I suspect folks use the 8mm smoothing kernel in SPM because it's the default. That said, there are some suggestions for using a larger kernel to play nicely with the Random Field Theory multiple comparison correction (which AFNI does not do). In general I see everything from smoothing the size of 1 voxel to nearly 3 voxels. So if you had a 3mm isotropic voxel size, you might use anything from 3mm to nearly 9mm. There's a decent r
eview on MindHive here.
Whichever size you use will impact the amount of smoothing in the data estimated by 3dFWHMx and the necessary cluster size needed for statistical significance in 3dClustSim.