Lattice Boltzmann CFD That Runs on Basically Any GPU or CPU
You've got a fluid dynamics problem to solve, and you'd rather not spend your afternoon fighting CUDA toolkits or realizing your laptop's integrated graphics don't support the one framework the tutorial assumes. FluidX3D takes a different approach: it's lattice Boltzmann CFD software that runs on all GPUs and CPUs via OpenCL, and it bills itself as the fastest and most memory efficient in its class.
What It Does
FluidX3D is computational fluid dynamics software built on the lattice Boltzmann method. It uses OpenCL (through the project's own OpenCL-Wrapper) to run across GPUs and CPUs from any vendor, which means you're not locked into a single hardware ecosystem. The project positions itself around two claims: speed and memory efficiency.
The feature set has grown meaningfully since the public v1.0 release in August 2022. Later versions added solid voxelization on the GPU, force and torque computation on objects, multi-GPU support on a single node, and particles using the immersed-boundary method (passive or 2-way-coupled). There's also an interactive graphics mode, including X11 support on Linux as of v1.4. The README includes several linked YouTube videos showing it in action, so you can see results before committing to a build.
Why It's Cool
-
Cross-vendor hardware support is the headline. OpenCL means you can run this on NVIDIA, AMD, or Intel hardware, discrete or integrated. If you've ever been handed a machine with an Intel iGPU and told to "just use CUDA," you'll appreciate why this matters.
-
The multi-GPU upgrade landed in v2.0 with cross-vendor support on a single node (PC, laptop, or server). That's a practical scaling path that doesn't require a cluster.
-
Voxelization got dramatically faster. Version 1.1 added GPU voxelization with what the README honestly calls a slow algorithm. By v2.1, they'd replaced it with a new algorithm that took it "from minutes to milliseconds." That kind of iteration is worth noting.
-
Moving geometry is supported. Version 2.2 added the option to voxelize moving or rotating geometry on the GPU, with automatic velocity initialization for each grid point based on center of rotation, linear and rotational velocity. Cells converted from solid to fluid during re-voxelization get their distribution functions properly initialized, which is the kind of detail that separates a working simulation from a broken one.
-
Force and torque computation arrived in v1.2, along with a Stokes drag validation setup. Having a validation case shipped with the tool is a good sign for anyone who cares whether the numbers are right.
-
Platform quirks are handled. The update history mentions workarounds for an Intel iGPU driver bug with triangle rendering and an AMD legacy driver bug with binary number literals. These aren't glamorous, but they're the difference between software that runs on your machine and software that crashes on it.
One thing to flag honestly: the README states the software is free for non-commercial use. If you're evaluating this for commercial work, check the licensing terms before you build anything on top of it. Also note that particles are only supported with single-GPU setups.
How to Try It
- Head to the repository and check the releases page for the latest version.
- Review the update history to see what's changed recently—the project moves quickly, and features like multi-GPU support and velocity voxelization are relatively new.
- Watch one or two of the linked YouTube videos to get a sense of what the output looks like and whether the visualization style fits your needs.
- Since it's built on OpenCL, make sure you have OpenCL drivers installed for your GPU or CPU. The project depends on its OpenCL-Wrapper, so that's part of the stack you'll be working with.
The README is truncated in the version provided here, so for full setup instructions and build details, go straight to the repo.
Final Thoughts
FluidX3D is aimed at people who want serious lattice Boltzmann simulation without hardware lock-in—students, researchers, and hobbyists working with whatever hardware they happen to have. The rapid release cadence (multiple versions in the first several months) suggests an actively maintained project, and the attention to driver workarounds shows someone's actually running this on real machines. If you've been curious about CFD but put off by the CUDA-only landscape, this is a reasonable place to start. Just read the license carefully if your use case is commercial.
Follow @githubprojects for more developer tools and open source projects.