Thanks for this post. This is a really interesting take on why we are where we are at with the Mac Pro.
However it does beg the question: shouldn't the Mac Pro be serving the mid-size job market as you define it given the iMac will eventually serve the small-size jobs? Your post provides a lot of insight into what Apple might be thinking. But it also further convinces me that the gap left in the Mac lineup is a mistake.
From a functionality standpoint, maybe, but from an addressable-market perspective it's less clear (especially since most of the actual need for workstations capable of doing "mid-sized" jobs isn't really in domains where Apple has much strength). It's also the case that Apple's not afraid of cannibalizing itself (consider the recent color gamut changes to the iMac lineup in this light).
I am not trying to justify the classic mac pro, my point was only that you can't justify the new mac pro design based on GPUs with HBM, since it currently uses GDDR5 and works just fine. Regardless, the size of many GPUs is dictated by how big the cooler needs to be instead of the actual PCB size. For instance AMD Fury boards are 12" long to maximize cooling.
If you're looking for strict logical entailment you won't find it; the cMP design is always an option, no?
You keep approaching this from the wrong angle, as if I'm claiming there's something magical about HBM that forces an nMP case, and failing to make that case. I've certainly failed to convince you, but that's not the claim I'm making.
I will try one last time here: the nMP case takes responsibility for cooling away from its GPU boards, at the cost of some trade-offs. The design timeline here is longer than you may realize, which should color what follows.
Without expecting a move to something *like* HBM, adopting an nMP-style case is risky, for all the reasons you guys have groused about for almost three years now: yes, it's an extremely-thermally-efficient case, but you give up *too much* to get that case.
If you expect a transition to something like HBM, you can justify the nMP design, because there's a reasonable argument to be made that adopting the case won't force significant compromises on the GPU.
It's, again, not because HBM is magic pixie dust; it's because, in a post-HBM world, you can be entirely confident that the *entire* functionality of the "GPU" part of a "GPU" board fits into a square ~1.5" (or less) on a side. This can leave you pretty confident you can fit *any* GPU into a case like the nMP without dropping any significant functionality (like, being stuck with half the video RAM or something like that); note that the rest of a board is either support electronics Apple usually does itself, or "dead space" used for cooling (which an nMP case handles).
This makes it look like the worst-case scenario is something like "your nMP has 2 top-end GPUs, but maybe 20% slower due to running at lower power levels"...which, yes, is a compromise, but it's at least a defensible one.
In a world where on-package GPU memory is still a decades-away pipe dream, an nMP design would be madness. Yes, in hindsight, GDDR-based GPUs have favored performance over memory capacity, but that wasn't necessarily a given; but, that's how it turned out, and even the GDDR-based GPUs have a smallish "functional block" these days, which makes it feasible to squeeze a reasonable GPU into the nMP case...but the compromise factor is a lot higher, and the amount of customization is a bit higher (if you're observant, you'll note the GDDR chips on the nMP are packed in closer-together than the norm, which increases the importance of very, very good cooling...consider this especially in light of reliability issues when worked hard).
But yes, nothing in HBM dictates the nMP design, but it makes it easier to justify that design as a low-compromise design (and again, the design timeline here is surprisingly long, making it easier to get things wrong).
Obviously the timing hasn't really played out too nicely; the bigger miscalculation is the power envelope being too low. It's a bit unexpected that "mainstream high end" GPUs have stabilized at such high power levels, but they have (for various reasons), and this means even in a best-case scenario the nMP is going to look like a pretty big GPU compromise compared to the obvious alternatives.
"Derivative" is very subjective. AMD would not redesign Fiji (Fury/Nano) with simply a new memory controller for HBM2. Usually when AMD creates a new high end chip, all the previous chips just get bumped down the lineup. So a Nano "derivative" would simply be the next generation of chips, which in this case will be Polaris. Rumors point to probably one high end GPU from AMD this year with HBM2 on it and 1 or 2 low to mid range GPUs with GDDR5. I am not sure what makes HBM1 "horrible" besides the fact that it is limited to 4 GB, making it unsuitable for professional graphics and the mac pro.
It's horrible from a supply-chain standpoint, like I said. Words aren't there just for decoration!
HBM1 is really only used by AMD, in only one product line; that product line has had at-best "ok" sales (at least relative to the overall market size).
HBM2 is better on essentially every relevant metric, has multiple customers already lined up, and is already beginning volume production. The HBM2 standard also includes a legacy mode making it easier to drive an HBM2 stack from an HBM1-targeted memory controller (it's not drop-in easy, and you lose some of the performance, but it's a lot less work than a full redesign).
If you're Apple, and you might not put out another update for 3 years, do you want to be on the hook to keep buying a dead-end product? Maybe you do, but probably you either stick with the tried-and-true stuff or you pick the obvious choice for the future (if it's ready on time).
If you're AMD, do you really want to be the only customer for a dead-end tech? Maybe, maybe not; it depends on sales projections and inventory levels into which I have no insight.
