For the love of god people are so overly pedantic. Here is the history of how things came about.
Everyone(almost everyone anyway
) kept saying 390x was 4GB because HBM1 was supposedly limited to 1GB stacks and people assumed you wouldn't be able to fit more than 4 stacks around a gpu for various reasons, yield, cost, power usage(1TB/s bandwidth is overkill and would use more power for no particularly good reason).
I from day one said, based on the entire history of memory, that HBM 1 was a MINIMUM spec, and HBM2 was also a minimum spec. Meaning if you memory meets all the requirements of HBM1 but not all of the HBM2 specs it can't be called HBM2, that is all. It doesn't mean HBM 1 couldn't have 8 hi stack, higher speed or double capacity, just that when you add them all together in one go you can call it HBM 2.
Those slides show two things, one being that 8-hi stacks on HBM 1 are very much doable and that two, what I had theorised for APU's ages ago but seems to be coming for GPU's is that they are going to effectively put two memory stacks on one connecting base die, then connect that to the interposer. So you still have 4 memory stacks on the interposer with the gpu and only a 4x1024bit HBM memory controller on the gpu, but each stack is effectively two stacks stuck together to double the capacity. That is why there are 8 2Gb's shown in two stacks on top of a base die.
Originally everyone was saying more than 1GB per stack would be impossible due to the HBM 1 limits, this is clearly untrue, they've both effectively doubled the capacity in one direction with double height stacks AND they've doubled capacity another way by melding two stacks together to mimic one bigger stack. Thus HBM 1 will get at least 2 and 4GB 'stacks'... regarding a stack as one chip you would stick onto an interposer(even though both the 2 and 4GB stacks would really be 2 stacks stuck together pretending to be one stack).
