What makes you say that? Apple has a number patents describing various ways to implement an Extreme design and there were reports from industry sources that devices were being tested. There is no doubt in my mind that a 4x SoC arrangement was in active development.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Like it or not, Apple needs Mac Pro.
- Thread starter 1d1otic
- Start date
- Sort by reaction score
What makes you say that? Apple has a number patents describing various ways to implement an Extreme design and there were reports from industry sources that devices were being tested. There is no doubt in my mind that a 4x SoC arrangement was in active development.
maflynn
macrumors Broadwell
I think there was plenty of substance to that rumor, I believe what stopped them was yields, and high cost of manufacturing that SoC
I agree. The entire enterprise was likely deemed to be rather unprofitable. They can't really sell a $50'000 Mac Pro that can only compete with a dual (or even single) high-end Nvidia GPU. There is just no market in that.
We also need to look at the route they took with their own Apple Silicon datacenters. They don't use complex, high-end packaging. It's just "cheap" SoCs arranged on boards and connected to a shared data network. Cheap, simple, and get's the job done.
Amethyst's friend had access to a Studio many months before it was announced from Apple. This was somewhat creditable 'rumor' info. ( yes it is hearsay bouncing through at least one person, but all the rumors are).
This report stated there was at least a test chip for a prototype board.
" ...
I have some info about next Mac Pro?? chips
- Total 40 cores, contains 32 P-Core and 8 E-Core.
- Total 128 GPU Core!!
- A sample board contains PCI-E slot but no ram slot (Doesn't know it exists on Production Mac Pro)
..."
40 core ASI with PCI-E slot is around the corner!!
I have some info about next Mac Pro?? chips - Total 40 cores, contains 32 P-Core and 8 E-Core. - Total 128 GPU Core!! - A sample board contains PCI-E slot but no ram slot (Doesn't know it exists on Production Mac Pro) - Try to put 6900XT on that slot, its not working at all. - Although it is in...
forums.macrumors.com
In retrospect, decent chance that the assumption that was four M1 Max might be misguided. Two larger, M2 gen chips put together with packaging tech that work on larger than reticle sizes. (e.g., 2-2.5x reticle CoWoS) would also be same core count , but not share the underlying costs of the Max die.
There was a path if wanted to make the most expensive SoC possible to build it. Which was rumored also not all that long after the Studio appeared. As mentioned in another post there was a somewhat Rube Goldberg patent. (that looked quite expensive to implement).
"... "Based on Apple's current pricing structure, an M2 Extreme version of a Mac Pro would probably cost at least $10,000 — without any other upgrades — making it an extraordinarily niche product that likely isn't worth the development costs, engineering resources and production bandwidth it would require," he wrote. ..."
Gurman: All-New Mac Pro Still in Testing, But 'M2 Extreme' Chip Likely Canceled
Apple continues to test an all-new Mac Pro with an M2 Ultra chip, but the company has likely abandoned plans to release a higher-end configuration with a so-called "M2 Extreme" chip, according to Bloomberg's Mark Gurman. Subscribe to the MacRumors YouTube channel for more videos. In the latest...
forums.macrumors.com
It would be a little odd for Apple to jump way, way ,way into the super deep end of the pool before anyone did a sanity check to see if the product could even 'swim' at those price levels.
A substantive fraction of the folks clamouring for commodity PCI-e GPU cards for a future Mac Pro solution are looking for 'commodity' (i.e., affordable) as much as the modularity. The modularity is what they count on to drive down the price. (e.g, 'used', 'not leading edge' but useful, etc. )
Err. Doesn't Polymarket ( and Kalshi) do Yes/No wagers/bet ( 'trades' ) . There would need to be a bet for every year (yes/no for that specific year).
The other substantive problem is that the range of chips that Intel offers is at least an order of magnitude larger than Apple's. Intel does lots of products and until recently tried to make almost everything for everybody in silicon.
( Intel has had their 'Atom' line up. Unknown if Wildcat Lake is going to dip quite that far in price. ) There have been automotive and embedded variants. there are 'chip' for DPUs. There are 'chips' for Ethernet. etc.
The Sonnet Technology Thunderbolt PCI-e expansion boxes all come with a PCI-e card capability list. For example, this new-ish box has link to the document.
Echo II DV T5 Desktop (2 PCIe slots for pro DV cards • Thunderbolt 5)
2-slot Thunderbolt 5 PCIe expansion system for pro DV I/O cards with dual Thunderbolt 5 interfaces delivering up to 6000 MB/s per slot for maximum performance.
www.sonnetstore.com
[ this is a direct link to the PDF.
creeping up on being a year old June 2025. ]
"work and actively supported" conjunction makes the list a bit smaller, but not a huge gap. ( mostly a combination of 'old' legacy card and/or smaller player )
Don't have time to download all the drivers for these and scan for DriverKit, but a quick proxy is to use the column which indicates M1/M-series support. Technically that is an overcount as there may be some old I/OKit drivers there that are in zombie status. However, most competent developers in it for the long term moved to DriverKit when reworking for M-series. I'd consider a card with DriverKit drivers in 'actively supported' status.
This list os really a subset. Sonnet doesn't really go way out of their way to list all of the competitors to their own cards. [ Lots of SSD caddy and Ethernet cards work for other vendors where there is basic driver support from Apple and/or the underlying chipset used for implementation. Also not sure they are up to date on ATTO's support where there is product overlap also. There is enough of these 'undercount' to offset the 'overcount' on zombie I/O Kit drivers that got minimal effort port and stopped. ]
So somewhere over 55 cards.
There are some vendors ( Dante, Bluefish, ) that have pragmatically dropped out of Mac PCI-e hardware game **. Folks who use those cards are more likely to claim 'doom and gloom' (as are Nvidia fans). Firewire (from Apple) support has been dropped by Apple so not a surprise that didn't make the M-series cut either.
Now, some folks might dismiss these cards as they work in TB enclosures and not exclusively in a Mac Pro. Would a Mac Pro just one one of these cards. That would be a very low volume usecase. The Mac Pro more so allows lots more combinations. From 55 choose 3 is a bigger number than 55 choose one. However, just because more combinations doesn't really mean an equal number of users.
There are some cards with a footnote of '8'. They have reduced performance on TBv3 which possibly alleviated on TBv5 (e.g., this new Sonnet Enclosure has one TB connection per slot. So slot gets the whole x4 PCI-e v4) There are 9 of those in the 55 ( ~14% ). Note though that this footnote 8 is just one card blowing out the whole bandwidth. If 2-3 of them .. still short even with TBv5. So it isn't just the how many types of cards, but also the number of cards in an instance of specific types.
Outside the list.
There are no 100GbE ( or better cards ) on that list.
But in 2023
June 6, 2023 – ATTO Technology to Support Connectivity for the New Apple Mac Pro - ATTO Technology, Inc.
ATTO hardware and software will also fully support new Mac Studio and macOS Sonoma ATTO Technology, Inc., celebrating 35 years as a global leader of network, storage connectivity and infrastructure solutions for data-intensive computing environments, announced its hardware and software will...
www.atto.com
Benchmark of ATTO FastFrame 4 versus Mellanox on a M2 Mac Pro.
https://www.atto.com/wp-content/upl...tFrame-100Gb-Competive-Mellanox-Benchmark.pdf
"... Test setup: NIC under test installed in a 2023 Mac Pro M2 Ultraconnected to an ATTO Fast Frame N412 installed in a LinuxServer running iPerf. MTU = 1500. No tuning. ..."
Sonnet's new PCI-e card. x16 PCI-e v4
Sonnet M.2 8x4 Silent PCIe Card (Eight M.2 NVMe SSD slots • Add your own SSDs up to 64TB)
Eight M.2 NVMe SSD slots on a PCIe 4.0 x16 card. Add your own M.2 SSDs. For Mac, Windows, and Linux.
www.sonnetstore.com
Not sure why Sonnet doesn't tag this card with footnote 8. A host interface of x16 PCI-e v4 almost certainly would be throttled on any of these TB enclosures.
Can add any PCI-e v5 SSD carrier cards as they come along also. (TBv5 isn't going to fully let those bandwidths work either.)
P.S. There is another thread about new driver that allows a AI compute card from AMD/NVidia to be used over USB4/Thunderbolt.
“Apple approves driver for Nvidia eGPU on ARM Mac” (for accelerated compute) news stories
https://byteiota.com/apple-nvidia-egpu-on-arm-mac-george-hotz-breaks-ban/ “On March 31, 2026, Apple officially approved George Hotz’s Tiny Corp TinyGPU driver extension, allowing Apple Silicon Mac users to run external Nvidia and AMD GPUs over Thunderbolt/USB4 for the first time since 2020.”...
forums.macrumors.com
Those may be quirky and be a virtualized PCI-e connection that wouldn't run a 'raw' PCI-e slot. But conceptually that could have worked on a M-series MPro also (wither the virtualized PCI-e tap dancing ). No video output, just compute.
** Dante seems to have a software virtual soundcard .
Bluefish444 seems to have dropped out before M-series even arrived. MacOSX download software dated 2017
Apple has LOTS of patents for things they don’t do. If an engineer works on it while being paid by Apple, then Apple’s going to get the patent even if it’s something they never plan to do.
Well you see, to me, if an engineering teems does as much work that the company goes through a trouble to submit a patent, means that that's an idea that was at least considered at some level. That might have never materialized as a concrete business plan, of course. So it depends a bit where one makes the distinction. I agree with you that the Eteme plans (whatever they might have been) likely did not even arrive at the pre-pre-pre-production stage. But I do think they were part of a potential roadmap, at least.
Yeah, but yields and high cost of manufacturing are BOTH things they would be clearly aware of before even taping out a single wafer, it’s not like this was the first time they’d ever made processors. By then, they’d already sold decent volumes of high performance processors in shipping systems, so they were also aware of how much they’d cost to build and deliver. With the Mac Pro already at the edge of viability and selling in small numbers, anything above that wouldn’t have been viable from day 1.
It’s greatest effectiveness was in spotting and firing leakers. Toss a few red herrings and see who bites!
maflynn
macrumors Broadwell
That is some crazy tin foil hat accusations - do you have any evidence to say that the M2 Extreme was only a figment, and that its only purpose to was to find leakers?
Who were the leakers and what were their punishments?
“My friend” or “My friend’s dad” stories are always entertaining!
The best part is that always it’s the friend and not the person, so they reasonably would never have to provide any further details than the sparse information they got from the friend. 🙂
I’m on board with this, but that would have been sorted out at the planning stages and would never be expected to be an actual product EXCEPT by those that Apple wanted to expose as leakers. Which they were able to do pretty effectively. They HAD been making chips and had actually designed and shipped the Ultra
Yeah, like something like, ”in 2040, Intel will still be shipping PC destined solutions less performant than the M1 base chip Apple shipped in 2020.” Hm… probably even further out than that.
Even if it was limited to “chips destined for use in retail consumer PC’s”, it’s STILL Intel’s business model to intentionally ship poorly performing solutions. It’s what makes people feel better about their high end solutions costing so much because “it’s multiple times faster than this really underpowered thing we ALSO make!” Make no mistake, Apple does the same, charging more for their higher end solutions, but their low end will always be very much higher than Intel’s.
Hee hee! Do you have any evidence to say that the M2 Extreme was anything more than a figment? This is important to you, so, I won’t disabuse you of that notion. It existed and was awesome. And leakers either don’t exist or are never punished. That’s why we get such accurate information these days from those folks connected to the leakers. They never say things like “Apple was going to, but decided not to”. Especially not Gurman. 😉
Who, of course, was one of the ones saying the M2 Extreme existed, but was canceled.
It’s one thing to know it’s something they’re working on that they never planned to ship as a product. I’d imagine there were a good number of trustworthy folks were aware of that. And, are still aware to this day that it was never going to be shipped even while obtaining valuable patents from the work on it. Someone not as trusted might have been let in on the project and they thought it was just TOO juicy to not tell anyone about it what, they assumed, was going to be an upcoming thing. “All these folks know about it, so it’s not like Apple would know which of us leaked it!” Except the part where everyone else was trusted. 😉
In the end, it’s a harmless enough thing to believe, I suppose. In the world where the Mac Pro has been canceled, just feels way less likely. (And, I’m sure there will be a rumor about an M6 Extreme next year!)
splifingate
Contributor
Last I checked, servers "serve" the stuff workstations *er* use in their 'work'.
It's not just a matter of scale....
keksikuningas
macrumors regular
to be fair, aren't most workstation class CPUs (Xeon W, Threadripper) descendant from the server counterpart (Xeon , EPYC) ?
Descendant? During the Mac Pro Intel era it is the exact same die (for given core count) as the servers in the exact same socket. It is literally the same thing, but binned slightly different (usually for goosing the clocks and perhaps flipping on/off some accelearation.). The large scale set of server customers largely cover the costs of developing the chip. No server customers... no development cost coverage. [ Intel also rebadged some consumer chips 'Xeon'. Those form the bulk of what gets sold as Windows "Workstations", but Apple never used those. ]
And it is that blurring, most non existence, difference why some higher end workstations are used as "one node" supercomputers for code development and then just deploy to more "nodes" on the supercomputer/server.
AMD Threadripper in the lastest generation is somewhat marginally different in the I/O chiplet is a slightly stripped down variant. The rest is really the same as the servers. It is pragmatically just a sibling SKU; not a descendant (child of).
They are coupled more so who is doing the major funding of the effort rather than some "parent child" kind of relationship. The high end workstation SoC is coupled to the Server chip in that they both need a funder to get made. But one doesn't 'father' the other.
What makes you so sure they never intended to ship it as a product? Business decision making at Apple can be fairly quick. Quote on contrary, I'd say that already the fact that the Mac Pro was left floating around for so long indicates that the final decision to limit horizontal scaling was not made that long ago.
I think it's just that Extreme chip was too expensive to make with TSMC's CoWoS that was the only option back in time. The yield could be a problem as well. The manufacture capacity today is also a problem considering everyone's booking it in advance for those AI chips.
Apple went chiplet design this year with their Pro/Max chips, so I assume the Extreme could be possible then with this design, along with Intel's EMIB and Foveros packaging tech.
Who knows, I'd be glad to see MP come back to life.
Apple went chiplet design this year with their Pro/Max chips, so I assume the Extreme could be possible then with this design, along with Intel's EMIB and Foveros packaging tech.
Who knows, I'd be glad to see MP come back to life.
If one of the major contributors was costs , then nothing has changed. Intel EMIB/Foveros are not ‘Crazy Eddie’s insanely low prices’. TSMC having far more customer demand than they have capacity means Intel does not have to offer money loosing pricing to pry customers away from TSMC . Additionally, Intel fab basically desperately needs higher margins ( because still bleeding cash ) . So Intel the bargain discount vendor….. probably not. ( tariffs forcing use of Intel is not going to lower end user costs either )
Apple had chunky ‘chiplet’ strategy before with thr ‘Max’ die. The fact that the ‘Max’ dies are still involved in a chiplet strategy is also not new . So again large reduction in costs when doing about the same thing …. probably not.
From Apple’s description of this new set up it appears much like their old one. For the M1 the Pro and Max basically shared the same ‘top’ ( relative to the die pictures). The Max was mainly just ‘more GPU and Memory controllers and Media accelerator ( and UltraFusion connector)” attached to the ‘chopped’ version. That was two different dies for manufacturing , but share a substantial amount of design costs. This new set up appears to be 3 dies. Mainly Apple decoupled the GPU/memory/media subset from the CPU part for the Pro version also . That is actually more dies to make. It is only cheaper if the yields go up , but the dies only have one UltraFusion connector for their other partner in a pair.
Wafer costs are steadily going higher for better yield can help make die costs tread water ( stay about the same) , but contributing cost to die is going up. Cheaper isn’t where high end chips are going right now.
One connector isn’t going to make Extreme cheaper at all. It is still a chunky ‘chiplet’ ; more than two is going to be partially Rube Goldberg contraption to cobble together. Unlikely that will be cheaper. Apple could use a different GPU chiplet from the Max ( and Pro) version, but have now detached even more from the laptop economies of scale. Again no cost lowering.
Pretty good chance Apple has stripped the Max chiplet duo of the extra PCI-e provisioning that only the Mac Pro needed. That would be a more effective use of the more expensive wafer area to assign that area to something that most Mac deployments for the chiplet.
If something came back it would probably have a new name so it wouldn’t have to swim upstream from the MP legacy features. ‘Mac Ultra’ maybe.
A chip that only fits inside of the relatively very , very low unit volume Mac Pro has major problems if needs different chiplet(s) than entire rest of the line up . The Mini Pro and Studio might have enough volume to differentiate, but the MP just by itself is likely too small. ( Apple AI servers SoCs likely look more like a Mac Studio SoC than an Extreme . Affordable and very numerous rather than maximally expensive. ) .