I have a simple scene composed by a plane with 6 vertices and 4 triangles. The plane geometry and textures are loaded as OpenGL objects and registered for use within OptiX. This plane uses 4 1024x1024 textures in its material. I already know that OptiX doesn't support mipmaps and my performance issues might be from incoherent texture cache accesses. If I reduce my textures to 512x512, my performance increase is significant (look below).

Texture dimension Frame Rate
1024x1024 49 fps
512x512 87 fps

My question is why OptiX is spending time runing a ray generation program that does nothing but fill my output buffer with a fixed color?

I suspect that, in each frame, OptiX need to copy texture data from OpenGL to its execution context, but, 4 textures of size 1024x1024 can give this overhead even not sampling anything?

Here is my ray generation program

RT_PROGRAM void pinhole_camera() {
output_buffer[ launch_index ] = make_uchar4( 255, 0, 255, 255 );
}

My system has 2 GTX580 with SLI disabled.