Hacker News Reader

Comment by allenrb | Hacker News Reader

This article confirms what some friends and I thought right at the time of announcement. You’re paying $3000 for some PCIe slots. And as others have pointed out, there’s very little demand for those slots.

It seemed like a given (to me, at least) that a new Mac Pro would have:

1. Slots 2. Storage expansion 3. The most CPU & GPU cores available in a Mac 4. DDR5 expansion, in addition to the SoC RAM

In fact, we got #1 and a little bit of #2 (two SATA ports). #3 is a miss, since the Studio offers the same configuration. #4 seems like the biggest miss of all.

Overall it just seems like a pointless machine. The number of use cases where a Mac Studio + TB4 enclosure(s) won’t do can’t possibly be large enough to justify keeping it in the lineup.

View on HN

You’re paying $3000 for some PCIe slots. And as others have pointed out, there’s very little demand for those slots.

I think those two details are related in the opposite way people would expect. I think Apple knows extremely well that very few customers want those slots these days. That $3000 takes into account the considerable NRE costs that went into developing the Mac Pro since no other model has PCI slots.

This is designed to be a very low-volume product. One of the few “Pro” products Apple sells where the name is sincere.

MBP is sincerely named too (excepting the weird 13"), and with the same familiarity with their user base. Lots of actual professionals choose these models for their higher spec configurations, and have done so for several years despite increased expansion limitations.

Marc Gurman reported that Apple was making a 4x Max SoC that likely would have landed inside this Mac Pro.[0]

For whatever reason, this did not come true. I suspect that Apple ran into technical issues trying to combine 4x Max dies into one SoC and have all the CPUs, Neural Engines, and GPU cores act as a single chip to macOS. Another reason could be that Apple decided that it's uneconomical to design such an "extreme" SoC for a niche market/machine.

Personally, I think Apple will release an M3/M4 "Extreme" that will go into the Mac Pro. I also suspect that Apple will try to create some sort of "Apple Silicon Cloud" in order to expand the market for such monstrous chip. The traditional workstation is a declining market. Such cloud would also be a way for Apple to serve customers who want cloud workstations.

[0]https://www.notebookcheck.net/Apple-is-reputedly-developing-...

Apple's privacy focus might have driven them to imagine a market for private clouds. Run your company's custom AI on one or two of these monster towers and never need cloud AI again.

(I confess I don't know enough about cloud tech or AI to know how far off the mark I am; I'm just weaving threads out of reading headlines)

Or buy an actual rackable server containing Nvidia's server offerings for neural networks and never need cloud AI again?

Your IT professionals will thank you for getting actual IT grade equipment with actual IT grade management interfaces.

Whenever Apple does something kinda weird, always remember: Tim Cook is a supply chain guy.

Just like the current 13” MBP, just like the old iPhone SE, the “new” Mac Pro exists to fulfill some manufacturing contracts and/or use up existing stock of the chassis.

A “real” Apple Silicon Mac Pro will come in 2-5 years (my best guess is 3), mark my words.

Just out of curiosity, what would you expect from a 'real' pro that couldn't be done in the studio? It seems the integrated RAM has put constraints on how powerful they can create the machine (ie. multiple ultra chips).

The point of the pro, at least in the past, was the ability to build an incredibly powerful, expandable computer, relative to the other offerings at the time. But Apple has moved further toward integration to get wins in perf/efficiency. It seems to be a product that no longer fits in their ethos.

That you can get a machine that's just 2x as powerful as a laptop - at 2x the cost - isn't particularly compelling.

Adding more GPU/ML capacity, and I seriously don't get why Apple was unable to put the SoC and RAM on some sort of exchangeable "carrier board".

From a "real pro" machine I expect to be able to upgrade it for at least three to five years. These things cost thousands of dollars ffs.

GPU, ML accelerator cards with real software support.

High speed interconnect, maybe proprietary.

Potentially multiple sockets if they can’t scale their current design.

Fast, redundant, hot swappable storage.

I super doubt that a low volume high end Mac is going to eschew the new unified memory architecture. Once Intel support leaves macOS, 100% of their kernel/gpu efforts across 100% of their devices will be based on an ARM SoC, so they can go on a code deletion/optimisation spree.

A few months ago there was a leak where they showed Apple had x86 servers running MacOS.

They have the source and can compile for themselves. I doubt x86 support will truly leave while it’s still the only way to do servers.

I'd be surprised if the Mac Pro sold more than 50,000 units/year.

There's no way Apple would invest in all that technology to sell its least popular Mac.

Investing in technology has an advantage of producing reusable tech down the line.

It is very difficult to predict which parts are going to be successful.

And if they manage to create a product definitively better that competitors (like the current gen ARM macbooks), they are going to sell a lot more than 50,000 per year. There are single companies which can afford to buy 50,000 units.

I think Apple would only invest in the tech that you listed if they also want to get into the cloud businesses. Other than that, I don't see it.

No idea, honestly. You're right, the product doesn't really make sense anymore. But at the same time, I can't see them simply dropping it. Maybe they'll surprise us.

Ok about the 13, but aren't the current 14 and the 16 already designed with apple silicon in mind?

Yes, and that's precisely my point.

They made the 14 and 16, but they presumably still had plenty of 13" frames left, so they're just selling both. Jobs-era Apple likely wouldn't have done that, as it causes the product lineup to be a good bit less clear, but in Cook-era Apple, that's an acceptable tradeoff.

We've been waiting for a "real" Mac Pro for a long time. I wouldn't hold my breath.

There is some demand for those PCIe slots. Far fewer without GPUs, but there are definitely use cases where people will pay $3K for mostly IO (audio being an obvious case, network being a more niche one, unaccelerated video is one I’ve heard and don’t really understand but admittedly I’m very far from the target market). And of course the fairly obvious PCIe storage expansion use case, which I don’t think should be discounted (although it probably has been for this gen of MP).

The speculation I’ve seen (on Ars forums) about RAM is generally a consensus that this is a stopgap, with future base volume increases expected and a future multiple of die connections (a so called “Extreme” being a 4x Max). I don’t think RAM expansion is coming back to the Mac platform for the foreseeable future outside of some kind of hypothetical way to add more SoCs.

What even is unaccelerated video?

Just I/O I guess? I’m definitely not the target market so I’m just filling in blanks in my understanding of this as people more knowledgeable talk about the remaining use cases.

More DDR5 is tricky, as the SoC very much depends on the packaged RAM. The M chips physically don't support it.

For $3000 I would expect a different SoC that has a second integrated memory controller with DDR5 support, even if it leads to an asymmetric memory configuration.

This is easy to say, but tape out costs for a reticle size slab of silicon on an interposer are ~10 figures, at least.

It makes zero economic sense for a low volume product like the Mac Pro, even with Apple margins.

EDIT: Also, they would have to seriously compromise performance. A number of GPU cores would have to be removed to make room for the bigger memory controller. The bus would have to be even wider with regular DDR5 DIMMs, and it would suck much more power.

The Studio should have upgradable RAM too IMHO, so they could use such a chip for that too, which would amortize the costs.

There is no space for ~16 DIMMs.

Make it taller by one Mac Mini height, no one would care. Also CXL memory modules don’t necessarily have to be the same dimensions of a DIMM.

The M2 series does not support CXL, as far as I know.

They would presumably use UDIMMs anyway, so maybe there is theoretically space?

Still, that would be one heck of a socket/motherboard. I can't imagine how many motherboard layers and socket pins they would need for 16+ channel DDR5 in such a tiny space... If its even possible to manufacture, it would be hilariously expensive.

Is that true even with a chiplet or north bridge approach? I feel like we’ve seen plenty of examples of CPUs laid out this way. Yeah with a performance penalty but that 192GB of L4 with direct GPU access would go a long way to making that better.

Apple could indeed fab a small memory controller chip, but I am not sure the M2 Ultra die has the interconnect to support such a thing? AMD makes this looks "easy," but it is not easy, and they built for Infinity Fabric from the start.

As said below, the answer is CXL. Apple would be stupid not to support CXL devices on the M3 or M4, though I would not be surprised if they skip it

I actually wasn't familiar with CXL, I'll check it out. Thanks!

It still wouldn't hurt to have as an option for, say, a swap disk in RAM instead of the SSD, or expose it as storage to macOS for use as scratch disks, or even just use it as a buffer/cache for the flash storage. There are all kinds of use cases where it would be beneficial to have a wicked-fast storage medium working between system storage and memory.

You are saying Apple needs CXL support.

Yes, this is 100% true. CXL devices are exactly what the Mac Pro needs.

Just curious but wouldn’t a Mac Studio paired TB4 PCI-E expansion board/dock fill this same purpose? Is there something inherently in the Apple ARM arch preventing this? I’m still on a Intel iMac where such a solution works a charm, so my natural upgrade path would be a Mac Studio and display. Of course you are still limited to your thunderbolt bus but I guess TB4 is good enough for a lot of PCI-E use cases.

TB4 only has 4 PCI-e lanes, so doesn't have the bandwidth for many applications.

Reply on news.ycombinator.com