Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Perhaps not, but (and this may blow your mind) other computers have upgradable memory as well as massive PCIe bandwidth. Max Studio maces out at 128GB memory which is pretty small.
In turn, it may blow your mind that none of the Nvidia Grace Blackwell logic boards have upgradable memory either. Neither will the next generation. That’s just not how modern integrated-chip systems with advanced packaging work. Nvidia and Apple are on the same page here.

Those computers you are referring to are yesterday’s systems.
 
Gemini 3.0 it is software not hardware. Nvidia supplies 90% hardware for AI. You could say it has a monopoly )
But Apple can lock itself in its ecosystem and pretend it's the best at everything.
Gemini was trained with TPU, not GPU. This will tell the whole story. Besides, there are many big companies started using their own chips to replace Nvidia due to inefficiency of GPU such as power consumption, performance, and workflow.

The reason why Nvidia supplies 90% hardware for AI is because of CUDA ecosystem and compatibility. GPU is just another way to use AI and since NPU started to emerge from many companies such as Amazon, Microsoft, Tesla, Meta, and more, it's just a matter of time before Nvidia lose their markets for AI.
 
EXX-IMG-3403692.jpg
NVIDIA-Grace-CPU-_Official.jpg

NVIDIA-Tesla-A100-in-HGX-A100-Board.jpg


Be aware that Nvidia already made their own workstations with their own ecosystem and even their own server cheats work like a SoC which can be inserted multiple of them at once. So why not Mac Pro with PCIe and proprietary slots/connections just like Nvidia? The 2nd image is an ARM CPU with GPU like a large SoC so it's totally possible to make their own Mac Pro with their own systems.

Making Apple's PC with proprietary connections along with regular PCIe slots is already possible and no reason to try if Apple thinks Mac Pro is important for professional markets.
 
I appreciate you digging that up. However, I'm perplexed by that report, as are many others in this thread:


It's difficult to tell if the numbers represent total sales or revenue, and, of course, that chart changes every quarter. Even more perplexing is that CIRP's data varies widely even within the same year: (both charts are from 2023)

View attachment 2592185View attachment 2592186
This proves why Mac Pro is still important after all.
 
Comparing apples and oranges though. That single NVidia card is drastically more capable than the studio. Get a low end or last gen NVidia and it’ll still cream the Mac. I’ve nothing against Apple, but they are what they are. If it’s powerful enough for your use-case then great, but denying that there are more appropriate systems out there isn’t achieving anything.
Sorry, you are just wrong. Low end nVidia cards have so little VRAM as to be much slower than m3 Ultra Studios.

 
  • Like
Reactions: throAU
Apple's rack mount MacPro with PCI slots is still a good choice for music-creation/commercial/industrial/defense-aerospace applications.

Mac Pro rack, front, Thunderbolt ports, power button, handles, lattice enclosure vents

I suspect Apple won't take this path but one option would be for Apple to only make this option. Perhaps build an Apple stand to hold it on its side for 'deskstide' deployment ( and skip the $400 wheels variant). . Also Perhaps make the Wifi-Bluetooth a plug module.

If just one chassis, I suspect Apple could bring the 'rack' version price down to deskside level. [ The major problem is what does a Studio+xMac-enclosure cost versus a Mac Pro. The wider the gap the deeper the Mac Pro problems. ]

It needs to be more than just a form factor. Apple will likely punt building alternative containers for their systems to 3rd parties. ( xMac Studio). I don't think Apple likely building just a container and is happy to leave that for other folks to do.

The real differentiator is whether Apple can enable a real higher bandwidth PCI-e ecosystem or not. If it was another Mac Pro and the PCI-e backhaul was still stuck at PCI-e v4 levels, then it really isn't viable. Music is isn't a high bandwidth driver ( most of those are legacy , very backwards looking in bandwidth perspective, cards). A fair amount of industrial is the same boat. The Mac Pro 2019 showed up with PCI-e v3 right when PCI-e v4 took off ; that didn't help it long term either. Apple is just going to keep repeating that slow adoption rate then it the system just won't be as competitive.

A likely limiting issue is that a really good PCI-e backhaul system is useless on a Mac Studio. So what need is a different SoC with the Mac Pro ( at least for I/O. Really probably do not need more than 4 Thunderbolt). Need a more affordable way to attach a different i/O subsystem to the the Mac Pro solution versus the rest of the line up.
Apple's 'too chunky' chiplet solution doesn't do that. If it changes it is unclear it will function decompose along the lines for a Mac Pro to benefit the most. (or whether the left over volume is going to be high enough).
If there is no M5 Ultra Mac Pro, then it isn’t on the back burner, it is dead.

The choice to not build an M3 Ultra Mac Pro can be understood, but the same can’t be said for M5 Ultra.

My *guess* is it is not dead, and it will have high-bandwidth networking capabilities beyond the constraints of the Mac Studio. But it will have the same M5 Ultra as the Mac Studio — the thing that is on the back burner isn’t the Mac Pro itself, it’s exclusive Mac Pro silicon.
 
  • Like
Reactions: lusty
That's nothing compared to Mac Pro 2019 with 1.5TB of RAM and 128GB VRAM

2.0TB actually on the Mac Pro 7,1, but I've never seen one configured in that way, I suppose 8x256GB DDR4 3200mhz. Happy to test this if someone will fund the testing for me. :D (What, you want me to give that lovely RAM back at the end....)

With the Puget work stations some of them can do 2.25TB RAM and extreme amounts of SSD storage. They look very hard to go past, especially considering their prices. They have a pretty good reputation of building very solid machines and good support.
 
If there is no M5 Ultra Mac Pro, then it isn’t on the back burner, it is dead.

That is not necessarily true. Depends upon how Apple constructs the Ultra. If the M5 Ultra is just two Max chips stuck together Apple may have sacrificed the ‘extra’ PCI-e lanes the Mac Pro needs but Studio ( and laptops ) do not for something like extra memory bandwidth/capacity ( to boost GPU performance and play the ‘affordable biggest VRAM ‘ game ) or just to bigger bigger function units. Or even just to shave some costs by being incrementally smaller ( as the big dies are now getting any cheaper on newer TSMC nodes ) .


Similarly if Apple shifted to more of a relatively smaller chiplet approach to buildin an’Ultra’ that could involve using more “UltraFusion like” connections and the ‘extra’ die edge space for PCI-e gets assigned to inter-chiplet comms . Again still could loose edge space to memory bandwidth/ capacity . (

Finally, if Apple put zero effort into upgrading the PCI-e backhaul provisioning for the Mac Pro logic board switch it likely won’t be competitive. ( rest of market on PCI-e v5 + CXL 2 or 3 and Apple stuck on 4 and less than 32 lanes . )

The choice to not build an M3 Ultra Mac Pro can be understood, but the same can’t be said for M5 Ultra.

But an Ultra optimized for what ? If it is focused on Mac Studio and maybe Private node compute ( PCC ) deployments then there is very weak PCI-e provision there ( pictures of PCC nodes being priced now don’t have standard slot openings on the back ).


My *guess* is it is not dead, and it will have high-bandwidth networking capabilities beyond the constraints of the Mac Studio. But it will have the same M5 Ultra as the Mac Studio — the thing that is on the back burner isn’t the Mac Pro itself, it’s exclusive Mac Pro silicon.

The reports of specific PCC node SoC makes that doubtful . That likely would be the higher priority design effort . The Mac Pro by itself is too small of volume to get something that can’t be placed in another product . That also cranks up overhead for the Studio’s exclusive ‘Ultra’ .

PCC nodes can be paid for by putting a $2-5 tax on 70-90 million iPhone/ipad/Mac users per year to run their escalated AI queries . There is no lack of volume problem there once Apple gets a decent Apple AI service running .

The PCC SoC likely will have high end Broadcom networking chiplet included , but that very likely will result in a PCI-e lane count ( and probably also Thunderbolt ) sacrifice . That would be a backslide for a Mac Pro. That even less multidimensional workload than things are now . What the PCC nodes need is not a GUI driven workstation for a single person ( would be prudent to dump Apple relative large display engines for PCC nodes. In the data enters there are no monitors hooked to any of those systems as a primary display ).

Apple might be about to substitute a different Mac Pro appropriate I/O chiplet for the Broadcom one , but I wouldn’t hold my breath and it wouldn’t be M5 generation. Apple doing a Rip van Winkle on the Mac Pro into 2027 would be relatively consistent with the last decade ( wake up every 4 years ) .
 
In turn, it may blow your mind that none of the Nvidia Grace Blackwell logic boards have upgradable memory either. Neither will the next generation.

Technically not . The DGX station is using SOCAMM memory modules ( Nvidia was pushing their own SOCAMM solution , but there is a better standard now )


This is more about ‘repair’ passibility than it is ‘upgrade’ though . Those SOCAMM are not going to be commodity parts . Much of the talk about ‘upgrades’ is really about commodity parts over extended time .


The DGX spark does not and likely will not next gen .
Grace hopper didn’t .


Where want to pack most capacity in least amount of space at better perf/watt levels … yes Apple and Nividia on the same page . Doubtful that Apple will pick up SOCAMM .
 
Now perhaps you personally would not care, but presuming that M6 card will probably cost over 50% of the price of the new M6 model, how many would just not buy a new machine? Or wait until the M7 or M8 arrives with even more performance and replace the entire machine then?
You mention one of the main shortcomings of the big box'o'slots PC towers...Whereas customers indeed customize these rigs at the time of purchase, at least two factors discourage customization further down the line:
1. Customers consider saving up for a completely brand-new machine that accommodates new standards and technology rather than spend on incremental upgrades.
2. Customers don't want to invalidate the original warranty that came with their customized PC. They'd rather leave it untouched—and the warranty intact—than install partial upgrades.

With those two shortcomings, Apple sees no problem selling fully integrated machines that are only customizable at the time of purchase since that's how customers typically buy them.
 
  • Like
Reactions: CWallace
2.0TB actually on the Mac Pro 7,1, but I've never seen one configured in that way, I suppose 8x256GB DDR4 3200mhz. Happy to test this if someone will fund the testing for me. :D (What, you want me to give that lovely RAM back at the end....)

With the Puget work stations some of them can do 2.25TB RAM and extreme amounts of SSD storage. They look very hard to go past, especially considering their prices. They have a pretty good reputation of building very solid machines and good support.

Wrong, the CPU of Mac Pro itself can only support up to 1.5TB.
 
  • Like
Reactions: throAU
We seriously need Mac Pro more than ever especially since the sale percentage of Mac Pro is actually equal or even higher than both Mac mini and Mac Studio.
 
Do you think Apple will also start to sell its own Ai compute power into the market?

No. ( as a compute service ) . Are they going to eventually charge for Apple Intelligence ? Probably yes for power users.

It is Doubtful Apple if Apple Intelligence gets good that they will have that much excess capacity. Howler they will want to charge users who almost constantly keep the hardware loaded up with ‘heavy’ workload. ( e.g., iCloud Drive free for 5GB and not free for 1TB ). If Apple AI flops they aren’t going to sell this to anyone else.

Apple very likely is not going into the generic IT in the cloud business . Xcode Clould is very narrowly focused on build to deploy to App Store service . It isn’t a robust code to the clould developer service ( e.g., store your code eslsewhere ) . iCloud storage for apps is largely what Apple Ned’s/wants for their own apps ( mainly running on their own OS variants ) .


If it has a power consumption advantage using it's own Apple Silicon it could price itself very aggressively compared to others.

Apple does want to run lots of hyperscale datacenters . They subcontract out elements of iCloud services workload now . Apple has a 1+ billion users . Right now only a narrow subset of them have Apple AI capable devices. However over next 8 that subset will get increasingly larger. They need zero addition clould compute sales team to expand to 3rd parties to get that growth .

Also pretty good chance Apple will push some subset of their internal training needs into their own clould once they have a more data enter optimized SoC to use. Probably lean on a large chunk of Google TPU cloud as long as Google owes apples several billion dollars per year ( e.g., take a subset as barter trade and don’t have to claim that as 100% renewal energy as not an Apple data center. ) .

There are multiple competitors coming with more effience inference.


Who is to say this hardware kit dev does not trickle down in some way to the retail product line either.

I don’t think the CPU design basic block is going to be different . Nor the GPU block , mememory block NPU block , ‘ minimal PCI-3 block etc are going to be different from M-series. It is more like the m-series is trickling down to this PCC SoC . The PCC will likely loose some I/O and gain some Broadcom. I/o by swapping a chiplet or two.

The Mac Pro SoC problem is primarily an economic one ( not enough volume to pay for effort and resources ) not a technical one . Apple isn’t trying to make everything for everybody .
 
I can drive 12 4k displays on my 2019 Mac Pro, and have PCI slots and USB ports to spare.

The most any Apple Silicon system supports is 8.

A Microsoft surface tablet can drive more displays (eGPU support) than an AS Max Studio.

"Compute" is not the be all and end all of GPU requirements.
 
  • Like
Reactions: basehead617
Mac Pro user (2019 16-core Mac pro, 96GB, W5700X) here. I think it's often not even about speed and processing power, but expandability, modularity and quietness. For me, the Mac Pro was the only feasible machine, since smaller Macs at the time (2020) either didn't support my multi display setup (the M1) or their fans were just always at 100% CPU, thermal throttling constantly, and the noise was just really distracting. The Mac Pro, even the intel one that I have can hold a sustained workload for hours on end while staying quiet. (although drawing an absurd 600+W when under load for what an M4 max can do at 60W. Even in Idle, the 2019 Mac Pro draws ~150W).

The clustering of Mac Studios has proven that Apple can still offer a very tempting option for LLMs, but the question is whether they’re happy with the cable mess or whether an enclosure with several Apple C/GPUs is more desirable.

The Mac Pro needs reframing as a science and LLM computer rather than the old thinking of a video editor.
What a stupid design by connecting tons of TB cables while Mac Pro can easily achieve that.

I wouldn't call the Mac Studio clustering stupid, but I agree with the idea that this could be also the architecture of the new Mac Pro and this has also been discussed in this thread in other places. I think apple could build on their clustering capabilities and build some kind of integrated cluster where, in the Mac Pro case, you can get up four "Mac Studios", e.g. 1-4 M3 Ultras or M5 Ultras or whatever the next gen is. It would essentially function like a Mac Pro, but with unprecedented performance for AI workloads, and there could also be a rack version. The additional price over the individual Mac studios would make sense for data centers, because paying someone to even mess with all the cables and third party rack mounts is is more expensive than just putting in the rack Mac Pro. It would also still make sense for smaller racks, like in recording studios, where you won't need 4 M3 Ultras, but you need the rack mount.

Another Tim Cook drop the ball IMO.

There are a bunch of reasons why

1. The M-chips don't have multiprocessor support, and at the same time, they pushed their top end chips into the Studio and lower tiered macs. So where is the opportunity to distinguish the model line from a performance perspective?
2. The major use of PCIe slots would have been GPUs. Guess what Macs don't support the latest and greatest of?

Looking at other uses of PCIe, not sure...

* Server drive cards - no Mac OS Server, why would you bother?
* Network cards - would be nice to have an ONT switch, but again, mostly server use
* Sound cards - Apple doesn't really give a **** about creator markets such as sound production anymore. If they did, things like SoundSource wouldn't be necessary. Get an external USB audio interface like everyone else.
* Capture cards - This has also moved to external usage.

Sure there are other niche usages. But the reality is Apple is not a "pro" company anymore in that sense. Hell, they are the iPhone company with a legacy PC business.
I have an additional USB card in my Mac Pro, a Raw video capture card to capture the HDMI from hardware samples I work on (or use a mirrorless as a nice webcam through HDMI) and I also installed an additional SSD. All of this would be annoying with a Mac Studio and it would be a bunch of clutter on my desk for this, with a PCI-E breakout box, USB hubs, hard drive enclosures... unless I start using some kind of rack again. And then the Mac Studio, additional external devices, rack and work time to put all of that together would be more expensive than just a Mac Pro tower that is essentially a "rack" of its own
I can drive 12 4k displays on my 2019 Mac Pro, and have PCI slots and USB ports to spare.

The most any Apple Silicon system supports is 8.

A Microsoft surface tablet can drive more displays (eGPU support) than an AS Max Studio.

"Compute" is not the be all and end all of GPU requirements.
I drive 3 4k 144hz displays, a 1440p60hz display and a load of other stuff... I'd have problems with the Mac Studio and would definitely need a big and really fast hub to connect everything. I'm not saying I couldn't make a Mac Studio work, but I think it would be so much pain/work that I'll at least miss the Mac Pro.

But also, it's funny how the 2019 Mac Pro was available with up to 2x ~23 32-bit tflops of GPU, something that even the M3 ultra is still far away from... (I think the M3 Ultra pushes around 28 tflops which is impressive, but still shy of what was available ~5 years before, with dedicated GPUs, for users that really need raw floating point processing power. That's not me, though, and the 5700 is sufficient for everything I do. The important thing for me is that I have a lot of I/O and it stays quiet driving all of it.
 
Last edited:
  • Like
Reactions: Omega Mac
Wrong, the CPU of Mac Pro itself can only support up to 1.5TB.
See the post in the Macrumors Mac Pro forums, W3265M and W3275M supports up to 2.0TB.

Note also that older 5,1 Mac Pros here are running with 128GB and even 256GB RAM, configurations Apple didn’t offer.

Perhaps you can head over here and tell them also they are all wrong?


 
  • Disagree
Reactions: 1d1otic
Intel clearly confirms @avro707 information. Perhaps you should investigate facts, before you make any assumptions?

View attachment 2592596

Intel Xeon W-3275M

It appears the 2TB configuration must be 8x 256GB DDR4-3200 on the Mac Pro using W3245M, W3265M and W3275M, with the memory running at 2933Mhz speed.

I haven’t seen anyone showing that configuration - obviously it is very expensive to obtain that RAM even before the AI boom started exploding prices into the stratosphere.

Note that Apple itself didn’t offer the W3245M, only the W3245 (non M).
 
See the post in the Macrumors Mac Pro forums, W3265M and W3275M supports up to 2.0TB.

Note also that older 5,1 Mac Pros here are running with 128GB and even 256GB RAM, configurations Apple didn’t offer.

Perhaps you can head over here and tell them also they are all wrong?


Intel clearly confirms @avro707 information. Perhaps you should investigate facts, before you make any assumptions?

View attachment 2592596

Intel Xeon W-3275M
Tell that to OFFICIAL info from Apple. Wanna blame me? Blame them.
 
Tell that to OFFICIAL info from Apple. Wanna blame me? Blame them.

There is a forum member using DDR4-3200 in his Mac Pro and it is working.

There is information from Intel and from users of these machines.
 
Last edited:
  • Haha
Reactions: 1d1otic
There is a forum member using DDR4-3200 in his Mac Pro and it is working.

There is information from Intel and from users of these machines. You have chosen to ignore that.
And yet, I provided the OFFICIAL info directly from Apple. Truth hurts?
 
Where does Apple explicitly say it does not support 2TB?
Even the iMac Pro supports up to 512GB, while the Official Apple info says up 256GB
Apple's official info literally mentioned that Max RAM is 1.5TB. Why dont you check that before you comment?
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.