I don’t understand why they can’t just focus on an advanced motherboard on the Mac Pro rather than building new SoC for such a niche low selling product?
Like the old Xeon Mac Pros why not just have the motherboard do all the work and allow you to drop in 1-4 ultras at your discretion? The design of the chipset of the motherboard would be the focus then. If they need to update the Mac OSX kernel, then that seems doable. Certainly that should be a technical no-brainer for them if it’s required to utilize the four ultra cores to the fullest? Maybe that will be a major feature of the next Mac OS release, enterprise server readiness.
Doesn’t this seem the most plausible step for the Mac Pro? I just can’t believe they will spend a fortune designing some frankenstein SoC for a product that makes up a few percent of the sales. But designing an awesome motherboard that supports multiple ultras and all sorts of extra enterprise features, like maybe multiple nvme modules in raid? And maybe like 16 thunderbolt ports instead of PCI slots? That sounds logical to me.
Think about this.
If you are going to have multiple chips in your computer, they have to have a high-speed way to talk to each other.
Whether this connection (think something like Infinity Fabric) is connected via packaging or via PCB is less important than that it simply exists.
There are second-order issues that also matter, like the interrupt distribution scheme has to know about the full number of chips available, not just those on the "local" chip, likewise the coherency scheme has to know about all the caches, not just the "local" caches.
Point is, you can't just slap down multiple chips that aren't designed for multiples to work together and expect anything useful -- how does information, for example synchronize between these different chips.
You can get something that's not optimal (in power, area, performance) by adding "glue silicon" (to provide the interrupt handling and coherency) and using a sub-optimal (slow, high power) connection like PCIe to transport information, but that's not a great solution. You might use it internally for debugging, to write the next version of the OS and drivers, and to test fabric protocols, but not beyond that.
The Ultra solved this by having a (single) row of connections to provide enough connectivity to one other ultra (along with things that had already been sniffed out, like having on-chip support for two sets of interrupts).
But going forward, Ultra does not look like it is designed to grow beyond two. Which is fine! Every sensible design starts by testing the concepts in the easiest case, then growing. Apple began with a single core on the A4, a dual core on the A5, then triple core on the A8X (three is different from two, because with two, there is just "me" and "the other guy"; with three you have to add in the functionality of "me" versus "which of multiple other guys?").
Ultra showed that the design concepts of Apple Fabric scaled across two chips. Next step is to see that they scale to more than two chips, which will be done with M2.
The packaging question (will M2 Ultra+ have four SoCs in a single package, or an MCM based on two M2 Ultra packages, or two M2's side-by-side on a PCB) is much less important than the fact that you need *something* new to scale beyond two SoCs.
In a way it shows something of how well M1 has been received, and how great the M1 brand is, that so many people are clamoring for more M1 rather than an M2! But great as the M1 is, it looks like we have reached the end of its bag of tricks.
