I’ve been genuinely curious what 🍎 planned to do for a Mac Pro/HEDT SoC ever since they announced the switch away from x86.

High core count machines need cache coherency too, and the best way is with a directory. (Maybe they can avoid building a directory with 24 CPU cores if they implement some hierarchy.)

#apple

https://www.tomshardware.com/news/apple-struggles-to-build-mac-pro-based-on-own-cpus

Apple Is Struggling to Build Mac Pro Based on Its Own Silicon: Report

The company is also reportedly considering moving production of the Mac Pro to Vietnam.

Tom's Hardware

The x86 vendors can take a server SoC and make a derivative HEDT chip, but 🍎 doesn’t have that option.

So are they going to design/verify/build all that logic for one (low volume) product line? Are they able to get by without a directory?

@philparkbot Honestly just do MCM at this point, probably the best solution available.
@matthewconnatser I’m not sure how MCM fixes this?
@philparkbot If you need a powerful processor but can't get it from above, you can try getting it from below and just scaling up from weaker processors. I don't think Apple can justify spending the tens or even hundreds of millions of dollars of R&D for a processor that's just going into workstations. Assuming they won't be going to Intel or AMD, anyways.

@matthewconnatser Right, but the issue here is that you need to keep 24 caches coherent, which is an interconnect issue.

So either every memory access needs to broadcast cache state updates/requests or you build a directory. The former costs power and doesn’t scale. The latter is expensive to design/build/verify.

But I completely agree with the last part - it doesn’t seem worth it for Apple to put in the design effort for an SoC that will only go into HEDT.

@philparkbot I would bet that Apple will just eat the power costs. Not a great solution but they can probably justify it, especially if the core processor side of things is efficient enough. Or maybe they're just not ready.
@matthewconnatser yup, as long as they don’t sacrifice performance, which is unknown atm