See PrinciplesSmoothing for background.
Choose your favourite smoothing kernel width. This tends to be more of a tradition than anything. Tradition seems to have it that 8mm is a reasonable smoothing for a single-subject analysis, but that a random effects analysis needs greater smoothing, say 12mm. This is said to be because the smoothing has the effect of increasing overlap of functional areas that have not been very well-matched with normalization, but I (MatthewBrett) know of no direct evidence to support this. Probably more importantly, it can increase power for larger areas of activation by pooling more data.
If your expected region of activation is small in size, you may want to use less smoothing. The idea is to match the smoothing kernel to the size of the expected activation cluster (this is not always known, or the size may differ across the brain).
Click on "Smooth", type in the smoothing in mm (say 8), select the normalised - wa* images, wait.
Note that SPM2 and SPM5 save the normalized images with a "w" prefix, where SPM99 used a prefix of "n".