I've no experience with Altivec acceleration, so don't know how much it would improve the speed of AFNI. If it isn't too costly, and the compiler is supposed to paralleize "for free" (without compiler directives), it would be worth a shot.
I hope you use a "mask" with 3dDeconvolve to speed it up - that can help a lot. Program 3dAutomask can generate a mask file for you, from an EPI dataset.