So you have a high-bandwidth, but no better latency, L4 chip. And that's kind of hard to deal with and optimized against.
Xeon Phi had a DDR4 + HMC (where HMC was a early competitor to HBM), and getting good performance out of that was unusual. DDR4 ended up having slightly better latency, so the HMC cache would hurt some workloads.
I think Apple would want to release a 64GB or 128GB Mac Pro... but only to upsell its customers to a 256GB or higher Mac Pro. The ability to configure more RAM into a system is key for marketing (even if this RAM is soldered onto the motherboard: like LPDDR4x, it wouldn't be too expensive to make different motherboards especially in a niche $10,000 computer like the Mac Pro).
But if Apple did go with HBM2 itself... well... then its physically unable to go above 32GB (today), or maybe ~64GB (if bigger HBM2 modules come out next year).
--------
Sticking with LPDDR4x just seems like the most obvious, and best, decision for Apple. So that's my expectation. But then again: making a 8-way decoder + fat-as-heck execution pipeline wasn't part of my expectation either. So maybe Apple will surprise us again.
First of all, that's ludicrously expensive. You need an active interposer that's larger than the chip to do something like that. 4x HBM stacks means 4096 wires to run between your compute-device and the RAM, that's a lot of wires in a tiny area, and only possible with advanced packaging (expensive) methodologies.
Second: HBM2 based designs have poor memory-capacity: 8GB per stack means 4x HBM2 caps out at 32GBs under most reasonable configurations.
Going into 2021 with a ludicrously expensive 32GB machine seems a bit weak. I don't think anyone in the high-end market is targeting that level of memory capacity. Even if we're generous and HBM capacity doubles between now and then, I'm not sure if 64GBs is enough for workstation tasks.
HBM works for supercomputers, because when 32GB isn't enough, you just buy another node. A64FX, the current CPU-based supercomputer that uses HBM, only has 32GB/node. Then Fugaku bought 150,000 nodes.