As you said, we already have buses that can do as much as even 32GB/s (2 lane PCIe 3). I think the bottleneck wrt main memory is the scalability of DRAM. Having lots of DRAM chips helps you scale _out_ just like graphics cards. To put in comparison, Geforce 8800 Ultra has an ~100 GB/s memory bandwidth available to 128 cores (shared).