Most recent drivers with 2.6 SDK?
Try vectors = 4 and worksize = 64 in cgminer as suggested
here.
If you're using 2.5 SDK, try vectors = 2 and worksize = either 128 or 256.
One of those should be the optimum setting.
Mind you, it's not only "vectors" in cgminer but also what vector width you want (1, 2, 4).
If you're editing the config file, the lines should look like that:
"vectors" : "2",
"worksize" : "256",
The default number of GPU threads is already optimal at 2, EDIT::
as is the default intensity of 9 but the intensity needs to be set to 9 manually.