Actually, I’ve found it depends quite a lot on the kind of kernel you are running. Your performance on a given card will either be FPU or memory limited. Assuming your CUDA kernel can use a large number of threads (512 or more), than the FPU performance will generally scale with the number of stream processors times the clock rate. If all of your working data fits into the on-chip shared memory (16 kB per multiprocessor), then most likely your program will be FPU-limited.
However, some people find that the memory speed is the limiting factor for their program. If your kernel requires a lot of data to be read or written to the global memory on the graphics card, then the width and speed of the memory bus is what matters most. Even the 8800 GTX memory bus can’t keep every thread supplied with data if they only do a few arithmetic operations per float.
Does anyone have information on when the next version will be released. “Sometime in May” as posted by Cyril is a kinda vage especially since it is already late May. 8600 support is of vital importance to our robotics stereo vision project.
If it turns out that the next release will be late we might need to make plans to find a workaround.