Now it is alot better

Yeah, SAH splitting did good job and I made some pointer magic ... so now performance runs from 1.2 MRays/s to 4.6 MRays/s (Hee yeah!).

Converged image from path tracing :P ... okay, next time I'll probably use another scene, this one is pretty boring and evil for KD-trees (made out of triangle fans). I'll probably compose mine, and compare results to offline renderers and my own OpenGL 4.1 engine.