The timings come from running a continuous loop and using hi res timer to check the start and end times of the individual code sections. The original problem was noticed while running the visual studio profiler, and the large performance hit that is taken on the amd system by inserting this kernel and accessing the fbo texture.
The AMD System is 4gb ram, Phenom II X4 955 3.2ghz, running 5870 on cat 13.4, win7 pro 64
The NV System is 16gb ram, i7-3770 3.4ghz, running GTX 670 on 320.49, win7 ultimate 64
The SimpleGL example seems to run fine, but it accesses a gl vertex buffer, not a gl texture which is attached to a framebufferobject. The NV opencl samples have demo showing a kernel accessing a gl texture but it uses pixel buffer objects rather then framebufferobjects.