its the size of the connection between the memory and the GPU. the larger that number, the bigger the connection is, and the more data it can move at once (faster), assuming the same clock frequency on the memory. there are a few things that complicate this though. GDDR3 is half the speed of GDDR5, thus a buss of 64bits hooked in to GDDR5 will be around the same speed as GDDR3 on a 128bit buss. also both Nvidia and AMD these days are including hardware compression on textures so the width of the buss has less of an effect on over all performance. for example, a GTX980 has a 256bit buss but is performing on par to better than a r9 290 that has a 512bit buss.
in short, its only good for comparing within generations of cards.