It’s an unaffordable Mac Mini. They used to make high performance minis and mini servers, now it’s the Studio.
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.
Gurman: All-New Mac Pro Still in Testing, But 'M2 Extreme' Chip Likely Canceled
- Thread starter MacRumors
- Start date
- Sort by reaction score
I also remember when they were unaffordable. Like a crippled LC2 for today’s equivalent of $5000.
No. The Mac desktop line has simply expanded and broadened. Calling a Studio "an unaffordable Mac Mini" is absurd. A base Studio now can do everything that I used to do with a loaded Mac Pro circa 2008 except much, much better. Everything. And the Studio is not "unaffordable." Anyone doing pro graphics expects to pay that or more for graphics competence.
What the next Mac Pro will be remains to be seen, but the Mac Studio is here now and it rocks. Just do not expect it to be a Mini or a Mac Pro, duh. The Studio sits in its own place in the product lineup.
With the "Extreme" chip being canceled, hopefully the multi-core performance on the chip being used will be at least on par with what was offered on the Intel Mac Pro.
While the extreme may have been $10,000, and Apple didn't want to spend money on the R&D on something so niche (and spent how much on self-driving car R&D), it would have been a nice bar to have set performance-wise, which itself would have probably been pretty good for Macs in general.
If Apple doesn't come close to Intel's offerings, it just won't look that good.
While the extreme may have been $10,000, and Apple didn't want to spend money on the R&D on something so niche (and spent how much on self-driving car R&D), it would have been a nice bar to have set performance-wise, which itself would have probably been pretty good for Macs in general.
If Apple doesn't come close to Intel's offerings, it just won't look that good.
14 years later, I would hope so.
Called it.
forums.macrumors.com
Never made much sense to only put this in a Mac Pro. The resources Apple would have to put in for a monster chip like this would be way bigger than what they could get in return from such a niche product.
You're Apple. You just built a 40-core SoC for Mac Pro. Now what?
Mark Gurman is saying that Apple is working on a 40-core SoC for the Mac Pro for 2022. You're Tim Cook, sitting in his nice office, looking at how much money you just spent to make this giant SoC for a relatively small market. In fact, you have to do this every year or every two years to keep...
forums.macrumors.com
Never made much sense to only put this in a Mac Pro. The resources Apple would have to put in for a monster chip like this would be way bigger than what they could get in return from such a niche product.
Of course. But you quote me out of context. I was responding to ikramerica calling the Studio "an unaffordable Mac Mini."
I doesn’t look that good now, if one doesn’t need macOS. If one DOES need macOS, it doesn’t matter how it looks, it’s not like there’s any other option.
Called it.
You're Apple. You just built a 40-core SoC for Mac Pro. Now what?
Mark Gurman is saying that Apple is working on a 40-core SoC for the Mac Pro for 2022. You're Tim Cook, sitting in his nice office, looking at how much money you just spent to make this giant SoC for a relatively small market. In fact, you have to do this every year or every two years to keep...
Never made much sense to only put this in a Mac Pro. The resources Apple would have to put in for a monster chip like this would be way bigger than what they could get in return from such a niche product.
How could you have called it? Your thread's stated premise was that Apple was not only use this in the Mac Pro but in the "apple cloud" to. If the SoC is cancelled not only would the upper end Mac Pro be falling into cancelation now... the data center project would also be screwed.
Without the Mac Pro, the data center SoC wouldn't pay for themselves either. If needed both to be economical then if one segment falls away the whole house of cards collapses.
What didn't make any sense was why Apple would need any data center Arm server packages at all when they can just pay $0.00 in proprietary R&D money and just by them off the shelf from Ampere Computing/Supermico/Inspire/etc. Apple's cloud that they rent from Amazon/Google/etc runs on Linux and the Ampere server packages run that just fine. (actually better than Apple SoC , since they run in 'raw iron' mode with readily availble linux images with supported drivers for data center type networking and storage. )
There likely was no Apple data center "Extreme" SoC at all.
'Called it' that the M2 extreme version was going to extensive? The sky is blue.
If there was no money to do a 'proper' Extreme , then there is pretty good chance there is also no money do a proper Ultra either. That what end up with here is two laptop optimized "Max" dies with some narrow I/O kludge to perhaps roll out a somewhat dubious Mac Pro that still has too chunky "chiplets" that will run into problems in the future. If Apple is stuck in a zone where they have to use the laptop volume to pay for the Mac Pro dies then they are likely going to run into substantive issues rolling forward.
If Apple did the proper diaggreation to save cost using more chiplet focus dies then scaling up would not be a huge cost explosion. AMD's Epyc and 7900 solutions scale just fine. They are more inexpensive than the rival Intel and Nvidia solutions. So the cost do not necessarily spin out of control just because using chiplets. Costs do tend to spin out of control if trying to pound a round peg into a square hole using some Rube Golberg level of complexity to force it to fit.
If was trying to force the laptop Max dies into a quad configuration ... yeah that would probably spin out of control on costs , because it is bad chiplet die. Too large and too chunky.
If what the M2 Ultra in the Mac Pro is getting is a "more affordable kludge" then I suspect there are going to be some disappointed folks.
Apple could have factored a desktop building block that could be spread over the Mac Studio , Mac Pro (and a high end iMac Pro if brought that back). They didn't necessarily need a data center deployment to generate deployment volume.
I suspect Apple has a growning number of chip design build issues ( > $1B cellular modem that is just burning money, AR/VR custom SoCs , etc. ) that probably are stretching them thin.
If the M2 Ultra is just two M2 Max dies slapped together the Studio also doesn't have the volume to support a specific SoC. ( It isn't just the Mac Pro).
There is no pragmatic good way to combine two Ultra packages. The Ultra package doesn't have any sufficient high enough off package bandwidth to connect the two packages is a Uniformly , Unified memory way. There are just eight x1 PCI-e v4 lanes and and a some TB4 controllers ( at x4 PCI-e v3 bundles ).
They could take the building builds that go into a Ultra ( two laptop Max dies ) and build a 4+ die single package .
That has some layout issue problems.
But if they took 4 chiplets that were design to scale to 4 ( or 2 , or 1 ) compute tile then would/could work. ( folks are already shipping designs like that. They'd need to disaggregate without thinking of "laptops first" , but it is doable.
The M1 Max Die is designed to be an elongated rectangular ship in part to fit the internal logic board of a MBP 14". That really isn't a good starting point to design a chiplet that scales. They'd need a different shape and size. ( even more so when get to TSMC N3 , N2 , etc. )
The Macs they sold 12 years ago suited my needs. The Macs they sell now do not.
Mac Pro even without the Extreme chip will still be very powerful.
Apple has priced the MPX W6900 ($6K ) higher than a MPX W6800 Duo ($5K). Both use the same fundamental die from AMD. So two dies are less than one (albeit better binned) die.
A 7900 replacement at the same price point would have a very similar ( so expensive ... how many folks are really going to buy it ) problem as the $10K M2 Extreme SoC.
Apple and AMD collaboratively worked on AMD GPU generational drivers when Apple both 2-3M AMD GPUs for iMac and MBP 15" per year. That volume to amortize development costs over is gone. That was being used to help defray the costs of the $2-6K GPUs for the 2019 Mac Pro driver development. When the Mac Pro has to pay for it all with only maybe 3 MPX modules the costs are going to go down? Probably not.
If the vast majority of end users bypass the MPX W7900 for the low cost, off-the-shelf version there is substantially even less money to pay for the driver work.
This story about "too few users to sell to leads to Apple not interested in making a product" impacts the high end GPU space also.
With the used GPU glut sales of W6800 and W6900 are likely falling pretty fast. That really doesn't do much to 'enable' a next generation for Apple. The MP 2019 user base likely isn't getting substantively bigger. (if leave the core system with a 3 year old CPU ... how is that competitive with 2023 workstation offerings coming from Intel/AMD? )
The Mac Pro was far more dependent upon the rest of the Mac market to help manage kernel/driver/software costs than most folks want to admit. The 580X , W5700X , 5500X all had placements in other Macs (iMac , MBP dGPU).
Not sure why AMD would be jumping up and down at opportunity window of selling another 10-30K units when they are shooting for 100's of thousands per quarter.
AMD Will Reportedly Have 200,000 RX 7900-Series Cards for Sale
The more the merrier!
The price points that Apple is selling at and the numbers likely to drive via Mac Pro sales might make some noticable impact on MI210 or future MI300 series cards. AMD position in the GPU market is not what it was back in 2016-2018. They don't really need for Apple to step in and sell GPU chips they can't sell elsewhere. Even more so to an Apple that would be buying in sub 25K/quarter volume rates. Buying another 20K/quarter MI210 might.
Pretty good chance that folks with a MP 2019 system that want a 7900 will do the same thing that the folks who wanted to run a 3090 30-70% of the time did. Boot the Intel system over into Windows and use the drivers there. Costs AMD/Nvidia/Apple an additional approximately nothing. Certainly what the hackintosh folks are doing.
Yeah, if you need Mac you don't have a choice, but potentially switching platforms or not being able to wasn't really what I was referring to.
I was talking about not looking good in terms of Apple switching to AS vs staying with Intel, at least for the mid to high range desktops.
The M1, M1 Pro, and M1 Max, was amazing at the time in terms of performance per watt and incredible single-core performance. Great for entry-level desktop Macs, and even better for entry-level to mid-range laptops due to the amount of performance for the energy used.
The issue with AS versus staying with Intel is with multi-core performance, Apple is starting to fall behind. There is also the GPU performance which has never been impressive with AS, at least so far.
With the Mac Pro, I suspect potential buyers want impressive raw performance, especially for multi-core and GPU. Unlike laptops and entry-level Macs, energy consumption isn't as big of a factor for these potential buyers.
If Apple unveils the AS Mac Pro, and focuses on energy consumption versus performance in the keynote, this just doesn't look that good, and people may start to question the switch to AS from Intel, at least for Pro desktops.
It isn't likely 'yield'. If a large number of tht 150 were missing for M1 Maxes that would be a problem for the Max sales also. The demand for the Max is going to be much higher ( < $1K per SoC chip being charged ).
If a wafer cost $24K to process then 30 bundles of 4 would be $800/bundle. If even 25% disappear that is still 22 bundles per wafer.... $1090/bundle. If selling them at $5K/bundle were is the problem? The high cost is going to drive down demand. The bundle cost could creep up the $3000 and only 8 bundles would hit break even. The yield is not gong to be that low. ( even more so if went TSMC N3 for just the compute chiplet and refactored a decent chunk of the I/O off onto another chiplet. Then would be sub 400mm^2 . The M1 Max is really not a good chiplet design; too big , too chunky.
Apple could bump yields by actually making a chiplet. Don't widly crank up the CPU or GPU core count. Move to N3 and the die would get smaller. They would get more 'desktiop M2 Maxes out of a wafer also. ( better keep up with Mac Studio sales with few wafers and yields would go up for Studio's and Mac Pros. )
Or TSMC's 3D packaging can't really do high volumes. It wouldn't be that there are tons of defective packages lying around. Just that the 3D production line isn't quite up to Apple volumes. ( More allocated for products that $4-8K price tag demand levels would generation )
There would need to be a shift for Apple. The M1 Max is so large that the 3D process that Apple is using Info-LSI doesn't scale past the 1x reticle limit. For quad set up with very large chunky 'chiplets' they'd have to drop that.
The problem with using those first hand accounts to validate the presense of PCI-e slots is that those accounts did not have any RAM. So that "could allow" really has no grounding with the prototypes.
There is a bunch of hopeful "they'll just slap in there later at the last minute" that isn't really grounded there either. Coherent Mixed memory types isn't 'simple' . That isn't something you toss at beta software at the last minute in a willy nilly fashion. ( Intel spent years getting Optane DIMMs working right with Linux kernel and various apps. )
The specific 'extreme' adjective is somewhat new , but the notion of the "4 chips combo" isn't really out of 'thin air'. Back in 2020 he talked about Apple working on Jade , JadeChip , Jade2C , and Jade4C. Jade (Max) , JadeChip (Pro) , Jade2C (Ultra) all showed up. So that isn't really 'thin air'. 3 of the 4 showed up.
Doing "jade4C" in the first generation did seem to me to be a likely "bridge too far" situation. And the Jade really didn't turn out to be a seperate die. Or the 'Jade' was codes for the overall packages. When it turned out that Apple had 'photoshopped' their initial Max die pictures to chop off the UltraFusion connection even more so. They were so eyeball deep in vast die reuse that they were shipping millions of UF connections that connected to nothing. That was kind of a red flag.
If they were going to apply extra hackery to try to pound the 'round' laptop die into a quad configuration that that would likely get into trouble ( via NUMA issues , scaling , etc. on the "Uniform , Unified Memory" front. )
I suspect this is more an expectation management leak as the Jade4C thing has been out there long enough to build up some pretty substantial following. If Apple has pulled the plug on a very expensive option ( the glare from the 4090 and 7900 + cache option that is bound to following next year being too much ).
Not happening
They started in 2019 to win back such an audience
Then decided to leave Intel
They cannot stick with Intel (otherwise cut off of macOS x86 so no developer will switch to Apple Silicon software)
At the same time they cannot shutdown an effort they started in 2019 otherwise passing off top tier audio/movie company
They are FORCED to make an Apple Silicon Mac Pro
Yes, but as it looks now, far from what the competition offers, and that will make people ask “What was the point?”
Apple 'painted them into a corner' with the Mac Pro 2013 design. It is really not hard to believe that they once again painted themselves into a corner with a Rube Goldberg 4 die design that was off in corner and didn't turn out well.
The "RAM DIMMs" suggestion that Gurman is making seems pretty good chance of being made up though. "he expects" isn't how you frame "source at Apple said". If Apple did something with DIMMs it is probably not what folks think it is (the main system RAM store). [ decent chance not directly exposed to normal apps as homogenous generic virtual RAM. ]
This Ultra could still be just two of those chiplets. Apple just doesn't want to do four in a package.
If they are using some non laptop M2 "Max core count sized" die as a building block it should build 2 just about as well as it could four. Apple could make twice as many 2 die packages with the same set of wafer flow as they could where 4 dies disappeared into a larger package.
They might need one UltraFusion connector to the other "compute oriented die" and another UltraFusion connector for I/O that they split off ( SSD controller , Secure element , Thunderbolt , PCI-e , Display Controller). The >400mm^2 M1 Max die really isn't all that good a 'chiplet'. It is very chunky and too integrated to be a good chiplet design. It is likely being used to save Apple money more so than it was the best chiplet design they could come up with (because it isn't. AMD , Graviton 3 , etc are better disaggreation approaches).
Apple is using the very high price in part to push down the demand volume on the 4 die solution. If it is supprssed too low then doesn't make sense to do it now.
However, if Apple is so die volume sensitive that they have to 'hammer' laptop M2 Maxes into the Mac Pro that is a different issue.
There was a rumor that there was a "one slot wonder" Mac Pro prototype. Shipping that and not cancelling the MP 2019 would make some sense. Apple is still selling the 2018 Intel Mac Mini two years after shipping the M1 Mini.
In that case it would not be 'cancel' as much as not being a more direct 1-to-1 replacement. Not that there was one coming anyway. However, this would have bigger gaps. Apple could sell something that was a better packaged
xMac Studio/Echo III
Pro 3U rackmount enclosure with three slot Thunderbolt to PCIe card expansion system for Mac Studio.
www.sonnetstore.com
just do it with a dedicated logic board than snap in Thuderbolt expansion module.
From the initial article in this thread
"... Gurman said the Mac Pro with the M2 Ultra chip will be available with up to a 24-core CPU, up to a 76-core GPU, ..."
16 P cores and 8 E cores is likely better than 3 year old 28 Xeon W-3200 series cores in CPU compute. Certainly better an 24 or less.
If Apple provisioned two x16 PCI-e v4 lane bundles out of the M2 Ultra then that would match the overall bandwidth of four x16 PCI-e v3 lane bundles of the W-3200 . If stick with the M1 Ultra's eight x1 PCI-e v4 lanes provisioning then not so much.
macOS doesn't go past 64 threads. So there are bigger software issues if want to get into absolute maximum ginormous mult-threaded scores. ( throw 90 or 128 cores at a problems). macOS pretty likely isn't going there.
No good reason to mutate the kernel to do that just for the Mac Pro when the iphones , ipads , and rest of mac line up don't need that at all.
Apple's trendline is to push more multihreaded compute at non CPU cores over time. Trying to keep up with the AMD Eypc/Threadripper Max and Xeon SP/Max is not really the main objective. It is more a balanced "high, but not crazy high" core count mixed with high , but not crazy high clock rate. Apple is probably going to be aiming for a better balanced approach.
Ok. I agree
But taking into account that 1200 in 2010 adjusted for inflation means 1650$ in 2022
So there was an increase but of 20% (definitely no of 65%)