Anyhow atm HBM doesn't seem to offer much other than increase the cost and thermal envelope of the package.
Nvidia for 3 generations now have been continuously shipping cards with lower memory bandwidth (with heavy focus on compression) than AMD even without HBM and continuously not only remaining competitive but managing in most cases to be the top performer.
Memory bandwidth is already not being saturated even with GDDR5/4 @ 256bits, HBM has some other benefits than raw bandwidth but those for the most part aren't very noticeable as far as gaming goes.
For compute the memory performance is almost completely irrelevant as you are working with very small batches and the biggest latency is in the compute operations them selves most people that build compute setups tend to under clock the memory considerably in order to save power (which can be used to further overclock the GPU) and to reduce the thermal output by as much as possible as they usually stack multiple cards into a single machine.