Thanks for the simple explaination
Now can you please interpret what this person in another forum is saying?
"As for the reason I was talking about memory bus width - because it has everything to do with HBM. HBM is a technology that allows to, in a nutshell, stack memory dies on top of one another - thus bypassing the engineering difficulty of creating a chip with a wider bus. Memory bus width for the GPU board is the total sum of memory bus width of the chips - 384 bit R9 has 12 500MB chips each 32bit wide bus, hence total of 384 bits and 3 GB. Now, the HBM basically allows to stack memory dies on top of each other, thus dramatically increasing the amount and memory bus of each chip. Effectively, HBM can reach orders of magnitude higher bandwith, and memory density, at the expensve of engineering difficulty and higher board design complexity. The current HBM allows stacking of 8 dies, each having 32 bit bus = 256 bit bus.Per chip. With typical boards having anywhere from 4 to 16 chips...lets say, if GTX 970 would have had 8 HBM chips instead of 8 GDDR5 - it would have had 2048 bit bus, 64 gigabytes of RAM, and same memory clock - it would have had memory bandwith of 1.792 TB/s. Terabytes per second. That is absolutely immense increase in memory bandwith. And if it had it, but with the same Maxwell core we have today - it would have performed ABSOLUTELY THE SAME. Not even 1% faster - because its 224 GBs coupled with compression is more than the chip can ever need. Even if a chip theoretically can pull more data, doesnt mean it can process more data.
So my point still stands. Nice PDF tho, I suggest you also read some articles about how memory bandwith affects actual performance of a processing unit, and to what degree. If you ever wondered why people still suggest 1600 DDR3 RAM for their PC builds, instead of 2133 that is already available.
EDIT : In case of GTX 970 with HBM, it might just get a tad bit better performance in 4K due to having bigger RAM. However, since its possible to manufacture 8 GB version of GTX 970 with traditional GDDR5 and there is no single GPU chip that can handle more than 8 GB(maybe the next titan will) - the point still stands. HBM will not bring anything by itself - you need to pair it with a GPU chip for which GDDR5 is not enough anymore."