Due to measurement errors there is always a bit of randomness involved. Occasionally it will even find a surprising new configuration that suddenly beats the one you were using previously.

It may depend on current overclocking settings as well (memory vs. core clock), that's why it's called "tuning".

Thanks to everybody who is posting their launch configs here. It allowed me to improve the heuristics code for people who do not have the patience to wait through an autotune session.

I am still lacking data points on compute 1.0 and 1.1 devices (8800GTX, etc...), and also compute 1.2. Really ancient hardware    And of course, the TITAN!

Due to good benchmark results, I just ordered myself a GTX 570 (Club 3D also, 149 Euros)
Anyone got a GTX 590? This dual GPU card could do 300 kHash/s I guess.

Also I posted one last binary+source code update for today, which includes the improved heuristics for --no-autotune. I found and fixed one more bug appearing on Linux in my 9600M GPU. Most autotune measurements were showing 0 kHash suddenly.

Christian