yeah it's amazing once you start to work with the gpu things can get really fast. the GPU is much more optimised to such things so calculating vectors is much faster than on the cpu and dependent on the graphics card you can have a high number of stream processors that work parallel so you have much better multithreading than with a multicore cpu.

Well I use shaders alot recently too but I didn't think that it makes such a big difference to put the physics computation on the GPU too, But now that I thought about it again its kind of obvious because you don't have to transfer the whole particles data to the GPU for rendering since it's allready there. I think that makes the big difference. kool work i might give that a try too. More complex particle systems are propably still a lot easier to programe on the CPU especially in terms of organizing things :)

I have implemented the particles in cg, there you can use dynamic arrays, structs and interfaces to stay flexible so I can plug together different forces dynamically as they are needed. Yes you are right having the data directly at the gpu for rendering makes it fast. But also calculation is faster this is why people use cuda to do work on the gpu. All the particle data is stored in textures so the simulation happens by rendering quads and combining values of different textures in different fragment shaders. This is really fast as the gpu works on this fragment stream with multiple processors parallel. I also started to work with the nvidia physx framework that you can use from inside java with jphysx here you have more complex things like rigid body collision, clothes and fluids also all calculated on the gpu.

wow that sounds awesome. I recently only use c++ because I think it's easier to implement your own shader and openGL code! I use straight GLSL though.

Ahh i just saw it was tagged as processing, so that question is answerd :P But source would still be awesome..

it is based on this paper: 2ld.de/gdc2004/
it is written in java you find sourcecode here:
code.google.com/p/creativecomputing/

yeah i also made a lot of progress with the system, so there are more forces, constraints, better performance and everything runs more stable, also tested it to run inside processing and hope to release as some kind of library soon