I understand now, I can say NVIDIA simply does not understand target CUDA developers. Almost every move in CUDA development cries about it. NVIDIA confused CUDA developers with 3D game engine developers, and may get setback. Point of no return was removing of device emulation mode. CUDA developers are scientists, who develop parallel algorithms. If a program does not work properly, they will nor debug every line of the code, they will simply rewrite algorithm. And warp tricks are not for them, cause their CUDA program is totally parallel anyway. CUDA has number of almost useless features, but main features are hard to get. Also they were troublesome too long. So, now CUDA is becoming more special architecture again, where was it some years ago.