Like it or not, Apple needs Mac Pro.

[tenthousandthings]

Are you referring to the new Spark? That’s literally a clone of Mac Studio, only with an anemic CPU, slow RAM, and a rebranded mobile 5070…

I don’t really follow you here. Are you telling us that Apple should make computers like Spark? They’ve been doing it years before Nvidia.

I think they mean the not-yet-available, price-unknown DGX Station (and clones) with the GB300 Ultra which is expected to start at minimum $40,000, possibly much higher.

That chip actually supports @mcnallym’s point about AI Training silicon, as it is derived from Nvidia’s massive investment in its core business. The idea that Apple has any interest in building a business to support something like that is absurd.

 

[deconstruct60]

Are you referring to the new Spark? That’s literally a clone of Mac Studio, only with an anemic CPU, slow RAM, and a rebranded mobile 5070…

Nvidia is at this point has had 3 generations of DGX stations.

https://en.wikipedia.org/wiki/Nvidia_DGX

DGX-1 2017 ... 2 years before the Mac Pro 2019. ( Xeon CPU and V100 GPUs)

NVIDIA DGX Platform

The best of NVIDIA AI - All in One Place.

www.nvidia.com

DGX A100 ( AMD and Ampere GPUs )

Introduction to the NVIDIA DGX A100 System — NVIDIA DGX A100 User Guide

docs.nvidia.com

DGX Spark and DGX Station

NVIDIA DGX Station: Ultimate AI Supercomputer

Powered by NVIDIA Grace Blackwell.

www.nvidia.com

May get another Intel-Nvidia set up in the future. ( AMD making competitive datacenter GPU put them in the 'don't help competitors' zone. )

I don’t really follow you here. Are you telling us that Apple should make computers like Spark? They’ve been doing it years before Nvidia.

Not really the same at the implementation level. The Spark consists of a Mediatek CPU/RAM/Display chiplet and a Nvidia GPU chiplet. (**) Likely will be the case that when Intel-Nvidia two chiplet combo comes along then the Nvidia GPU chiplet will be the same for both.

The 'box with slots' stations are more modular. Even the GraceBlackwell version. Not generic PCI-e slot though. It is really a non standard NVLink socket/connector. ( for the GB generation station have to put a more standard Nvidia GPU card in a PCI-s slot to get video out of the system. The Blackwell 'GPU' unit doesn't really have any. )

Apple extremely unlikely to go down this road though. Amazon is also slapping NVLink variant on some of their stuff. IBM Power had an early gen NVLink deal that largely went no where. Intel took the $5B and doing NVLink also. Nvidia would like for folks to settle on NVLink instead of something standard and portable. Apple just isn't going to go down that road for the Mac Pro ( or any of their SoCs .. Even a AppleData center only one. Probably Broadcom for networking there. Decent chance RoCE (RDMA over converged Ethernet. ) or something more standardized like that. All the inference and/or 'iCloud' data is following is flowing in over regular Ethernet. )

I think the original poster is highly misguide in how AI in the cloud is somehow going to 'force' Apple to do a revised Mac Pro. What Apple needs in their "AI Cloud" is not a box with legacy, generic standard slots.


** Marketing labeled something else in the Spark (GB300 ?) , but essentially the rumored N1/N1X SoC that was suppose to roll out for Windows on Arm solution too. It is still waiting on Windows to catch up to the hardware. (and likely NVidia can make more money selling them as Linux AI stations this year anyway. ). They'd likely need another gen 2 revision of upgrades to be viable in generic Windows box world.

 

[leman]

Nvidia is at this point has had 3 generations of DGX stations.

https://en.wikipedia.org/wiki/Nvidia_DGX

DGX-1 2017 ... 2 years before the Mac Pro 2019. ( Xeon CPU and V100 GPUs)

NVIDIA DGX Platform

The best of NVIDIA AI - All in One Place.

www.nvidia.com

DGX A100 ( AMD and Ampere GPUs )

Introduction to the NVIDIA DGX A100 System — NVIDIA DGX A100 User Guide

docs.nvidia.com

DGX Spark and DGX Station

NVIDIA DGX Station: Ultimate AI Supercomputer

Powered by NVIDIA Grace Blackwell.

www.nvidia.com

May get another Intel-Nvidia set up in the future. ( AMD making competitive datacenter GPU put them in the 'don't help competitors' zone. )

I must admit that I was not aware about the upcoming DGX station — curious about pricing, performance, and availability.

I knew about the older DGX stations, and I don't see much relevance for the current discussion. These systems were more like stand-alone Ml servers for companies who wanted on-premise hardware for training, hardly personal workstations. From what I gather the GDX A100 pricing was in six-digits, the box weights over 50kg, and dwarfs even the largest workstation towers on the market. And it was widely considered to be extremely niche and overpriced system that was never sold through retail channels. I think if we are discussing workstations we should limit the attention to end-user systems, not niche enterprise-only stuff.

 

[mcnallym]

Nvidia DOES have workstations and it calls DGX.

And again, Apple is only limiting their own hardware so that many professional apps are either gone or restricted. Do you really think that's normal? Since Apple needs AI more than ever instead of idiotic Apple Intelligence, it's just a matter of time before they make their own severs with their own chips.

You might say why then tell that to many companies using their own chips for their own models and workflows such as Google, Meta, Amazon, Microsoft, Tesla, and more.

You mean people have developed there own chips to use in DataCentres to do AI, which is very different to end use machines which is what a Mac Pro is. What you need for the AI development and what you need to access AI from End User Computers are not the same. Hence why they are all specialist solutions.

Whilst you can attach a keyboard mouse and monitor to DGX they are also designed so that access them over the network. You need never attach keyboard, mouse, monitor to one as can even setup over the networks and use as a network appliance. They are not anything like a Mac Pro but specialised systems and try setting up a Mac Pro without keyboard, mouse, screen.

If you look at the DGX Spark then very Studio like even down to unified Memory.

IF Apple decided to build their own AI hardware for AI Development then would be DataCentre solution not a workstation like a Mac Pro or Mac Studio. So same as the above companies wouldn’t need to be an End User Computer which is what Apple sells as a company.

IF they then decided to sell that hardware out to other companies to use for AI then again would not be a Mac Pro. DGX Spark resembles a Studio more then a Pro.

Apple doesn’t even see the Studio as worthy of every M generation and the Pro even less. So whilst they will put hardware in such as the NPU and added Tensor Cores to M5 then they are more for accessing AI than developing AI. AI usage is useful for every machine from phones to Studio how ever AI development is specialised so would be niche use within what is for Apple an already low volume product.

 

[deconstruct60]

Apple doesn’t even see the Studio as worthy of every M generation and the Pro even less.

It is probably not 'worthiness', as much as economics. Throwing a SoC that large into the trash can every year likely would be extremely expensive. Almost nobody does that. There is a high catch-22 with being no other customers for Apple Silicon other than Apple. When the Studio ( and/or Mac Pro) walk away from an Ultra what other product can that SoC go into to be continued to be sold. The iPhone passes chips down to the plain iPad ( and Mini) , AppleTV. The Watch at this point passes chps down to the Home Pod . The plain Mn is passed from iPad Pro to iPad Air.

The Max in the MBP 14/16" helps sometimes help pay for UltraFusion that it is never going to use. Flip side of that Apple runs the Ultra version long or just skips (so the laptop Max isn't saddled with that and other I/O it isn't going to use at all; 'extra' PCI-e ).

Now that the 'Ann Pro' is a separate die from plain Ann, Apple is also looking for more products to put that in because tossing it into the trash can each year probably is too much even for iPhone Pro yearly volumes.

Similarly, folks who spend $8-10K on their systems do no churn their hardware at the same rate folks who pay $80 , or $800 do . Yes there is a 'low end' Studio variant that is close to $2K but it is pragmatically also coupled to $4-10K models also.

 

[deconstruct60]

Take a look at Gemini 3.0. Google trained their own AI with their own chips, TPU. Which means, Apple can also do it as long as they can make their own powerful chips.

Err, there is an even simpler solution there. Apple just trains its large models on TPUs also. Google is cutting Apple a double digit billion dollar 'check' every year. Apple could easily take $1-2B in TPU time on Google servers and still have a horde of money to throw into the Scrooge Mc Duck money pit.

Apple doesn't need to do training on their own equipment with their own OS. They don't run the vast majority of Apple services on macOS (Linux) ... why would the AI training have a required need to be on macOS? Corporate accounting and ERP solution. Not run on macOS either. Very good chance a big chunk of Apple Silicon compute design simulation/tools isn't on macOS either.

There are reports that Apple Intelligence Inference will run on Apple Silcion (PCC nodes). But that is different and little absoluted required need for that to be the same vendor for inference as to compute. Apple inference since deploy on Apple Silicon has more of a pressing need there, but the training can be done on whatever is appropriate. ( if the data set isn't enormous the data center centric TPU/Nvidia solution are needed either. )

The core of PCC node OS is about as much iOS as it is a subset of macOS. It is neither one and isn't deployed at all as a consumer retail system OS.

 

[1d1otic]

Today's news.

Now who said Apple wont gonna develop their own chips, huh?

 

[tenthousandthings]

Today's news.

Now who said Apple wont gonna develop their own chips, huh?

More like last year's news, since this story broke on December 11, 2024. The codename is Baltra, and the project has been known since then. It is reported to include Broadcom networking. It is most likely a variant of M5 Ultra designed for Apple's Private Cloud Compute servers.

There is no evidence it will result in the sort of Mac Pro workstation you envision, because it does not mean Apple is building silicon for AI training. Instead, Apple is building silicon for AI processing. They are not the same thing. You are jumping to conclusions. Several people above have tried to explain this distinction to you, but you reject it, which is okay, but this old story is absolutely not proof that you are correct.

 

[leman]

Today's news.

Now who said Apple wont gonna develop their own chips, huh?

Good morning! They’ve been running their private cloud on their own chips for over a year now. Nice of you to finally catch up!

 

[theorist9]

More like last year's news, since this story broke on December 11, 2024. The codename is Baltra, and the project has been known since then. It is reported to include Broadcom networking. It is most likely a variant of M5 Ultra designed for Apple's Private Cloud Compute servers.

When the story broke Apple was most likely using M2 or M3 Ultras (the M3 Ultra Studio was released in March 2025, so it's possible Apple had some M3 Ultra chips available beforehand).

But even today, I'd be surprised if they have M5 Ultras (or a variant thereof) to use in their servers. Why do you think they likely do?

The base M5 was only released two months ago, and the Ultra is the most difficult chip for them to make, and hence always appears much later than the base chip.

For instance, the M3 Ultra wasn't released until 16 months after all the other M3's, and they weren't even able to make an M4 Ultra in time to release it alongside the M4 Max Studio.

 

[tenthousandthings]

When the story broke Apple was most likely using M2 or M3 Ultras (the M3 Ultra Studio was released in March 2025, so it's possible Apple had some M3 Ultra chips available beforehand).

But even today, I'd be surprised if they have M5 Ultras (or a variant thereof) to use in their servers. Why do you think they likely do?

The base M5 was only released two months ago, and the Ultra is the most difficult chip for them to make, and hence always appears much later than the base chip.

For instance, the M3 Ultra wasn't released until 16 months after all the other M3's, and they weren't even able to make an M4 Ultra in time to release it alongside the M4 Max Studio.

Well, yes, Cook said recently they are shipping M2 Ultras in Arizona right now for that purpose.

I didn’t mean to suggest they have an M5 Ultra right now. The news story I linked to said Apple was developing a next-generation chip codenamed Baltra for the PCC servers, to be introduced in 2026. So if we get the M5 Pro/Max in January, and the M5 Ultra in June, then this new server could also be announced at WWDC.

Ultra has always launched five months after the Max. M1. M2. Even the M3 Ultra launched five months after the M4 Max. This could be a result of the production volume of the Max (not just the Ultra) — they can’t launch both the MacBook Pro and the Mac Studio at the same time, so they stagger the release.

There’s no indication an M4 Ultra was ever in the works. Apple said as much. M5 Pro/Max/Ultra is thought to incorporate SoIC “chiplet” architecture, so it’s a watershed, a kind of reset or redesign. This major development effort took precedence.

 

[dgdosen]

Jeff Geerling posted a video about Mac Studios and RDMA.

 

[maflynn]

Jeff Geerling posted a video about Mac Studios and RDMA.

Good video, I like how he made a case to say that a single Studio is more then enough then clustering 4 of them - still seeing a 4 node cluster was really cool.

 

[theorist9]

Jeff Geerling posted a video about Mac Studios and RDMA.

Can't resist a comment on this:

"Apple doesn't have any say about what I'm saying in this video and if they did I wouldn't be making it."

Nope, if reviewers rely upon getting loaner review samples from companies, rather than purchasing the product themselves, that introduces bias. They know if they speak very negatively about a product they'll be cut off, which will affect their income stream if their business relies on continuing to get those loaners for product review.

That's why reviewers that truly are independent, like Consumer Reports, purchase all their review samples at retail using anonymous buyers.

A reviewer that doesn't recognize the bias introduced by relying on receiving review samples is fooling themselves. Only by recognizing the potential for bias can they minimize it.

 

[avro707]

The usages we used MPs for 20 years ago are mostly now superbly well served by MBPs and Studios.

Those could be dropped as well and the power they have provided by cloud computing on a subscription basis.

Apple offers just the one basic laptop CPU and memory spec and bigger sizes of screens, or perhaps even just a basic 14” screen and if you need bigger you buy a Studio Display.

And your computing needs are covered by various cloud subscription models per month or per year. More flexible for the customer, more profits for Apple.

 

[theluggage]

A reviewer that doesn't recognize the bias introduced by relying on receiving review samples is fooling themselves. Only by recognizing the potential for bias can they minimize it.

Where in the review does he claim to be unbiassed? He's just completely transparent about the stuff being loaned by Apple, says that no money changed hands and that Apple didn't get any say in what he published.

Anyway, I've certainly seen Jeff Geerling reviews of loaned/donated kit where his conclusion is "you probably shouldn't buy this"... and his review of the Mac cluster is hardly uncritical.

That's why reviewers that truly are independent, like Consumer Reports, purchase all their review samples at retail using anonymous buyers.

Maybe, but if you want Consumer Reports-level reviews you're going to have to pay a subscription for it, because the money to purchase $40k computer clusters has to come from somewhere. If it comes from advertising, then the reviewer daren't publish anything that will upset a large advertiser.

...then, plenty of people are happy to be biassed for free, gratis, nothing. Unless you feel that all the Apple-bashers and equally absurd Apple-can-do-no-wrong posters on this site are being paid...

 

[theorist9]

Where in the review does he claim to be unbiassed?

He effectively says that by claiming receiving review samples will not affect his reviews. That's tantmount to claming no bias is introduced, even if he doesn't use the word. I thought that was clear from my post.

That's just the nature of being human. Even if you don't think it introduces bias, it does, even if the reviewer has the best of intentions. That's been shown to be an issue for medical studies, and those are done by trained scientists.

Anyway, I've certainly seen Jeff Geerling reviews of loaned/donated kit where his conclusion is "you probably shouldn't buy this"... and his review of the Mac cluster is hardly uncritical.

Can't have been that critical, since tech journalists who have been strongly critical, even if that criticism is reasonable, have been blacklisted by Apple. So Apple allows a certain level of crticism, but only up to some limit. And this has been going on for quite a while. Here's a 2012 article from ZDNet:

Apple PR's dirty little secret

Apple PR maintains a blacklist of journalists that it refuses to talk to. This includes any media outlet that posts anything even remotely negative or heaven help you, a rumor.

www.zdnet.com

And this is not a fringe view. Jason Snell, former editor-in-chief of Macworld magazine, has stated that Apple (like other big consumer-product companies) has a blacklist of journalists and publications to whom they won't provide product or invitations to their events.

Maybe, but if you want Consumer Reports-level reviews you're going to have to pay a subscription for it, because the money to purchase $40k computer clusters has to come from somewhere.

Yup, and I'm not saying he needs to move to that model. I'm saying that readers should be aware—contrary to what you seem to be claiming—that receiving review samples inherently introduces bias, unconscious or otherwise.

The exceptions would be if the reviewer is in such a powerful position (e.g., employed by the WSJ) that they know they can say whatever they want, so long as it's good professional journalism, and it won't affect their access to review samples. Or they have enough subscribers that they can purchase the stuff themselves (though the latter still risk being cut off from advance review samples of new product, which is bread-and-butter for many tech journalists).

And of course there is more to this than just product access. Being a WSJ reporter, for instance, doesn't eliminate other forms of bias—e.g., maybe they've developed close professional relationships with people at Apple, and don't want to burn those relationships. That's the other advantage of CR—they keep an arm's-length distance from all the companies they review.

In sum, unless the reviewer goes to extraordinary lengths (e.g., a CR-type model), there are obvious sources of bias. And if we want to be sophisticated consumers of these reviews, we need to recognize this. [And CR surely has its own biases, from other sources; they can never be eliminated, only minimized.] [And no, this does not make CR just like everyone else because they all have biases—it's a matter of degree, and the matter of degree is quite large betwen what CR does and what the vast majority of tech journalist do.]

 

[maflynn]

He effectively says that by claiming receiving review samples will not affect his reviews. That's tantmount to claming no bias is introduced, even if he doesn't use the word. I thought that was clear from my post.

Given what he covered and how he covered it, I don't see how bias would have ruined the technology review. What did he skim over, misrepresent or sweep under the rug?

I think what apple did with rdma is fantastic and shows, not only how potent a single studio is, but how they can be used for LLMs. Do you disagree?

 

[leman]

I think what apple did with rdma is fantastic and shows, not only how potent a single studio is, but how they can be used for LLMs. Do you disagree?

This sounds like a great point to discuss some shortcomings of the technology as implemented. And probably the most significant one is that it’s barely usable at scale. And I’m not even talking about lack of professional cabling or routing, but basic hardware/software management. These are still independent computers communicating over the network, so you need to make sure that they are properly configured and that the software is running. As a hobbyist or a small-time user you can reasonably manage a few studios using a kvm switch or remote desktop, but I don’t see how it’s feasible in a larger environment.

 

[maflynn]

And probably the most significant one is that it’s barely usable at scale

No question, and I don't think any youtuber has even implied that this functionality is something that will replace large scale clusters, but rather for smaller more focused solutions.

 

[leman]

No question, and I don't think any youtuber has even implied that this functionality is something that will replace large scale clusters, but rather for smaller more focused solutions.

My impression is that this technology was primarily developed for their in-house servers, and that we are lucky they decided to expose it to the end user. Other than that, I don’t think it will change the value proposition much. This is not something professionals will buy in volume to satisfy their ML needs. GPU matmul acceleration will be much more important to drive purchase decisions.

 

[theluggage]

He effectively says that by claiming receiving review samples will not affect his reviews.

No he doesn't.
It's a standard disclaimer given by every responsible reviewer who receives free/loan review samples (and, I believe, required by YouTube codes of practice anyway). He's just telling you that he's not being paid and that Apple aren't censoring the result. You're twisting that into some sort of evidence of unconscious bias - without providing one shred of argument as to what was "biassed" about the review.

That's been shown to be an issue for medical studies, and those are done by trained scientists.

I'm not asking Jeff Geerling to take out my appendix. I'm certainly not going to buy a $40k stack of Studio Ultras on his say-so.

Here's a 2012 article from ZDNet:

So, let's see. A even-handed sounding, mostly technical, review of an (already announced) Apple product must be biassed because the reviewer says that it isn't. Meanwhile, a rant about Apple blacklisting journalists brimming with more sour grapes than an opened bottle of Beaujolais left over from last New Year, from a journalist who claims they have been blacklisted (for undisclosed reasons) is trustworthy because the writer says it isn't sour grapes... To be fair, back in 2012, the links providing some background & the relevant blog posts might still have worked, which would have helped somewhat, but today they're not very helpful.

I mean, I wouldn't be surprised if Apple do have a blacklist of outfits that break review embargoes and NDAs on new products or re-post negative rumours, and that it can get "abused" at times. What I don't see in that article is anything sustaining the "anything remotely negative" claim. I do see a lot of pejorative language ("Apple stealing user data") and talk of re-circulating rumours from other blogs.

I think someone who got put on Apple's naughty list for rage-baiting (a reliable source of clicks) might be biassed against Apple... You can get a lot of clicks out of being blacklisted by Apple, too...

Note that "biassed" doesn't mean wrong - but there are lots of clues in the language that should warn a reader this person is not going to give you both sides of the story - and there's always another side.

 

[theluggage]

As a hobbyist or a small-time user you can reasonably manage a few studios using a kvm switch or remote desktop, but I don’t see how it’s feasible in a larger environment.

I believe there are "fleet management" systems for Mac... The various outfits running racks of Minis presumably have some solution for installing OS upgrades etc.

Most of the software in use seems to be written for Unix/Linux with a text-based UI, rather than being Mac OS desktop Apps so you don't need a KVM, just Terminal (or maybe XQuartz)... and presumably any software designed for clustering has some sort of management tool...

I don't think any youtuber has even implied that this functionality is something that will replace large scale clusters, but rather for smaller more focused solutions.

...a point that is made quite clearly in the video.
The take-home is very much "this is an interesting development" and "maybe the Mac Studio has a niche in desktop AI"...

 

[maflynn]

The take-home is very much "this is an interesting development" and "maybe the Mac Studio has a niche in desktop AI"...

No question that this only serves a niche or very targeted use case - still I think its incredibly cool.

 

[Basic75]

Most of the software in use seems to be written for Unix/Linux with a text-based UI, rather than being Mac OS desktop Apps so you don't need a KVM, just Terminal (or maybe XQuartz)... and presumably any software designed for clustering has some sort of management tool...

Right, but you'd also want an ILO management interface, at least at scale.