RapidCFD

Developed by SimFlow CFD
https://sim-flow.com

CFD toolbox running on CUDA

OpenFOAM solvers ported to Nvidia CUDA. All the calculations are done on the GPU giving a huge speed-up. Still in development stage, waiting for your contribution!

Features:

• most incompressible and compressible solvers on static mesh are available
• all the calculations are done on the GPU
• no overhead for GPU-CPU memory copy
• can run in parallel on multiple GPUs

Build status (this fork)

Successfully built and smoke-tested on:

• Host: Ubuntu 24.04 LTS
• GPU: NVIDIA RTX PRO 6000 Blackwell Max-Q (compute capability 12.0 / sm_120)
• CUDA Toolkit: 12.8
• MPI: system OpenMPI (SYSTEMOPENMPI)
• Platform tag:linux64NvccDPOptSM120

Verified: icoFoam lid-driven cavity (20×20×1) runs to completion on GPU with orthogonal schemes.

Compilation (Ubuntu 24.04 + modern CUDA)

1. Install CUDA toolkit (nvcc) and a compatible NVIDIA driver. Ensure nvcc is on PATH:

   export PATH=/usr/local/cuda-12.8/bin:$PATHexport LD_LIBRARY_PATH=/usr/local/cuda-12.8/lib64:$LD_LIBRARY_PATH

2. Install build deps: g++, make, flex, bison, openmpi (libopenmpi-dev, openmpi-bin).

3. Clone so the install root contains RapidCFD-dev (and optionally ThirdParty-dev):

   export FOAM_INST_DIR=/path/to/parent # e.g. $HOME/dev# layout: $FOAM_INST_DIR/RapidCFD-dev

4. Set GPU architecture in etc/bashrc / etc/prefs.sh (WM_GPU_ARCH):

   GPU family	WM_GPU_ARCH
   Ada (RTX 40)	sm_89
   Hopper	sm_90
   Blackwell (RTX PRO 6000 / RTX 50)	sm_120

5. Source and build:

   export FOAM_INST_DIR=/path/to/parent
   source$FOAM_INST_DIR/RapidCFD-dev/etc/bashrc
   export WM_NCOMPPROCS=$(nproc)# or a lower number
   mkdir -p "$FOAM_EXT_LIBBIN"
   ./Allwmake 2>&1| tee build.log

6. Run solvers with a GPU device:

   icoFoam -device 0 -case /path/to/case

Notes:

• This tree only builds solvers (no blockMesh / mesh utilities). Generate meshes with system OpenFOAM or another tool.
• Prefer snGradSchemes { default orthogonal; } and matching laplacian schemes for now; corrected snGrad can hit a known CUDA invalid device function path on Tensor gradients under CUDA 12.8 / sm_120.
• Multi-GPU still benefits from ThirdParty CUDA-aware MPI (optional).

Original notes (Ubuntu 16.04 / CUDA 8)

• ensure CUDA 7.5 is not installed from Ubuntu repositories
• ensure you are using an nVidia driver compatible with CUDA 8
• download CUDA 8.0 from NVIDIA's archive
• to compile in parallel, export WM_NCOMPPROCS=10
• ThirdParty-dev is needed for multiple GPUs with the bundled OpenMPI build