It depends. On one hand it's better to have lower memory bus width if the company can achieve the same memory bandwidth because lower memory bus reduces the complexity of the memory controller (i.e., saves transistor space) and the PCB. Therefore, it's normally cheaper to manufacture a card with 256-bit bus than a 384-bit bus. On the other hand, if you are limited by GDDR5 memory speed, you may have no choice but to increase memory bandwidth by means of widening the bus. Higher performing GDDR5 modules may also cost more $ than widening the bus, depending on a generation and some memory controllers (i.e., Fermi) do not work with higher speed GDDR5 memory. Therefore, a lot of has to deal with engineering limitations. Generally speaking though, lower memory bus saves energy and reduces costs. For overclockers a wider bus can be better since for every 1mhz increase in memory speed, there is a larger increase in memory bandwidth if you have a wider memory controller.
This round, NV did well with only having to use a 256-bit bus, which is a win-win from a power consumption and cost perspectives. Pairing 256-bit bus with 2GB allowed NV to hit the sweet spot while AMD used a more exotic 384-bit bus on a pixel / ROP limited architecture and had to incur larger costs of 3GB GDDR5 memory that doesn't have much benefit yet. Unless NV widens the bus though, their chance of increasing performance substantially at higher resolutions is very limited as tviceman keeps alluding, unless we are talking in modes with FXAA/MLAA or TXAA.
From a strategic perspective, I also think NV wants to leave a larger performance increase for Maxwell. Lower manufacturing node on a more advanced architecture will allow that GPU to be much faster than GTX780 compared to how 780 will look vs. 680. With most games most likely continuing to be console ports with high resolution textures and sprinkled DX11 effects in 2013 (just look at 2012), there is also less pressure for both AMD and NV to deliver 50% more performance.
thanks 
Facebook
Twitter