MP All Models Apple sleeping on the mac pro is comically bad timing

[theluggage]

I disagree, though, with part about I/O bandwidth being fantastically expensive.

It's not the cost of I/O bandwidth per se or that it would push the envelope of technology. The cost of getting any complex chip design into production is high and the price of the resulting product is hugely dependent on economies of scale. The cost of a new Apple Silicon SoC with significantly increased PCIe lanes - however straightforward - would have to be recouped from sales of the Mac Pro alone. Even the Mx Ultra has the shortcut of being two Mx Max dies fused together and that has a $2k markup over the Mx Max. Obviously AMD and Intel have no difficulty making CPUs with massive PCIe bandwidth - but they have a far larger marketplace.

 

[theluggage]

Remember the video where Steve Jobs asked by the chap about Flash, and Steve Jobs basically says that when doing the ecosystem that make choices that some people may not be happy with but fits th Ecosystem.

O/T but Adobe & Google went on to prove Jobs dead right: Flash for Android appeared with great "iPhone killer!!!" fanfare and then quietly fizzled out. For one thing, it turned out that (big surprise - not) lots of existing Flash content designed for PC/Mac just wasn't usable on a tiny touchscreen... Even simple "just video" Flash apps often had their own custom video controls that were either too fiddly or flat out didn't work with touch.

I believe Henry Ford didn't actually say "If I'd asked people what they wanted they'd have said 'faster horses'" but it still rings true.

 

[Fallforward]

People love to get all up in arms about which Macs they think Apple should be building, but we're talking about a 7.6% segment of their revenue. And the vast majority of that 7.6% is from MacBooks. That leaves, what, maybe 2% of their overall revenue coming from these desktops they’re doomed because they didn’t bring to market? Please.

View attachment 2634855

Yes. If non-macbooks are an estimated 15% of total Mac $ sales, they equate to 1.2% of Apple sales. Take out the Mac mini, cheaper price but higher volumes, then the high end non-macbooks - iMac, Studio and Pro - combined are probably less than 1% of total sales. Of those three sharing less than 1%, the Studio probably brings in half and the Pro probably brings in the least $ sales, perhaps 10% of that less than 1%.

Of course, Pro sales might be much higher if Apple was not "sleeping" on it, but not enough to have much impact.

 

[BSDnostalgia]

The cost of a new Apple Silicon SoC with significantly increased PCIe lanes - however straightforward - would have to be recouped from sales of the Mac Pro alone

I agree with you 87%, but, I think the "PCIe" moniker is a little distracting. While PCIe is a kind of "I/O intermediate currency", the object isn't a chassis full of PCIe slots, it is higher-bandwidth I/O to support more/faster SSD inside a "Pro" system like the higher-tier Studio, and, for (ahem) clustering, like Nvidia is doing with Spark, via the QSFP form factor. The increased I/O wouldn't be for one product, but, for the entire upper tier of Studio systems.

 

[Flint Ironstag]

...and here's the point that seems to be sailing clear over your head:

The M5 Max outperforms the mobile 5090 RTX at least partly because ...

Here's what continues to sail over your head: Nobody in the Mac Pro forum cares about your comparisons with mobile RTX GPUs. We can use multiple full-powered GPUs.

Sure thing, you might even get 8x GPUs! Just buy four M5 Ultra Studios (when they come out), connect them together, and write some software that figures out how to coordinate work using the new RDMA feature.

There's plenty of people building up serious machines.

 

[BSDnostalgia]

Yes. If non-macbooks are an estimated 15% of total Mac $ sales, they equate to 1.2% of Apple sales. Take out the Mac mini, cheaper price but higher volumes, then

(etc. quote mainly for context).

I love the way people write off a $33B enterprise - MacOS systems - as insignificant.

Perhaps it would make sense for Apple to split off Macs into a separate company, mainly so that people would pipe down about the market size disparity. Although, at least for the moment, Apple Silicon is one of those very rare instances of true synergy. Splitting them apart would have the benefit, I would hope, of slowing down the pressure to make the Mac systems UI more dumbed-down, more phone-like.

 

[Bug-Creator]

This is the Mac Pro forum.

This the sub section of a forum dedicated to a niche product that has been taken out behind the shed after years of neglect by Apple.

HP Z8, Lenovo Thinkstation, etc

Yep those exist, those can be made by HP/Lenovo/etc with minimal development cost by just using existing PC or server parts. So how do you think could Apple compete in such a small niche when they would need to pretty much throw aways their existing HW stack to develop something from the bare silicon up. Something they couldn't use anywhere.

So no they didn't kill the MPro because they were stupid, also not because they were out to get you. They killed it because it made no economic sense.

 

[BSDnostalgia]

Here's what continues to sail over your head: Nobody in the Mac Pro forum cares about your comparisons with mobile RTX GPUs. We can use multiple full-powered GPUs.

Serious question: what is the "program goal" for "Mac Pro"? Is it limited to discrete GPUs on Intel-based PCIe card cages? Dual-boot MacOS and Windows only? Or, is there a functional/performance goal instead? e.g.

- Faster gaming
- Faster FP64 etc engineering simulations and scientific computing and etc
- Faster crypto-mining
- Faster AI
- Low idle power

What if-- new unified-memory platforms could deliver all of the listed items in the lower section?

 

[SevenTheGamingKitty]

There is as new wave of people setting up computers locally (both personal and business) for AI inference.

The Mac Studio works fine for personal reasons?

For business reasons, that'd be reason for apple to bring back the Xserve, not the Mac Pro.

 

[teh_hunterer]

Not everything is AI.

Ok but if not for AI why would you need 1.5TB of user upgradeable memory?

But even with AI, paging data between main RAM and GPU/VRAM is better than nothing. Certainly better than swapping to disk. Having a giant block of RAM (e.g. VRAM) directly accessible to the GPU is of course ideal but then the question is what is your fallback when you have more data than will fit?

No AI focused build is going to use 1-3 GPUs on a motherboard in a desktop tower and a whole bunch of user upgradeable RAM. You'd be better off going with a system with as much unified memory as you could get with the money instead. This scenario you're talking about is just not a thing.

 

[Fallforward]

(etc. quote mainly for context).

I love the way people write off a $33B enterprise - MacOS systems - as insignificant.

Perhaps it would make sense for Apple to split off Macs into a separate company, mainly so that people would pipe down about the market size disparity. Although, at least for the moment, Apple Silicon is one of those very rare instances of true synergy. Splitting them apart would have the benefit, I would hope, of slowing down the pressure to make the Mac systems UI more dumbed-down, more phone-like.

My writing skills might have failed because I didn't mean to write off all MacOS systems, just the tiny % of MacOS computers called Mac Pro, not to be confused with MacBook Pro.

Integrated System on Chip designs are reducing (not eliminating) the need for PCIe cards. The direction for solving challenges today seems to be more "how can I solve that better within an SoC." So the Mac Pro’s dependence on those cards is becoming more of a limitation - or an investment misdirection - than an advantage.

 

[benholt]

Agree 100%, Apple officially pulling the plug on the Mac Pro in March was spectacularly bad timing. They abandoned their top-tier workstation precisely when local LLM inference made massive unified memory a killer feature.

Nvidia’s RTX Spark announcement this week capitalizes on that exact void. By combining a 20-core ARM CPU, a Blackwell GPU, and up to 128GB of memory (all with native CUDA support) Nvidia is bringing Apple-style memory efficiency to Windows. Apple effectively ceded the local pro AI market just as demand exploded. The Mac Studio is capable, but up against equivalent unified-memory ARM PC backed by Nvidia's dominant AI software stack, it's completely outmaneuvered.

 

[leman]

Agree 100%, Apple officially pulling the plug on the Mac Pro in March was spectacularly bad timing. They abandoned their top-tier workstation precisely when local LLM inference made massive unified memory a killer feature.

Nvidia’s RTX Spark announcement this week capitalizes on that exact void. By combining a 20-core ARM CPU, a Blackwell GPU, and up to 128GB of memory (all with native CUDA support) Nvidia is bringing Apple-style memory efficiency to Windows. Apple effectively ceded the local pro AI market just as demand exploded. The Mac Studio is capable, but up against equivalent unified-memory ARM PC backed by Nvidia's dominant AI software stack, it's completely outmaneuvered.

Nvidia Spark is an equivalent of an Apple Max chip, with considerably worse CPU and less memory bandwidth. I’m afraid you got this exactly backwards - it’s Nvidia who is trying to capitalize on Apples success by copying the Apple Silicon hardware model. The top-tier workstation from Apple is the Studio, and a potential M5 Ultra should have double of Spark’s ML performance for half-precision compute.

 

[Flint Ironstag]

This the sub section of a forum dedicated to a niche product that has been taken out behind the shed after years of neglect by Apple.



Yep those exist, those can be made by HP/Lenovo/etc with minimal development cost by just using existing PC or server parts. So how do you think could Apple compete in such a small niche when they would need to pretty much throw aways their existing HW stack to develop something from the bare silicon up. Something they couldn't use anywhere.

So no they didn't kill the MPro because they were stupid, also not because they were out to get you. They killed it because it made no economic sense.

It's the Mac Pro forum. Sometimes things need to exist even if they're not your biggest seller. But people worship money too much these days and forget this.

Serious question: what is the "program goal" for "Mac Pro"? Is it limited to discrete GPUs on Intel-based PCIe card cages? Dual-boot MacOS and Windows only? Or, is there a functional/performance goal instead? e.g.

- Faster gaming
- Faster FP64 etc engineering simulations and scientific computing and etc
- Faster crypto-mining
- Faster AI
- Low idle power

What if-- new unified-memory platforms could deliver all of the listed items in the lower section?

Whatever the unified memory platform can give, I want 10x that in a tower. And give me 5 more in a rackmount. The goal is to have a halo product. A statement of the best you have to offer to the commercial market. It can't be 2 minis stacked together 🤣. That's a damn disgrace for a company with Apple's capabilities.

The Mac Studio works fine for personal reasons?

For business reasons, that'd be reason for apple to bring back the Xserve, not the Mac Pro.

Xserve and Mac Pro enjoy excellent synergy. You ever installed any together?

Ok but if not for AI why would you need 1.5TB of user upgradeable memory?



No AI focused build is going to use 1-3 GPUs on a motherboard in a desktop tower and a whole bunch of user upgradeable RAM. You'd be better off going with a system with as much unified memory as you could get with the money instead. This scenario you're talking about is just not a thing.

There's all kinds of builds doing exactly this on a budget. What are you talking about?

My writing skills might have failed because I didn't mean to write off all MacOS systems, just the tiny % of MacOS computers called Mac Pro, not to be confused with MacBook Pro.

Integrated System on Chip designs are reducing (not eliminating) the need for PCIe cards. The direction for solving challenges today seems to be more "how can I solve that better within an SoC." So the Mac Pro’s dependence on those cards is becoming more of a limitation - or an investment misdirection - than an advantage.

For now, however much performance you can get from SoC is great, but if you can stick some REAL PCIe lanes on there, it would be infinitely more useful - even if the stuff on the backplane runs slower... we accept that today as the tech is new.

He will never answer that question. I've asked too and he just skips over it as usual.

Admitting that he just likes experimenting with that form factor purely because he finds it neat, runs counter to the identity he's trying to project of being this super interesting tech guy who's doing things that are so Pro that we're too ignorant to comprehend

🤣 Read my thread history - I speak quite freely about a bunch of my use cases. 🤣. And yes, tinkering is one of them. 😂

 

[Bug-Creator]

Sometimes things need to exist even if they're not your biggest seller.

Sure and sometimes a product is just a pointless waste of resources dragging a company down.

I want 10x that in a tower. And give me 5 more in a rackmount.

To do what?

The goal is to have a halo product.

It would be as much of a "halo product" as the TAM or those wheels for the last MPro.
A clear sign that Apple had lost the plot (again).

If they were to invest the billions into such a monstrosity they would at least try to make a use case for it and since there isn't a viable/plausible way to make it into a (commercial) success for whatever that use case that would be it would be seen as a failure.


If you can't do something right it is better not to do it at all.

 

[mcnallym]

Agree 100%, Apple officially pulling the plug on the Mac Pro in March was spectacularly bad timing. They abandoned their top-tier workstation precisely when local LLM inference made massive unified memory a killer feature.

Nvidia’s RTX Spark announcement this week capitalizes on that exact void. By combining a 20-core ARM CPU, a Blackwell GPU, and up to 128GB of memory (all with native CUDA support) Nvidia is bringing Apple-style memory efficiency to Windows. Apple effectively ceded the local pro AI market just as demand exploded. The Mac Studio is capable, but up against equivalent unified-memory ARM PC backed by Nvidia's dominant AI software stack, it's completely outmaneuvered.

Other than the fact that Apple limited this year to the memory that can buy then the Studio and RTX Spark are not far apart.

Both are SoC systems that cannot be expanded internally like a Tower. Clearly Nvidia not felt the need for a Mac Pro like product with it AI workstation.

Under Apple Silicon then Mac Pro is just a Studio with PCIe slots.
Nvidia not making a Mac Pro but a Mac Studio format machine.

The key difference is as you say is the CUDA support.
Apple has never natively supported CUDA and the last Mac supplied with Nvidia GPU was 2014 MBP. Just rechecked as previous AI search said iMac 2013 was the last.
Mojave was the last Mac OS that would take the Nvidia Web Drivers to add support for Nvidia.
CUDA is not important to Apple and the Apple Ecosystem.

So what would a Mac Pro actually bring, if you want CUDA then won’t be looking at Mac OS as your platform in the first place. (Or you shouldn’t be)

It isn’t that the Mac Studio is outmanoeuvred by RTX Sparx, if you want a CUDA AI hardware system, then would not start with a Mac OS system full stop. A Mac Pro in its Apple Silicon generation is not going to offer for CUDA anything that a Studio doesn’t unless expect Apple to somehow bring forth the Extreme Chip which not tried since M1 when failed either due to not being able to produce or could but not at a price people would pay.

Even if you get a Mac Pro M6 Extreme and internal expansion capabilities it still won’t natively support CUDA, so wouldn’t be suitable for CUDA system like the RTX Sparx is.

 

[leifp]

The Mx Ultra only has enough spare PCIe for 1 16-lane GPU card - often lagging a PCIe generation behind. It won't compete with Xeon/AMD CPUs offering 128 lanes (...and probably PCIe6 in the near future).

The laptops and SFF systems making up the vast majority of Mac sales only need PCIe for SSD, USB, ethernet etc.

So, again, the only "design choice" is between the huge expense of designing a whole new die exclusively for the Mac Pro or adding unnecessary complexity (and power consumption) to the laptop-friendly Mx Max die.

With the M5 and move to chiplets this changes the equation. Adding a Northbridge etc or similar addresses the concerns. These are not huge expenses but reimplementations of tried and true tech knowledge once the monolith is split into chiplets already.

It’s a chicken and egg problem. Their choice to take the Mac Pro and shake its users for over a decade played into it, of course, but releasing a Mac Studio with some PCIe slots (that couldn’t GPU) for a lot of money was probably killed by the decisions of non techies (e.g. marketing) at Apple. Instead of half-***ing it they could just have killed the Mac Pro instead of releasing the mediocre-at-best M2 Ultra version.

 

[dmr727]

There's all kinds of builds doing exactly this on a budget. What are you talking about?

Not for AI there's not. If the GPUs can't share the memory, there's no point.

 

[theluggage]

Whatever the unified memory platform can give, I want 10x that in a tower. And give me 5 more in a rackmount. The goal is to have a halo product.

Apple already have plenty of "halo" products, partly thanks to Apple Silicon and its precursors. The iPhone and iPod pretty much popularised the term "halo effect". The M5 MacBook Pro is setting records for laptops in its class. The MacBook Neo is being copied by Asus, Microsoft and others. The NVIDIA RTX Spark that people keep going on about is clearly inspired by Apple Silicon (and is primarily aimed at laptops and SFF so I don't know why people keep raising it in the context of tower systems).

The 2019 and 2023 Mac Pros were more like albatross products!

Outside a few small niches where they may actually meet a need (e.g. hardware-accelerated ProTools for the 2023 MP), the price and spec of the base model 2019 nd 2023 MPs only attracted mockery (even before the $800 wheels). If you want a PCIe tower then more powerful* , with a wide choice of CPUs and GPUs are available and supported by more software. If you want rackmount, proper machine-room-grade systems designed for rackmount are available, in 1/2/3U cases, with lights-out management, hot-swappable drives, redundant PSUs etc. The "rackmount" MP was just a kludge - a tower system turned on its side (with access to those must-have RAM sockets hilariously underneath the machine...)

* ...or sometimes cheaper, less powerful machines, if you just need the expansion. The price of the base Mac Pro was ludicrous for a config no more powerful than a iMac (2019) or Studio (2023) that could take a GPU and internal storage.

Apple have always swiftly exited markets where they no longer had a unique selling point and generic alternatives had become just as good/better, even where they had once been pioneers. Printers, WiFi/NAS, XServe... nostalgia aside, it was far more important to make Macs work smoothly with generic products. Uncomfortable truth - the 2019 MP was nothing special in the high-end tower workstation world. Apple Silicon is something special - but its totally unsuited to PCIe towers.

 

[theluggage]

Even if you get a Mac Pro M6 Extreme and internal expansion capabilities it still won’t natively support CUDA, so wouldn’t be suitable for CUDA system like the RTX Sparx is.

Yup, CUDA is the potential stumbling block - regardless of whether it is in a laptop, SFF or tower.

I'm sure much time will be spent benchmarking RTX Sparx with CUDA against M5 Max/Ultra and Apple's AI/ML frameworks.

One ace in the hole for Apple is that the iPhone/iPad/MacBook Air/Neo represent a too-big-to-ignore market for AI consumption. So third parties do have an incentive to support Apple AI/ML.

 

[mcnallym]

Yup, CUDA is the potential stumbling block - regardless of whether it is in a laptop, SFF or tower.

I'm sure much time will be spent benchmarking RTX Sparx with CUDA against M5 Max/Ultra and Apple's AI/ML frameworks.

One ace in the hole for Apple is that the iPhone/iPad/MacBook Air/Neo represent a too-big-to-ignore market for AI consumption. So third parties do have an incentive to support Apple AI/ML.

For consumption as well then Apple confirmed that used Google’s hardware for training Apple Intelligence so perfectly possible to train the AI that will then be used by Apple End User devices on non Apple Systems, so Apple has no need to produce systems for training to allow the consumption of AI on Apple devices.

 

[theluggage]

it is higher-bandwidth I/O to support more/faster SSD inside a "Pro" system like the higher-tier Studio, and, for (ahem) clustering, like Nvidia is doing with Spark, via the QSFP form factor. The increased I/O wouldn't be for one product, but, for the entire upper tier of Studio systems.

Don't disagree - but in context many people here are quite clearly asking for a new Mac Pro with PCIe slots, specifically for GPUs which (if you're looking for max bandwidth between CPU and GPU) consume 16 PCIe lanes each. I'd expect clustering between integrated SoCs with substantial local unified RAM and storage to do more with less bandwidth.

A Mx Ultra already has enough "spare" PCIe to accommodate several more SSD modules (either Apple flash-only modules or M.2). It has 8 (?) independent 80Gbps thunderbolt controllers for clustering etc. That's not negligible, even if you just take 4 of them for clustering. The problem would be Apple pulling 100 or so lanes of PCIe out of their hat to drive a bunch of x16 GPUs (which also broke unified RAM).

In the video I linked earlier the main complaint about Thunderbolt for clustering seemed to be the lack of switches resulting in a nest of cables since each node had to be wired to every other node. That sounds like a plumbing problem to me (or put switches on the wishlist for USB4v3) rather than a whole new die. QSFP, Inifiniband etc. may offer higher bandwidth, but that would be offset if the main point of them was to run several cluster links over a single cable... Anyway, a Thunderbolt transciever in QSFP format doesn't sound like rocket surgery...

A new high-end Mac could be something like a 1U rack device with one or more Mx Ultra processors internally connected by next-gen Thunderbolt and maybe modular thunderbolt sockets that could accept adapters for various interconnects. and/or maybe they could add some internal storage expansion to the Mac Studio. That's nothing like the sort of monolithic tower Mac Pro that people here are demanding.

 

[FilthyMuppetInnuendo]

There is as new wave of people setting up computers locally (both personal and business) for AI inference.

But we’ve already seen the results of that. The Mac mini and Mac Studio are completely out of stock. That’s what people are buying to run things locally. When Apple saw they weren’t buying the Mac Pro for this purpose, that gave them the excuse needed to discontinue it.

 

[leman]

I've long supported the idea of the Mac Pro as a Halo product. However, more recently I came to believe that Apple already has a very convincing Halo product — and that's the MacBook Pro. You can get a performance of the traditional tower workstation with the portability and battery life of an ultracompact. These are real benefits, in contrasts to the "I want all the PCI lanes just because it sounds fun" attitude some posts here show.

 

[Flint Ironstag]

...

Both are SoC systems that cannot be expanded internally like a Tower. Clearly Nvidia not felt the need for a Mac Pro like product with it AI workstation.

Under Apple Silicon then Mac Pro is just a Studio with PCIe slots.
Nvidia not making a Mac Pro but a Mac Studio format machine....

Nvidia and their partners just announced a 3 tier lineup:

- laptops
- desktops (mini and studio sized boxes)
- workstations (MP sized towers)

Not for AI there's not. If the GPUs can't share the memory, there's no point.

Please tell that to everyone running discrete GPUs with performance that crushes AS.