Uploaded October 2016 | Updated November 2026, 5 hours ago
I have implemented GPU computation via aparapi in my PhysX3D project. At high body numbers the Nvidia GeForce GTX 960M (1.40Tflops, 640 shader units @ 1.097GHz) is 7x faster compared to the Intel Core i7-4720HQ (0.44Tflops, 8 threads @ 3.56GHz).
The simulation is being calculated in real time. At 10k bodies the GPU accomplishes 5 steps per second with Runge-Kutta 4th order.
The program currently is 3953 lines of code.
Special thanks to ThePhysics1234 for helping me out with aparapi.
youtube.com/user/ThePhysics1234
More information at: projectphysx.de
Music: Chris Zabriskie - Cylinder Three
I have implemented GPU computation via aparapi in my PhysX3D project. At high body numbers the Nvidia GeForce GTX 960M (1.40Tflops, 640 shader units @ 1.097GHz) is 7x faster compared to the Intel Core i7-4720HQ (0.44Tflops, 8 threads @ 3.56GHz).
The simulation is being calculated in real time. At 10k bodies the GPU accomplishes 5 steps per second with Runge-Kutta 4th order.
The program currently is 3953 lines of code.
Special thanks to ThePhysics1234 for helping me out with aparapi.
youtube.com/user/ThePhysics1234
More information at: projectphysx.de
Music: Chris Zabriskie - Cylinder Three

![PhysX3D [GRAVITY SIMULATION] Vertical Galaxy Collision
Two galaxies containing 2.5k bodies each are set up above each other spinning in opposite directions. Everything is calvulated in real time on a 1.4Tflops notebook GPU (Nvidia GeForce GTX 960M) achieving 22 steps per second using Runge Kutta 4th order integration.
Note how the center parts of the galaxies are pulled in first.
More information at: http://www.projectphysx.de
Music: Chris Zabriskie - Cylinder Five PhysX3D [GRAVITY SIMULATION] Vertical Galaxy Collision](https://i.ytimg.com/vi/BXVAHdS_xUY/mqdefault.jpg)


![PhysX3D [GRAVITY SIMULATION] 5k Body Galaxy @10x Speed
A galaxy containing 5k bodies is simulated on a 1.5Tflops notebook GPU (Nvidia GeForce GTX 960M). The GPU achieves 22 calculation steps per second using Runge Kutta 4th order integration in 3D space. The recording is played at 10x speed. The simulation took roughly 5 minutes.
More information at: http://www.projectphysx.de
Music: Bastille - Requiem for Blue Jeans PhysX3D [GRAVITY SIMULATION] 5k Body Galaxy @10x Speed](https://i.ytimg.com/vi/COha4PW6yZo/mqdefault.jpg)
![Line3D [GRAPHICS ENGINE] Procedural Landscape Rendering
This is GraphicsEngine, the part of Physx3D which draws everything on the screen. It is a complete 3D graphics engine that I programmed myself. It can even render in stereoscopic 3D.
In terms of calculation resources, I do only provide one CPU core for it, meaning it has to run sequential and thus be extremely efficient.
The grid in this demo consists of 100x100 points that are moved up and down via a 3D SimplexNoise (x, y, time) function. Every 4 points are connected with 4 individual triangles and a point in the middle of the four. Yes, you are seeing 40000 triangles being sequentially rendered in real time at more than 15fps. And it is all pure Java.
More information at: http://www.projectphysx.de Line3D [GRAPHICS ENGINE] Procedural Landscape Rendering](https://i.ytimg.com/vi/CPmD9UOVLJY/mqdefault.jpg)


![FluidX3D [FLUID SIMULATION] Taylor-Couette Flow
FluidX3D is a GPU-accelerated real time 3D fluid simulation based on the Lattice Boltzmann Method.
This demo shows the famous Taylor-Couette flow which occurs when in a cylindrical container an inner cylinder is rotated. The fluid then forms convection-like vortices with a specific width and periodicity, depending on the rotation frequency of the inner cylinder, the exact geometry and the fluid viscosity.
FluidX3D has been completely overhauled, all velocity sets (D2Q9, D3Q7, D3Q13, D3Q15, D3Q19, D3Q27) as well as all collision operators (single relaxation time (SRT), two relaxation time (TRT), multi relaxation time (MRT)) have been added and can be freely combined. In my implementation, neither TRT or MRT have any measurable performance impact since the bottleneck is still the memory bandwidth, capped by the hardware to about 420GB/s on the Titan Xp. The swap algorithm is OS-Pull (fastest algorithm, but not as memory efficient) and the data layout is the faster struct-of arrays (SoA). Ive reached the optimization limit with 2.9 Giga lattice updates per second (GLUPs) for D3Q19 on the Titan Xp.
Ive rewritten my graphics library Line3D in OpenCL with the Bresenham algorithm for drawing lines on a bitmap. It can handle up to 2 billion lines per second, which is way faster than all other graphics libraries for primitive shapes out there.
For more information about FluidX3D and my other projects, visit
http://www.projectphysx.de FluidX3D [FLUID SIMULATION] Taylor-Couette Flow](https://i.ytimg.com/vi/DTEc-CfI3b4/mqdefault.jpg)

![PhysX3D [GRAVITY SIMULATION] Osculating Orbits vs. Paths
Today I updated my osculating orbit function. For the internal calculation I changed from Euler angles to the more common angles used as orbital elements. I also added the calculation and visualization of the eccentric anomaly for both elliptic and hyperbolic orbits. With the ecentric anomaly I can paint the orbits so that the color is brightest behind each body and than gets dimmer, which lets one spot bodies on their orbits even without their names being displayed.
More information at: http://www.projectphysx.de
Music: Civil Twilight - Letters from the Sky PhysX3D [GRAVITY SIMULATION] Osculating Orbits vs. Paths](https://i.ytimg.com/vi/FCoBsC3Sols/mqdefault.jpg)