Not true re GPU bandwidth/dictionaries. see cudahashcat - we're using it on a switched PCIe bus feeding dictionaries with plenty of bandwidth. You're not limited to mask attacks, which is what you're describing.
Are you sure about that? I'm more familiar with oclhashcat, and it certainly can't keep the GPU busy with a pure dictionary attack. It'll feed the GPU with a dictionary, then apply rules on GPU to generate variants.
