I don't really believe they were ever going to use HMC, it's basically non viable for a huge amount of products.
An HBM stack has a area footprint of something nuts like 42mm^2, it's TINY, it's significantly smaller than a single gddr5 chip. HMC is god damned huge for a reason I can't even begin to understand. HMC in the full 4 link 160GB/s bandwidth layout is 31mm x 31mm package, it is god damned ridiculous, that is 961mm^2, the "small" package at 120GB/s is still 16 x 19.5, which is still 312mm^2. It's completely and utterly non viable right now as really on package memory(with the processor). The few designs I've seen are say multiple FPGA's on a PCB with the HMC separate in far memory mode(not on package).
Currently speaking afaik the maximum interposer size is somewhere between 650-800mm^2.... so putting a high end gpu of say 450-600mm^2 and even one HMC stack is completely impossible, I honestly don't know how they made it so big and thought it was viable.
I also can't imagine a more mainstream product like a discrete gpu using something that isn't a Jedec standard and can't really be relied upon. What if only one memory maker bothers making it(afaik Samsung aren't even sampling it or doing much with it), Micron being very very closely aligned with Intel, what if you make a run of HMC products and Intel/Micron just stop selling it to you, or stop producing it, you're screwed.
Basically there is zero viability to use HMC with the on package variant now or in the near future, but being that the full speed variant is almost 23 times larger... that is a LOT of shrinking to do.
HMC's only real advantage is signalling. It's entire thing really is that it's replaced the parallel signalling with serial signalling. Much like hdd's going Pata to Sata, reducing signalling makes it easier to increase speed and drastically reduces power of the signalling(if done right). Aside from that HMC is all downsides, it's more expensive than any other type of memory, it's monumentally and ridiculously large, it's not a standard, it's long term future is unknown.
Alongside HBM which is cheaper, a jedec standard, it's well, not 900mm^2+ per stack, it will scale in capacity and speeds very nicely. the only slim advantage HMC may have is with how it accesses the memory internally, it may be a little more efficient though I've seen no direct comparison. But the complexity which may have increased efficiency is probably also responsible for the near ridiculous size of the thing so.... meh.
