So before we dive in we need to talk a little about HBM. As we mentioned HBM has 128GBps of bandwidth per chip. If you drop in four chips you get 512GBps of bandwidth for your GPU to play with. Right now the most widely available density for HBM is a four stack 1GB module at 1Gbit. In fact SK Hynix only shows this as available in their Q1 2015 product datasheet. This means that to hit 8GB you need eight HBM chips, but you will be pushing 1024GBps of memory at the GPU. Unless AMD designed the memory controller on the GPU to handle this much bandwidth you are going to have some serious issues. There is also the pin consideration, eight 1GB HBM memory chips will take up real-estate on the PCB and the GPU.
Everything we have seen shows AMD built the new GPU for the 3xx series to handle four HBM chips and 512GBps of HBM bandwidth. This means there is a deficiency somewhere or that AMD managed to pull off a massive win and get enough 2GB HBM modules from someone (summoning Loki maybe?) so that they could make production level status. This would be very important if they wanted to launch a non-vapor ware card to the public that can go toe-to-toe with NVIDIA’s latest.
So far the people that we know around the net have not been able to confirm the rumors, but are all thinking that AMD might be scraping the bottom of the supply barrel to make this card work. SK Hynix shows Gen2 HBM having the available density for this move, but it is a ways out on the roadmap. The other possibility is that this is all just a rumor to get people talking about the Radeon 3xx and not the NVIDIA Titian X. Time will tell what the truth is, but so far the rumors just do not seem to hold water.