There are three ways to increase memory bandwidth in a system, generally. First, you can increase memory clock rate. This has its drawbacks—some memory types become error-prone beyond certain clock frequencies, and require more power to run at these high speeds. Memories that work at higher frequencies without losing coherency or using more power require substantial changes and new standards, and that's basically what GDDR5 is.

Second, you can increase the bus width. The Radeon HD 2900 (or "R600" chip) used a 512-bit memory interface, and the GeForce 8800 GTX and Ultra (or "G80" chip) used a 384-bit interface. This requires the chip to have a lot of pins for the memory interface, which is undesirable. No matter how small the lithography of your manufacturing process, you can only fit so many physical pins in so little space, so these wide memory busses with their many pins guarantee a large chip. Large chips means fewer per wafer, and higher costs. This isn't so bad if your chip was going to be really big and expensive anyway, but it's murder on those mainstream and budget graphics cards (which is why they all have 256-bit or 128-bit memory interfaces).

Third, you can transmit more data per pin, per clock. In many ways this is the ideal solution. You get more bandwidth with lower power and fewer pins. Of course, the more data you try to cram into the signal, the trickier it gets. The goal is simple—get to massive bandwidth with small memory busses that can fit on smaller chips.

In data transfer technology, you basically have two kinds of signaling: single-ended and differential. Most of the memory in a PC is single-ended (DDR2, DDR3, RDRAM, and GDDR3/4/5). Many of the bus interfaces are differential (PCIe, HyperTransport, USB), and some memories use differential signaling as well (XDR, XDR2). The difference between these two signal interfaces is that single-ended requires fewer pins or wires.

You can read the entire article at ExtremeTech.