The main change in this release is infrastructure: AFF3CT's continuous-integration and delivery pipeline has moved from GitLab CI to GitHub Actions, alongside the code and pull requests. For a simulator whose results need to be correct and reproducible, this changes how every contribution is verified.

Changes are verified before they merge

• Pull requests must pass CI to be merged. Static analysis (cppcheck, clang static analyzer, clang-format) and a cross-platform build matrix — Linux (GCC / Clang / ICPX, plus GSL and MPI variants), Windows (MSVC), and macOS (Apple Silicon + Intel) — run on every PR and are required status checks on develop.
• Regression and coverage testing on real hardware. A self-hosted runner replays the reference simulations across each code family (BCH, Convolutional, LDPC, Polar, Turbo, …) and compares the BER/FER curves against stored references, so a change that degrades decoding performance is caught. A fast profile runs on each push; a thorough profile runs weekly.
• Approval gate for external contributions. Because the regression suite runs on our own hardware, an external pull request does not execute on the self-hosted runner until a maintainer approves it. Collaborators with write access run immediately; others wait for review.

Tag-driven releases

• Documentation, the Launchpad PPA package, and binaries (Linux / Windows / macOS arm64 & x86-64) are built and validated on every branch and PR as a dry run, so packaging problems surface early.
• A real publish happens only when a v* tag is pushed: the PPA upload, the GitHub Release, and the website download page update. This release was produced by that pipeline.

Notes for research use

Building on every supported compiler and OS, re-checking reference curves on real hardware, and accepting external contributions without manual infrastructure steps all support the reproducibility of results produced with AFF3CT and make the toolbox easier to extend.

Also in this release

• 5G NR LDPC (QC) matrices now ship with the toolbox, enabling 5G LDPC simulations out of the box.
• Windows / MSVC 2022 support, with the associated portability fixes.
• Prebuilt binaries for Linux, Windows, and macOS on both arm64 and x86-64 (SSE4.2), available from the download page.

Full changelog: v4.5.0...v4.6.0

What’s New

Changes listed bellow:

• Add support for 5G NR UL-SCH FEC (LDPC QC)
• Add optimized 5G LDPC QC encoder implementation
• Add native 5G puncturing and rate-matching compatibility
• Fix Sum-Product Algorithm (SPA) initialization for zero-valued LLRs to avoid division by zero

Performance Benchmarks: AFF3CT vs. Sionna

Nvidia’s Sionna library has recently gained significant traction within the digital communications research community. Starting with v2.0.0, it transitioned from a TensorFlow to a PyTorch backend, significantly improving framework interoperability and developer adoption. Sionna claims that "NVIDIA GPU acceleration provides orders-of-magnitude faster simulations". To verify this, we conducted a rigorous, apples-to-apples comparison focused exclusively on 5G NR UL-SCH LDPC decoding chain. This benchmark evaluates AFF3CT (CPU-optimized) against Sionna (GPU-optimized) across various frame sizes, decoding algorithms, and quantization formats.

Figure: 5G NR UL-SCH FEC simulation chain used for benchmarks.

Hardware Configuration

• CPU: AMD Ryzen 9 7945HX (AMD Zen 4 architecture, 16 cores, 75 W TDP)
• GPU: Gainward GeForce RTX 4090 Phantom GS 24 GB (Nvidia Ada Lovelace architecture, 128 SMs, 450 W TDP)
• Platform: Dalek cluster (az4-n4090 partition)

This benchmark contrasts a high-performance mobile workstation CPU (approximately 800€ for the motherboard, processor, and 16 GB dual-channel RAM) against a flagship desktop GPU (1,779–3,000€). While the cost and target markets differ significantly, this asymmetric comparison provides valuable insight into the practical trade-offs between CPU-optimized and GPU-accelerated FEC decoding across different deployment scenarios.

Software Environment

The benchmarks were executed on Ubuntu 24.04.4 LTS HWE (Linux kernel 6.17) with Nvidia driver 595.71.05 and CUDA 13.2.1. All C++ components were compiled using GCC 13.3.0. The evaluation utilizes AFF3CT v4.5.0 and Sionna v2.0.1 (PyTorch v2.12.1). AFF3CT was built with the following configuration to maximize CPU throughput and vectorization through well-spread AVX2 SIMD extension:

cmake -S . -B build -DCMAKE_BUILD_TYPE="RelWithDebInfo" -DCMAKE_CXX_FLAGS="-Wall -funroll-loops -faligned-new -mavx2 -mfma"

Decoding Configuration & Measured Metrics

To ensure reproducible and stable measurements, all LDPC decoders were executed without early stopping (or syndrome checking). The following parameters and metrics were recorded in the following tables for each experiment:

• ID: Unique experiment identifier
• K: Number of information bits per frame
• Decoder: Scheduling and update rule (BP-F: Flooding schedule; BP-HL: Horizontal Layered schedule; SPA: Sum-Product Algorithm / boxplus; OMS: Offset Min-Sum approximation)
• Format: Data precision and format (f32: 32-bit floating-point; i16: 16-bit fixed-point Q10.4)
• Ite: Number of decoding iterations in the BP
• Threads: CPU thread count (AFF3CT only)
• Batch Size: Number of decoded frames in parallel. For AFF3CT, formatted as (A×B)=C, where A is the inter-frame SIMD width and B the number of CPU threads. For Sionna, this denotes frames processed per GPU kernel launch.
• Throughput: Average information throughput (accounting for K bits)
• Latency: Average decoding time per frame
• Power: Average instantaneous power draw
• Energy: Average energy consumption per frame

Power Measurement Methodology

Power was logged as a stable average during simulation execution. GPU power was captured via nvidia-smi, while CPU power was measured using RAPL (pkg-0 domain) with powerstat. RAM power was excluded as RAPL does not support the memory domain on this platform; however, estimated DDR5 SO-DIMM draw (2–6 W in dual channel) is negligible and does not affect the observed trends.

Metric Measurement & Derivation

To ensure methodological rigor, we distinguish between directly measured and derived quantities:

• Directly Measured: Average information throughput and average instantaneous power consumption are captured directly during simulation execution
• Derived Quantities: Latency and energy are computed to accurately reflect per-frame performance across different parallelization schemes:
  • Latency for AFF3CT = (K × Threads) / (Throughput / inter-frame SIMD width)
  • Latency for Sionna = K / (Throughput / Batch Size)
  • Energy = (Power / (Throughput_Mbps / K)) / 1000 (<=> (Power / (1 / T_μs)) / 1000 <=> (Power × T_μs) / 1000 = Energy_mJ)

Running Sionna

To run Sionna in a way that closely matches the AFF3CT simulator behavior, we developed a minimalist Python script called "MiniFEC". See the minifec.py file included in the release assets. The script supports results reproducibility and enables potential corrections.

Benchmarks Philosophy

The following benchmarks evaluate decoding performance (FER) for a rate-1/2 LDPC code compliant with the 5G NR UL-SCH FEC specification. Three frame sizes are considered, spanning the full range defined in the standard—from short payloads to large transport blocks—allowing us to assess how CPU and GPU architectures scale in terms of throughput, latency, and energy efficiency across realistic deployment scenarios.

Each subsection presents FER results, followed by a summary of key performance metrics and a pie chart showing the distribution of execution time across processing tasks. While this breakdown is easily obtained in AFF3CT using the --sim-stats option, it is more challenging in Sionna, where operations are fused into PyTorch mega-kernels, making per-block analysis less straightforward.

Short Code Length - N = 40, K = 20 (BG2, N_LDPC = 40, K_LDPC = 208)

Open the detailed BER/FER plots on the AFF3CT online comparator.

Changes listed bellow:

• Add polarization-adjusted convolutional (PAC) codes support with state-of-the-art SCL decoder (thx @Murad1997)
• Add a generic convolutional encoder (thx @zsoerenm)
• Add support of generic convolutional simulation with Viterbi and Viterbi list decoders (thx @zsoerenm and @alvalade)
• Improve the speed of the RS encoder (up to 17%) (thx @sudo-LuSer)
• Fix transform_H_to_G_identity for rank-deficient parity-check matrices (thx @zsoerenm)
• Add a new template to customize LDPC syndrome checking (useful to implement coset decoding for instance)
• Update MIPP, StreamPU and references
• New fancy theme for the documentation (Furo)

New major version.

Changes listed bellow:

• Add Viterbi+TBCC and Viterbi list decoders (for RSC codes)
• Add hard input majority vote repetition decoder
• Update from AFF3CT-core to StreamPU
• Improve CI
• Tons of small code improvements and fixes

Changes listed below:

• add operator= method to the module::Socket class to simplify the binding
• add an automated create_reset_task method in the module::Module class
• add get_n_input_sockets and get_n_output_sockets methods in module::Task
• add tools::Interface_is_done to tools::Sequence: enable auto-stop in tools::Pipeline
• add exec_step method to tools::Sequence: enable task by task execution
• improve tools::Sequence: new task is added to graph if all input sockets are fed
• implement get_info_bits_pos method in module::Extractor_RSC class
• improve module::Switcher: enable different type and number of elements in commute and select tasks
• plug CWD socket to the RS decoder (to know if the decoder found a codeword)
• replace reference members by copy in modules (to simplify the binding with py_aff3ct)
• add documentation of some tasks
• fix doc requirements
• fix the LDPC BP flooding/layered decoder (inter-frame SIMD)
• fix the puncture method prototype of the module::Puncturer class
• fix AZCW simulations
• unbind adaptors when a tools::Pipeline is destroyed (avoid bugs)
• disable backtrace on Android NDK

Changes listed below:

• add tools::Sequence and tools::Pipeline

• add 0-copy module::Adaptor for efficient tools::Pipeline implementation

• add module::Binaryop, module::Unaryop, module::Reducer

• add module::Switcher, module::Controller and module::Iterator to support
  loops and conditional structures (if, switch, ...)

• add the hwloc dependency for thread pinning in tools::Sequence and
  tools::Pipeline

• implement clonable modules (required for automated duplication in
  tools::Sequence)

• add a status module::Socket for each task automatically

• add a priority during the binding of a module::Socket

• add an unbind method to module::Socket

• add BCH and LDPC decoding status

• add FRA, BE, FE, BER and FER as output sockets for monitor BFER

• add Multi-kernel Polar encoders and decoders

• add module::Extractor to extract information bits in a codeword

• add module::Initializer, module::Incrementer, module::Sleeper and
  module::Finalizer

• add module::Probe for very precise feedback

• add module::Sink (write end of a communication system into a file)

• add module::Subsequence to regroup multiple tasks in a single one

• add tools::Interface (clone, reset, set_n_frames, is_done, set_seed, ...)

• add the C++ API doc in Sphinx (Doxygen + Breathe + Sphinx)

• add an inter-SIMD implementation of the BCH encoder

• add new code rates to DVB-RCS2

• add a set_n_frames method in the module::Module, tools::Sequence and
  tools::Pipeline

• add native support of inter-frames (in the module::Task)

• add a new socket CWD (codeword) to get the decoding status per frame

• add the operator[] method for string in module::Module

• add scripts (based on Clang AST) to simplify the automation of AFF3CT wrappers
  (Python, MATLAB)

• add auto-stop feature in tools::Sequence and tools::Pipeline (based on the
  is_done interface)

• add tools::Digraph (directed graph)

• add a new operator= to module::Socket: sockets can be bound in both
  directions (SIN = SOUT or SOUT = SIN)

• replace tools::Noise by a channel parameter (CP) socket in module::Channel
  and module::Modem

• command Line Interface (CLI) has been externalized

• split source code (src/) in public include/ and private src/ folders

• simplify tools::Interleaver: no need to call the init method anymore

• dockerize the CI/CD

• fix a bug in the LDPC Gallager A decoder for irregular matrices

• fix the bug on the majority vote of the LDPC Gallager B decoder

• fix bug en the DVB-S2 constants (thx Ben Mathews)

• fix Galois field segmentation fault

• speedup the LDPC Gallager E (~x2)

• remove SystemC/TLM interfaces

• remove EXIT chart simulations

• remove legacy simulation (without tools::Sequence)

• remove SIN_SOUT socket type

Changes listed below:

• add tools::Sequence and tools::Pipeline

• add 0-copy module::Adaptor for efficient tools::Pipeline implementation

• add module::Binaryop, module::Unaryop, module::Reducer

• add module::Switcher, module::Controller and module::Iterator to support
  loops and conditional structures (if, switch, ...)

• add the hwloc dependency for thread pinning in tools::Sequence and
  tools::Pipeline

• implement clonable modules (required for automated duplication in
  tools::Sequence)

• add a status module::Socket for each task automatically

• add a priority during the binding of a module::Socket

• add an unbind method to module::Socket

• add BCH and LDPC decoding status

• add FRA, BE, FE, BER and FER as output sockets for monitor BFER

• add Multi-kernel Polar encoders and decoders

• add module::Extractor to extract information bits in a codeword

• add module::Initializer, module::Incrementer, module::Sleeper and
  module::Finalizer

• add module::Probe for very precise feedback

• add module::Sink (write end of a communication system into a file)

• add module::Subsequence to regroup multiple tasks in a single one

• add tools::Interface (clone, reset, set_n_frames, is_done, set_seed, ...)

• add the C++ API doc in Sphinx (Doxygen + Breathe + Sphinx)

• add an inter-SIMD implementation of the BCH encoder

• add new code rates to DVB-RCS2

• add a set_n_frames method in the module::Module, tools::Sequence and
  tools::Pipeline

• add native support of inter-frames (in the module::Task)

• add a new socket CWD (codeword) to get the decoding status per frame

• add the operator[] method for string in module::Module

• add scripts (based on Clang AST) to simplify the automation of AFF3CT wrappers
  (Python, MATLAB)

• add auto-stop feature in tools::Sequence and tools::Pipeline (based on the
  is_done interface)

• add tools::Digraph (directed graph)

• add a new operator= to module::Socket: sockets can be bound in both
  directions (SIN = SOUT or SOUT = SIN)

• replace tools::Noise by a channel parameter (CP) socket in module::Channel
  and module::Modem

• command Line Interface (CLI) has been externalized

• split source code (src/) in public include/ and private src/ folders

• simplify tools::Interleaver: no need to call the init method anymore

• dockerize the CI/CD

• fix a bug in the LDPC Gallager A decoder for irregular matrices

• fix the bug on the majority vote of the LDPC Gallager B decoder

• fix bug en the DVB-S2 constants (thx Ben Mathews)

• fix Galois field segmentation fault

• speedup the LDPC Gallager E (~x2)

• remove SystemC/TLM interfaces

• remove EXIT chart simulations

• remove legacy simulation (without tools::Sequence)

• remove SIN_SOUT socket type

Changes listed below:

• Fix link issues due to missing 'inline' keywords.
• Add a custom name to the abstract 'Module' class.
• Fix SystemC compilation error.
• Remove the outdated completion script.

Changes listed below:

• Rename PPA packages:
  • 'aff3ct' becomes 'aff3ct-bin'.
  • 'aff3ct-dev' has been split in 'libaff3ct' and 'libaff3ct-dev'.
  • 'aff3ct-doc' stays 'aff3ct-doc'.
• Remove the automatic swap of M, N and K in the alist G and H matrices.
• Add an exception when the path to the polar best channels is wrong.
• Move versioning functions from 'aff3ct' to 'aff3ct::tools' namespace.
• Remove all the code from the '.h' and '.hpp' header files.
• Replace includes relative paths by absolute paths.
• Clean includes, remove useless ones and add others:
  • speedup re-compilation time in many cases.
  • allow to easily include specific headers from external project.

Changes listed below:

• Add AFF3CT PPA repositories for Debian/Ubuntu OSes:
  • ppa:aff3ct/aff3ct-stable
  • ppa:aff3ct/aff3ct-dev
• Doc: explain how to use AFF3CT as a library.