AFNI uses 32 bit ints as array indexes, which means that no array can have more than 2 G entries. This is not a 2 GB llimit, though, since each "entry" (voxel value) could have more than one byte. The upper limit on memory would be imposed by a couple of things. The first would be the operating system, but I don't think IRIX has a 2 GB upper limit (but that may be configurable?). The second thing is pointer size. Since AFNI uses 32 bit pointers for memory access, it would be limited to using 4 GB of memory at once.
It might be possible to compile AFNI to use 64 bit pointers. I've not tried this, since I don't use a 64 bit computer system much. I suspect there might be a few problems in places, but probably not too hard to overcome.
If you are processing a lot of datasets, but each processing thread is less than 2 GB, then on a 64 bit system such as IRIX, you could simply run multiple processes at once (e.g., several 3dDeconvolve scripts).
I don't have such high end systems, so can't go on much further. --
bob cox