Classical and Learned Decoders for URLLC: A Reliability-Latency Comparison
Anas Aljahdali *
Department of Electrical and Computer Engineering, King Abdulaziz University, Jeddah, Saudi Arabia.
Mohammad Awedh
Department of Electrical and Computer Engineering, King Abdulaziz University, Jeddah, Saudi Arabia.
*Author to whom correspondence should be addressed.
Abstract
Aims: This study evaluates which short-block channel decoders can jointly satisfy ultra-reliable low-latency communication (URLLC) requirements, defined here as BER/BLER ≤ 10−5 and mean decoding latency ≤ 1 ms. The comparison focuses on reliability-latency feasibility rather than error-rate ranking alone.
Methodology: A matched benchmark was conducted for Viterbi decoding of convolutional codes, plain Min-Sum belief propagation for LDPC codes, single-scalar normalised Min-Sum (NMS), learned neural normalised Min-Sum (NNMS), and CRC-aided successive-cancellation-list (CA-SCL) decoding for polar codes. All decoders were assessed at K = 64, 128, and 256 with rate 1/2 over BPSK/AWGN channels. Reliability was estimated from Monte-Carlo simulations with up to 106 blocks per SNR point, while latency was measured as mean per-block wall-clock decoding time in compiled MATLAB on an Intel Core i9 CPU.
Results: Polar CA-SCL required the lowest SNR across the tested lengths, but its sequential critical path increased latency from 0.47 ms to 1.47 ms and exceeded the 1 ms budget at K = 256. The BP-class decoders remained within approximately 0.16-0.18 ms. At K = 256, NMS recovered about half of the Min-Sum-to-Polar gap while remaining latency-feasible; NNMS matched NMS within 0.03 dB.
Conclusion: Under the stated conditions, decoder selection for short-block URLLC should be treated as a joint reliability-latency design decision. NMS provided the strongest feasible trade-off at K = 256, and learned per-edge weighting did not improve over a tuned scalar normalisation for the LDPC codes considered.
Keywords: URLLC, short-blocklength codes, finite-blocklength regime, channel decoding, reliability-latency trade-off, belief propagation, normalised Min-Sum, neural decoder, polar CA-SCL, LDPC codes, decoding latency