There are several benefits to this technology, most notably is for MP machines.
Think about this, the more bandwidth going in and out of the processor, the more data you can crunch. For example, getting to main memory is faster etc. (e.g. this isn't just between the CPU and main memory).
But, you say, isn't 12.8 gigabytes/second enough even for a 3GHz processor? For almost every application that would be more than enough even with 64 bit accesses. In all likelihood it would be pretty good for even a dual 3GHz machine, even using vector instructions, for 99% of the desktop applications. However, for dual or more processors, running at >3GHz, it might not.
I would say it would be particularly useful for things like Xserves used in a supercomputer where there is lots of data to be crunched. This would include things like the Virginia Tech machine, rendering farms (e.g. think Pixar) and the like. This is where most of the benefit would go. (It wouldn't help the speed *between* machines in the cluster, but would speed up each individual machine).
The other area it will help is for dual core CPUs when we see them. For example, you might have a two dual core CPUs which would roughly equate to a quad processor. They would need lots of bandwidth.
I'll be interested in seeing these machines whenever they come out - probably 9-12 months, although like everyone, I'd love a G5 PB (or an Apple branded cell phone) with it tomorrow. ;-)