP4 outperforms LS–LMMSE while INT4 performs poorly; analytic modeling projects 66× fewer bit-operations and 8.8× less weight storage for weight-activation quantization followed by pruning.
Source: §4.2 · p. 3, §5 · p. 4Takeaways & Limitations
The reported comparison supports microscaling formats for neural receivers, with FP4 outperforming LS–LMMSE while INT4 performs poorly.
Source: §5 · p. 4LS–LMMSE and LMMSE–K-Best use floating-point arithmetic, and fixed-point arithmetic may increase BLER through additional quantization error.
Source: §4.1 · p. 3
Abstract
from arXiv · showhide
Neural receivers outperform conventional 5G NR processing chains, but their compute and memory demands hinder real-time deployment. For a standard-compliant multi-user MIMO neural receiver, the 4-bit number format, not merely the bit width, determines whether compression preserves the gain over classical receivers. We apply weight and activation quantization-aware training (QAT) and, separately, 50% magnitude pruning, comparing INT8/INT4 and FP8 (E4M3)/FP4 (E2M1) weights with INT8 post-ReLU activations. Trained on 3GPP UMi channels and evaluated on TDL-B and TDL-C, 8-bit weight-activation models remain within 0.05 dB of FP32 at 10% and 1% block error rate (BLER). At 4 bits, uniform INT4 loses 3.3-3.7 dB and falls below LS-LMMSE, whereas FP4 more than halves this loss (1.3-1.4 dB) and still outperforms it by about 0.5 dB, even after pruning. FP4's denser near-zero grid matches the trained weight distribution, and FP4 avoids the residual-path over-pruning seen with INT4. An analytic cost model projects 66x fewer bit-operations and 8.8x less weight storage for pruned 4-bit-weight inference.
1. INTRODUCTION
Neural receivers can improve BLER but their computational and memory demands challenge real-time deployment. This work tests weight and activation quantization and pruning in a standard-compliant MU-MIMO receiver, emphasizing whether FP4 preserves performance better than INT4.
1. INTRODUCTION
- Reported neural-receiver compute demand of ∼26 TFLOP/s motivates quantization to reduce storage, memory traffic, and bit-operation complexity.The introduction notes that real-time RAN compute relies on GPUs.Source: §1 · p. 1
- Prior integer weight-only QAT used uniform grids and full-precision activations, leaving floating-point microformats, post-ReLU activation quantization, pruning, and standard-compliant MU-MIMO reception untested.At the same nominal 4-bit budget, FP4 places reconstruction levels more densely near zero than uniform INT4.Source: §1 · p. 1
- Adding INT8 post-ReLU activation quantization to 8-bit weight-QAT models degrades required Eb/N0 by at most 0.08 dB.Weight–activation QAT and pruning are evaluated separately.Source: §1 · p. 1
- At 4 bits, FP4 (E2M1) outperforms INT4 before and after 50% unstructured pruning.The study also characterizes residual-kernel sparsity.Source: §1 · p. 1
- An analytic model quantifies weight storage and bit-operations for the compressed receiver.The paper evaluates weight–activation QAT and pruning separately.Source: §1 · p. 1
2. SYSTEM MODEL AND NEURAL RECEIVER
The receiver processes post-FFT data and pilot grids from a multi-user uplink, then produces per-bit LLRs for conventional decoding. Its GNN–CNN architecture is trained with a bitwise loss and an auxiliary channel-estimation objective.
2.1. Uplink Setup
- Single-layer UEs transmit on shared time–frequency resources to an Nrx-antenna base station; after OFDM demodulation, the receiver uses the received vector and pilot grid to output per-bit LLRs.DMRS tones provide an initial LS channel estimate, and conventional 5G NR decoding follows.Source: §2.1 · p. 1
2.2. Neural Receiver Architecture
- The GNN–CNN receiver uses eight unrolled blocks to exchange information across users and the time–frequency grid, then produces LLRs through a dense readout.Shared weights allow training on 4 PRBs and evaluation on a 132-PRB grid; pruning excludes depthwise and readout kernels.Source: §2.2 · pp. 1–2
2.3. Loss and Training Schedule
- Training minimizes bitwise BCE with an auxiliary channel-estimation loss, using random active-user counts, velocities, and rate-adjusted Eb/N0 values.The FP32 baseline is trained first, and quantized models warmstart from that checkpoint.Source: §2.3 · p. 2
3. QUANTIZATION-AWARE TRAINING AND PRUNING FOR MU-MIMO NEURAL RECEIVERS
The receiver uses quantization-aware training with integer and floating-point weight grids, INT8 post-ReLU activations, and a separate 50% magnitude-pruning procedure. Training starts from pretrained FP32 weights, with quantization-aware fine-tuning and fixed-mask recovery for sparse models.
3.1. Weight Fake Quantization
- Weight fake quantization uses per-output-channel abs-max scaling and a straight-through estimator to train a full-precision master copy against quantized weights.The quantization grid is recomputed from current master weights at each training step; biases use the corresponding weight format.Source: §3.1 · p. 2
3.2. Floating-Point Microformats: FP8 and FP4
- FP4 (E2M1) and INT4 both provide nominal 8× weight compression relative to FP32, but use different grid spacing.FP4 uses a fifteen-level grid with maximum finite magnitude 6; FP8 uses E4M3 with maximum finite magnitude 448.Source: §3.2 · p. 2
3.3. Activation Quantization
- Post-ReLU activations use per-tensor unsigned INT8 quantization, while network inputs and outputs remain in FP32.The activation scale is max(a)/255; all-zero tensors use a scale of 1, and training jointly fake-quantizes weights and activations.Source: §3.3 · p. 2
3.4. Unstructured Magnitude Pruning
- Global magnitude pruning targets pointwise, aggregation, and non-readout dense kernels, while depthwise and readout kernels remain dense.The removed fraction is set to 50%, and the keep mask is frozen during recovery fine-tuning and enforced after each update.Source: §3.4 · p. 2
3.5. Training Protocol
- Weight-only QAT starts from a pretrained FP32 checkpoint; weight-activation models and sparse models are fine-tuned separately from their corresponding weight-only checkpoints.Sparse models use fixed-mask recovery fine-tuning.Source: §3.5 · p. 2
4. EXPERIMENTAL RESULTS AND DISCUSSION
The experiments compare four weight formats under WbA8 QAT and separate 50% pruning, measuring BLER on a 3GPP-compliant two-user PUSCH setup and estimating inference cost. FP4 preserves more of the 4-bit receiver’s performance than INT4, including after pruning.
4.2. FP4 Halves the 4-bit Loss
- FP4 needs 1.32 dB and 1.40 dB more Eb/N0 than FP32 at 10% and 1% BLER, respectively, and beats LS–LMMSE by about 0.5 dB; INT4 needs 3.3–3.7 dB more and trails LS–LMMSE.After 50% pruning, FP4 changes by −0.07/−0.10 dB and remains ahead of LS–LMMSE, whereas pruning adds 0.95–1.29 dB to INT4, which trails LS–LMMSE by more than 2 dB.Source: Figure 1 · p. 3, §4.2 · p. 3, §4.3 · p. 4
4.1. Activation Quantization is Radio-Neutral on Link Performance at 8 Bits
- FP8 and INT8 WbA8 receivers remain within 0.05 dB of FP32 at 10% and 1% BLER, and adding INT8 activations changes required Eb/N0 by at most 0.08 dB.Both receivers outperform LS–LMMSE by approximately 1.83–1.91 dB and remain within ∼0.2 dB of LMMSE–K-Best.Source: §4.1 · p. 3, Figure 1 · p. 3, Table 2 · p. 3
4.2. FP4 Halves the 4-bit Loss
- FP4’s denser grid resolves the trained weight distribution’s near-zero region better: over [0, 0.25], it provides four magnitudes while INT4 provides two.Among 0.404 M normalized prune-eligible weights, 65% have |w| < 1/7 and 79% have |w| < 0.25; FP4 adds a level at 1/12 ≈0.08.Source: §4.2 · pp. 3–4, Figure 2 · p. 4
4.3. Post-QAT Magnitude Pruning
- After global 50% pruning, INT4 residual update-to-state kernels average 86.5% zeros, compared with 53.8% under FP4; the Input CNN state-write kernel reaches 78.5% and 36.1%, respectively.The paper describes INT4 sparsity as concentrated in residual paths, while FP4 sparsity remains comparable to INT8; FP4’s denser near-zero grid is consistent with its higher pruning robustness.Source: §4.3 · p. 4, Figure 2 · p. 4
- The modeled W4A8+P configuration reduces BOPs 66× and weight footprint 8.8× relative to FP32.These are analytic cost projections for the combined configuration; radio experiments evaluate WbA8 and pruning separately.Source: §4.3 · p. 4, Table 3 · p. 4
5. CONCLUSION AND FUTURE WORK
Eight-bit weight–activation QAT preserves link performance, while FP4 outperforms LS–LMMSE and supports substantial modeled compute and storage reductions. Future work targets structured sparsity, native MXFP4 scaling, and distillation-augmented QAT.
5. CONCLUSION AND FUTURE WORK
- 66× fewer bit-operations and 8.8× less weight storage are analytically projected for weight-activation quantization followed by pruning.The projection applies to the pruned inference model.Source: §5 · p. 4
- FP4 outperforms LS–LMMSE, whereas INT4 performs poorly; this supports microscaling formats for neural receivers.The conclusion aligns FP4 robustness with its denser near-zero grid and lower residual-path sparsity.Source: §5 · p. 4
- Future work will pursue 2:4 structured sparsity, native MXFP4 block scaling, and distillation-augmented QAT for ultra-low-bit receivers.Distillation-augmented QAT is proposed as a complementary route to improving link performance.Source: §5 · p. 4