Introduction
Standard 5G New Radio (NR) Low-Density Parity-Check (LDPC) codes are the backbone of modern wireless error correction. While these codes are theoretically designed for "infinite" decoding iterations, real-world hardware—constrained by power, area, and latency—typically performs only 5 to 20 iterations. This paper identifies a significant performance gap: puncturing patterns optimized for theoretical limits are often sub-optimal for practical, low-iteration decoders. To bridge this gap, the authors introduce SPAT (Swapping of Punctured And Transmitted blocks), an algorithm that optimizes bit selection to boost performance in hardware-constrained scenarios.
The Puncturing Performance Gap
In 5G NR, "puncturing" is used to achieve various data rates by calculating certain bits but omitting them from transmission. The standard patterns were designed using asymptotic tools like density evolution, which assume an ideal decoder. However, practical mobile devices and base stations use message-passing (MP) decoders that run for a limited number of iterations.
As a result, the standard patterns do not account for the convergence speed of the decoder. The authors argue that by carefully selecting which blocks to puncture, one can significantly improve the Signal-to-Noise Ratio (SNR) required to reach target error rates without altering the underlying code structure.
The SPAT Algorithm
The proposed solution is a greedy optimization heuristic called SPAT (Swapping of Punctured And Transmitted blocks). Rather than redesigning the LDPC code—which would break compatibility with the 5G standard—SPAT rearranges the existing blocks.
Starting with the standard 5G pattern, SPAT iteratively attempts to swap a punctured block with a transmitted block. Using Monte-Carlo simulations, it evaluates each swap based on a hardware-aware objective: the lowest SNR needed to achieve a target Block Error Rate (BLER) of $10^{-3}$ within a fixed, low number of iterations (e.g., 10 iterations). This ensures the resulting pattern is tailored to the specific behavior of the actual decoder implementation.
Key Insights and the "Head Start" Effect
To understand why SPAT works, the authors utilized empirical average Binary Cross-Entropy (BCE) to track decoder convergence. They discovered a "head start" effect: optimized patterns prioritize transmitting bits that reduce initial uncertainty for the decoder. By providing more informative bits early on, the decoder converges to the correct solution much faster than it would with the standard 5G pattern.
This reveals a fundamental trade-off between "early" and "late" performance. While SPAT-optimized patterns excel at 10 iterations, they may actually be outperformed by standard patterns if the decoder were allowed to run for 1,000+ iterations. However, in the context of real-world hardware, the early convergence provided by SPAT is far more valuable.
Results and Impact
Evaluated using the NVIDIA Sionna library across various code lengths and rates, SPAT-optimized patterns consistently outperformed the 5G NR standard:
- Short codes ($n=128$): Achieved gains of 0.28 dB to 0.55 dB.
- Medium codes ($n=512$): Achieved gains of approximately 0.20 dB.
- Long codes ($n=4096$): Achieved gains of approximately 0.24 dB.
Crucially, because these gains are achieved by simply choosing a different set of bits to transmit at the design stage, they come with zero additional computational cost or complexity overhead during real-time transmission and decoding.
Conclusion
The SPAT algorithm demonstrates that hardware constraints should be a primary consideration in code design. By optimizing puncturing patterns for low-iteration scenarios, significant SNR gains can be realized for free. Future work includes extending this methodology to Hybrid Automatic Repeat Request (HARQ) systems to optimize retransmission patterns, potentially leading to even greater efficiency in 5G and future 6G networks.
Summary updated