Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Memory-Assisted Quantized LDPC Decoding

Philipp Mohr, Gerhard Bauch

TL;DR

This work tackles information loss in coarse-quantized LDPC decoding by preserving and reusing previous CN messages as side information within a memory-assisted reconstruction framework. A side-information aware information bottleneck algorithm designs CN quantizers that maximize preserved information, while a simple merge rule minimizes memory and lookup complexity. The approach yields tangible gains, including up to $0.23$ dB improvement for 2-bit decoding and up to $32\%$ area efficiency at high rates, demonstrated on 5G-LDPC codes with flood- and layered-scheduling, indicating practical benefits for high-throughput decoders. The results support a hybrid decoding strategy that leverages memory for high-rate operation and non-memory decoding for low-rate scenarios to optimize performance and hardware efficiency.

Abstract

We enhance coarsely quantized LDPC decoding by reusing computed check node messages from previous iterations. Typically, variable and check nodes update and replace old messages every iteration. We show that, under coarse quantization, discarding old messages entails a significant loss of mutual information. The loss is avoided with additional memory, improving performance by up to 0.23 dB. We optimize quantization with a modified information bottleneck algorithm that considers the statistics of old messages. A simple merge operation reduces memory requirements. Depending on channel conditions and code rate, memory assistance enables up to 32 % better area efficiency for 2-bit decoding.

Memory-Assisted Quantized LDPC Decoding

TL;DR

This work tackles information loss in coarse-quantized LDPC decoding by preserving and reusing previous CN messages as side information within a memory-assisted reconstruction framework. A side-information aware information bottleneck algorithm designs CN quantizers that maximize preserved information, while a simple merge rule minimizes memory and lookup complexity. The approach yields tangible gains, including up to dB improvement for 2-bit decoding and up to area efficiency at high rates, demonstrated on 5G-LDPC codes with flood- and layered-scheduling, indicating practical benefits for high-throughput decoders. The results support a hybrid decoding strategy that leverages memory for high-rate operation and non-memory decoding for low-rate scenarios to optimize performance and hardware efficiency.

Abstract

We enhance coarsely quantized LDPC decoding by reusing computed check node messages from previous iterations. Typically, variable and check nodes update and replace old messages every iteration. We show that, under coarse quantization, discarding old messages entails a significant loss of mutual information. The loss is avoided with additional memory, improving performance by up to 0.23 dB. We optimize quantization with a modified information bottleneck algorithm that considers the statistics of old messages. A simple merge operation reduces memory requirements. Depending on channel conditions and code rate, memory assistance enables up to 32 % better area efficiency for 2-bit decoding.
Paper Structure (12 sections, 10 equations, 6 figures, 1 table, 1 algorithm)

This paper contains 12 sections, 10 equations, 6 figures, 1 table, 1 algorithm.

Table of Contents

  1. Introduction
  2. Preliminaries on LDPC Decoding with Mutual Information Maximizing Quantization
  3. Design with Memory-Assisted Reconstruction
  4. Information Bottleneck Algorithm with Side-Information Awareness
  5. Merging Costs with Side Information
  6. Performance and Complexity with 5G Codes
  7. Evolution of Mutual Information
  8. Boundary Placement for Memory-Assisted Reconstruction
  9. Complexity Reduction of Memory-Assisted Decoding
  10. Frame Error Rate Performance using a Flooding Schedule
  11. Complexity vs. Performance under Layered Decoding
  12. Conclusions

Figures (6)

  • Figure 1: Quantizer design with memory-assisted reconstruction.
  • Figure 2: Evolution of mutual information with code rate $1/3$. All results obtained with same design $E_b/N_0{=}1.0$ dB.
  • Figure 3: Evolution of quantizer boundary placement each with code rate $1/3$ and individual design $E_b/N_0=.55$, $.73$ and $1.0$ dB.
  • Figure 4: The compression operations $\Upsilon$ and $\Psi$ reduce the complexity for the VN's reconstruction $\phi$ and memory to store side information $s^\Psi$.
  • Figure 5: FER performance of memory-assisted (solid lines) and non-assisted (dotted lines) setups for 5G-LDPC decoding (flooding schedule, $\iota_{max}{=}30$).
  • ...and 1 more figures