Rama Hoetzlein (c) 2007-2025
👉 Project Website: https://ramakarl.com/fluids3
Fluids v.3 was the first very fast, interactive >1 million particle, Smooth Particle Hydrodynamic simulator.
This repository contains the complete history of Fluids.
🚀🚀 NEW (2025): The latest Fluids 5.0 has been published in Nov 2025!
The current version is even faster, uses the CUDA Driver API, can handle >4 million particles, uses full smart CPU/GPU pointers, buffer-based structure of arrays, integrated CPU/GPU switching, with cleaner, simplified code.
Date: 11/23/2025
Hardware: GeForce RTX 4090
# Particles: 4,000,000
- Fluids ver 3.2 ➔ 32 fps
- Fluids ver 3.9 ➔ 71 fps
- Fluids ver 4.0 ➔ 98 fps
- Fluids ver 5.0 ➔ 160 fps
(same hardware & num particles for all versions)
I recommend you build the latest Fluids v5.0.
Run CMake on the Fluids v5.0 folder.
The CUDA Toolchain must be setup correctly for your Graphics Cards model, CUDA Toolkit version and NVIDIA Driver.
Set the CUDA_ARCH_BIN and CUDA_ARCH_PTX cmake variables to match your hardware.
For example: GeForce RTX 5060, CUDA 13.0, NV Driver 581.80 => CUDA_ARCH_BIN=compute_120, CUDA_ARCH_PTX=sm_120
Use ChatGPT to help install the proper toolchain for your specific hardware, or if you don't know what I'm talking about.
Fluids 5.0
- Modernized, 2023
- Cuda Driver API (cuLaunchKernel), no longer requires CUDA Run-time linkage
- Large-scale prefixSums, custom code, >4 million particles
- Full Smart Pointers, CPU & GPU (DataPtr)
- Buffer-based kernels
- Abstract Buffer management (DataX)
- Removed run modes
- Integrated CPU & GPU switching per func (instead of ver 4.0 Run modes)
- Modern folder struct (src, assets)
- Major code cleanup and reduction
Fluids 4.0
- Rewrite, 2022
- Fluid struct removed
- Fluid as buffers (SofA, see fluid.h)
- Buffer-based kernels
- Allows for arbitrary per-particle properties
- New kernels: sampleParticles, computeQuery
- Write to disk (SavePoints)
- Equivalent CPU & GPU code
- Run modes: SEARCH, CPU_SLOW, CPU_GRID, VALIDATE, GPU_FULL, PLAYBACK
Fluids 3.9
- Major update, 2021
- Switched to Cuda Driver API (cuLaunchKernel), no need to link to Cuda Run-time
- Large-scale prefix sums, custom code, >4 million particles
- Allocation helper, cpu & gpu (AllocateBuffer)
- Source at top level
Fluids 3.2
- Minor update, 2013
- Application code (abstracted main)
- GUI added
- Still using Cuda Run-time API (cudaLaunchKernel)
- Identical kernels to 3.0
Fluids 3.0
- Major update, 2012
- New Insertion-sort based
- Entirely on GPU
- Explicit memory allocs
- Cuda Run-time API (cudaLaunchKernel)
- Static/mixed host code (fluid_system_host.cu)
- 3rd party Prefix scan
- Fluid as struct (AofS)
- Kernels: insertParticles, countingSort, computePressure, computeForce, advanceParticles
- Introduced insertion-sort and prefix scan, as faster alternatives to radix-sort for particle search
- GTC 2014 Talk: 2014, Hoetzlein. Fast Fixed-Radius Nearest Neighbors: Interactive Million-Particle Fluids. GPU Technology Conference. Santa Clara, CA. 2014.
Fluids 2
- First GPU version, 2011
- Radix-sort based (Simon Green, 2010)
- CPU-GPU transfers per frame (pipeline bottleneck)
Fluids 1
- First CPU version, 2007
- Smoothed Particle Hydrodynamics (SPH)
- Proof of concept
- Self learning
H Turn scene info on/off
N, M Change number of particles
[, ] Change example scene (loaded from scene.xml)
F, G Change algorithm (CPU, GPU, etc.)
J Change rendering mode (points, sprites, spheres)
C Adjust camera (using mouse)
L Adjust light (using mouse)
A,S,W,D Move camera target
1 Draw acceleration grid
2 Draw particle IDs (be sure # < 4096 first)
~ Start video capture to disk (tilde key)
-, + Change grid density
- Example scenes are loaded from the scene.xml file
All parameters are permitted in either Fluid or Scene sections.
DT Simulation time step
SimScale Simulation scale (see website)
Viscosity Fluid viscosity coefficient
RestDensity Fluid rest density
Mass Fluid particle mass
Radius Fluid particle radius, only for boundary tests
IntStiff Fluid internal stiffness (non-boundary)
BoundStiff Fluid stiffness at boundary
BoundDamp Fluid damping at boundary
AccelLimit Acceleration limit (for stability)
VelLimit Velocity limit (for stability)
PointGravAmt Strength of point gravity
PointGravPos Position of point gravity
PlaneGravDir Direction of plane gravity
GroundSlope Slope of the ground (Y- plane)
WaveForceFre Frequency of wave forcing
WaveForceMin Amplitude of wave forcing from X- plane
WaveForceMax Amplitude of wave forcing from X+ plane
Name Name of scene example
Num Number of particles to simulate
VolMin Start corner of Domain Volume
VolMax End corner of Domain Volume
InitMin Start corner of Initial Particle placement
InitMax End corner of Initial Particle placement
Feel free to contact me if you have any questions, comments or suggestions:
Rama Hoetzlein
Website: ramakarl.com
Email: ramahoetzlein@gmail.com
MIT License
FLUIDS - SPH Fluid Simulator for CPU and GPU
Copyright (c) 2007-2023. Rama Hoetzlein