Sorting of GPU Particles

did i ask something too strange? should i clarify?

Another problem is that different alpha blended geometries could have different blend modes and/or textures/shaders.

Imagine a car with transparent windows that goes through some smoke.
The windows of the car might have some fancy fresnel reflections and refractions while the smoke might be using some simple billboarding approach.

This means that the windows needs some special textures (maybe dynamic) and some pixel shader math that the smoke doesn't need.

If you sort the individual triangles for the glass and smoke they might be interleaved and in the worst case you need to switch shaders / textures between every single triangle (or use an ubershader containing all shaders).
Worse is that you can not access any number of textures in a single shader (maybe you can in DX10 using texture arrays).

Even sorting individual triangles isn't enough since they could intersect.

All of the above boils down to the fact that it's extremely hard to handle alpha blended geometry 100% correct (especially if you need to have some sort of performance).

I think you should accept that you won't be able to handle things 100% correct and start thinking on how you can minimize artifacts instead (while still maintaining some performance).

Another problem with per triangle sorting is that billboarded triangles (i.e triangles facing the camera) will show popping artifacts on camera / particle movements.

In some cases (smoke for instance) it could look better to keep the wrong sorting order (this avoids popping).
Popping is quite bad IMO since it some cases it could cause a large portion of the screen to change from one color to another within a frame.

Radix sort works for positive floating point numbers, and is much faster than bitonic merge sort. I think cudpp's sort routine uses radix sort, and that should be fast enough for even a very large number of particles.