Requirements | Compilation | Integration | Documentation | Roadmap |

NeoN

NeoN is an open-source, high-performance C++ CFD library for modern heterogeneous computing systems. It currently provides data structures, parallel algorithms, and numerical infrastructure for the finite volume method.

By combining finite-volume abstractions with modern C++ and performance-portable backends, NeoN enables developers to build maintainable, scalable, and high-performance fluid-flow solvers without writing architecture-specific code for each hardware platform.

Its modular architecture enables performance-portable execution and solver backends, currently using Kokkos and Ginkgo while remaining flexible to adopt alternative technologies as the framework evolves.

Key Features

• Execution on
  • serial CPU
  • multithreaded CPU
  • MPI-based distributed systems
  • GPUs from NVIDIA, AMD, and Intel
• Portability across Linux, macOS, and Windows
• Performance-portable parallel and memory abstractions
• Unified GPU execution model for NVIDIA, AMD, and Intel GPUs

---

Important

The NeoN project needs you! If you're interested in contributing to NeoN please open a PR! If you have any questions on where to start please contact us here or on gitter.

Requirements

NeoN has the following requirements

• cmake > 3.22
• gcc >= 13 or clang >= 19
• Kokkos 5.0.2

For GPU support

• NVIDIA: CUDA 12+
• AMD: ROCm 6.4.1
• Intel: oneAPI Base Toolkit 2024.2

For development it is required to use pre-commit.

C++ dependencies

C++ dependencies like Kokkos are handled via CPM and are cloned at the configuration step. However, the cmake build process will prefer system wide installed C++ dependencies like Kokkos, cxxopts, etc. If you prefer to clone, configure and build dependencies your self consider setting -DCPM_USE_LOCAL_PACKAGES = OFF, see CPM for more details.

Compilation

NeoN uses cmake to build, thus the standard cmake procedure should work. From a build directory you can execute

cmake <DesiredBuildFlags> ..
cmake --build .
cmake --install .

Additionally, we provide several Cmake presets to set commmonly required flags if you compile NeoN in combination with Kokkos.

cmake --list-presets # To list existing presets
cmake --preset production # To configure for production use
cmake --build --preset production # To compile for production use

Executing Tests

We provide a set of unit tests which can be executed via ctest or

cmake --build . --target test

Installing the Python bindings

NeoN ships Python bindings as the neon_pde distribution (imported as neon). The package requires Python 3.9–3.13 and NumPy.

CPU (from PyPI). Pre-built CPU wheels are published to PyPI for Linux (x86-64/ARM64), Windows (AMD64) and macOS (Apple Silicon/Intel):

pip install neon_pde

Conda / pixi (from prefix.dev). The same releases are published as conda packages named neon-pde to a prefix.dev channel, which also installs the NeoN C++ runtime, headers and CMake package files into the environment:

pixi add neon-pde -c https://prefix.dev/exasim-project -c conda-forge

# CUDA 12.8, CPython 3.12, NVIDIA Ampere (sm_80), linux-64
pixi add "neon-pde=*=cuda_py312*" -c https://prefix.dev/exasim-project -c conda-forge

Conda packages cover CPython 3.10–3.13 on linux-64, linux-aarch64, osx-64 and osx-arm64; Windows and Python 3.9 are wheel-only.

CUDA (from GitHub Releases). GPU wheels are not published to PyPI because they are large and depend on the NVIDIA driver. They are attached to the corresponding GitHub Release and carry a local version suffix such as +cuda128. A release carries a single GPU wheel, built for CPython 3.12 on Linux x86-64; install it directly by URL:

# CUDA 12.8, CPython 3.12, Linux x86-64, NVIDIA Ampere (sm_80)
pip install https://github.com/exasim-project/NeoN/releases/download/v0.3.0rc2/neon_pde-0.3.0rc2+cuda128-cp312-cp312-manylinux_2_35_x86_64.whl

The CUDA wheel needs a host NVIDIA driver providing libcuda.so.1 (it is intentionally not bundled). It does not require a local CUDA toolkit at runtime, but it does require glibc 2.35 or newer (Ubuntu 22.04+, Debian 12+).

From source. Building the bindings uses scikit-build-core and compiles the C++ library, so a C++20 compiler and CMake ≥ 3.22 are required:

# CPU build. Both GPU backends must be disabled explicitly: with neither set,
# AutoEnableDevice.cmake probes for nvcc/hipcc and enables that backend.
pip install . --config-settings=cmake.define.Kokkos_ENABLE_CUDA=OFF \
              --config-settings=cmake.define.Kokkos_ENABLE_HIP=OFF

# CUDA build (matching the released wheels)
pip install . --config-settings=cmake.define.Kokkos_ENABLE_CUDA=ON \
              --config-settings=cmake.define.Kokkos_ENABLE_COMPILE_AS_CMAKE_LANGUAGE=ON \
              --config-settings=cmake.define.Kokkos_ARCH_AMPERE80=ON \
              --config-settings=cmake.define.CMAKE_CUDA_ARCHITECTURES=80 \
              --config-settings=cmake.define.CMAKE_CUDA_STANDARD=20

Verify an installation with:

import neon
print(neon.__version__, neon.__has_serial__, neon.__has_cpu__, neon.__has_gpu__)

See the documentation for the full install-and-release workflow.

Python Wheels

The Python distribution name is neon_pde, which produces wheel filenames starting with neon_pde. The import package remains neon.

The package version in pyproject.toml is the source of truth. CMake reads that version during configuration, and the generated neon.__version__ uses the same value.

GitHub Actions uses cibuildwheel to build wheels for release tags and explicitly requested manual builds. Stable releases use tags named like v0.1.0. Manual non-tag builds use development versions like 0.1.1.dev202605270217123. The wheel workflow does not currently run for ordinary branch pushes, pull requests, or on a nightly schedule.

CPU wheels are built for:

• Linux x86-64
• Linux ARM64
• Windows AMD64
• macOS Apple Silicon
• macOS Intel

The CPU matrix covers CPython 3.9 through 3.13. CPU wheels use the plain package version and are published to PyPI only for stable v*.*.* tags.

CUDA wheels are currently limited to CUDA 12.8 on Linux x86-64 with CPython 3.12. They use a local version suffix such as 0.1.0+cuda128, are uploaded as workflow artifacts, and are attached to the GitHub Release for stable tags.

Conda packages

Conda packages are built from recipe/recipe.yaml with rattler-build by .github/workflows/conda_packages.yaml, on the same triggers and with the same version scheme as the wheels, and are published to prefix.dev. Build one locally with:

ci/install_rattler_build.sh                       # optional, pinned release
ci/build_conda_packages.sh --python 3.12 --gpu cpu

The GPU flavour is part of the build string (cpu_py312_*, cuda_py312_*, rocm_py312_*), so all flavours coexist in one channel. The ROCm flavour is experimental and manual-dispatch only: conda-forge has no rocBLAS/rocSPARSE/ rocThrust/rocPRIM, so it ships Kokkos HIP without the Ginkgo solver backend. GitHub-hosted runners do not provide a GPU, so the CUDA wheel build does not run runtime tests against the resulting wheel.

Integration with other CFD Frameworks

Currently, NeoN is not a standalone CFD framework. It is designed to be used with other CFD Frameworks. Examples how to integrate NeoN into CFD frameworks and how to write applications is demonstrated in the NeoFOAM repository.

Documentation

An online documentation can be found here, be cautious since this repository is currently evolving the documentation might not always reflect the latest stage.

For building the documentation further dependencies like doxygen and sphinx are requirement. The list of requirements can be found here