Like it or not, Apple needs Mac Pro.

[1d1otic]

https://forums.macrumors.com/threads/mac-pro-reportedly-on-back-burner-and-largely-written-off-at-apple.2471279/

Despite that Mac Pro is being on the back burner, Apple will need to develop high-end and/or workstation grade Apple Silicon chips for their own future. I would skip the performance and specs of Macs cause it's gonna be a long conversation despite the fact that Mac's max performance is poor but the most important fact is, they need workstation grade chips to make their own servers for AI.

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. M3 Ultra? It's extremely slow compared to others especially Nvidia. Furthermore, since Apple Silicon is SoC based, it's too expensive to mass produce especially with ultra-fusion due to the die size. Dont forget that Apple does NOT make chips to sell so they'll need to use chips from Mac, not custom chips just for servers.

Tho Apple is planning to make new chips with a whole new design thanks to TSMC's SoIC, they will need Mac Pro grade Apple Silicon chips one way or another cause in order to compete and improve their Apple Intelligence.

So what Apple is doing with Mac Pro is totally unacceptable.

 

[leman]

Just a few comments on this:

- What makes you think that Apple wants or needs to train models on their own hardware (inference is a different matter)?
- M3 Ultra is slow for ML because it still lacks ML acceleration. M5 Ultra might be less slow.
- The latest macOS beta introduces Infiniband support, which is the protocol Nvidia uses to build large AI clusters. This would allow you to connect multiple Studios together and use them as a distributed ML accelerator at a lower price than an equivalent hypothetical Mac Pro. And it is very possible that this is what Apple will use internally to link multiple Max or Ultra class chips into coherent compute clusters.

 

[!!!]

The Mac Studio is essentially the replacement for the Mac Pro, there's no difference aside from upgradability/extendability.

The problem with Apple's chips in terms of AI is that training requires high-precision GPU-like cores, which apple struggles at (and NVIDIA does not). Apple does have the NE-cores, but they are geared towards running the model, and thus have worse precision and don't have all of the necessary functions for training.

I don't see how the Mac Pro would help, unless Apple gives in and allows NVIDIA/AMD GPUs, or Apple goes on to make their own dedicated GPUs. The SOC model just does not work when you need high performance GPUs, you'll run into thermal limits.

- The latest macOS beta introduces Infiniband support, which is the protocol Nvidia uses to build large AI clusters. This would allow you to connect multiple Studios together and use them as a distributed ML accelerator at a lower price than an equivalent hypothetical Mac Pro. And it is very possible that this is what Apple will use internally to link multiple Max or Ultra class chips into coherent compute clusters.

If Apple were to go as far as designing their own AI compute cluster, using full M-series processors would be a waste, you don't need so many CPU cores. It might be cost-effective to use binned chips with good GPU cores and bad CPU cores, but even then they may as well design a completely custom board/system that can integrate multiple chips in a better way than separate computers. Take a look at NVIDIA's DGX systems, they're essentially a bunch of GPUs with tons of VRAM all plugged into a single powerful CPU.

Basically, Apple could be far more optimized than just plugging in Mac Studios together. In fact, I believe it's probably more profitable for them to sell every M-chip to consumers and just rent out some of Google's TPUs to train with, which is what they're currently doing.

 

[leman]

The problem with Apple's chips in terms of AI is that training requires high-precision GPU-like cores, which apple struggles at (and NVIDIA does not). Apple does have the NE-cores, but they are geared towards running the model, and thus have worse precision and don't have all of the necessary functions for training.

It’s a matter of priorities, really. Apples new GPU matrix accelerators for most parts have performance parity with Nvidia on 16-bit floating point data (sans sparsity). Nvidia has much higher compute on lower-precision formats, and their professional GPUs have additional FP32 matrix acceleration. Still, nothing stops Apple from building a larger matrix accelerator, they have the tech. Just not the business need.

If Apple were to go as far as designing their own AI compute cluster, using full M-series processors would be a waste, you don't need so many CPU cores. It might be cost-effective to use binned chips with good GPU cores and bad CPU cores, but even then they may as well design a completely custom board/system that can integrate multiple chips in a better way than separate computers. Take a look at NVIDIA's DGX systems, they're essentially a bunch of GPUs with tons of VRAM all plugged into a single powerful CPU.

They could also make custom chips consisting mostly of GPU cores or AMX cores or scaled-up ANE…

Basically, Apple could be far more optimized than just plugging in Mac Studios together. In fact, I believe it's probably more profitable for them to sell every M-chip to consumers and just rent out some of Google's TPUs to train with, which is what they're currently doing.

For training, yes. But they also have growing demands for private server-side LLM execution. Can’t really outsource that.

Plugging in Mac studios together is for the user (if you want to run a large LLM locally or build a render farm). Server-side tech will obviously look very different. What’s relevant is that they are building the software foundation to do both.

 

[!!!]

They could also make custom chips consisting mostly of GPU cores or AMX cores or scaled-up ANE…

True, and I hope they do, as someone who would rather game on a mac than have to deal with windows. And I did state:

Apple goes on to make their own dedicated GPUs

It's debatable as to if they'd ever go for a separate card, my guess is that they will at least move to a chiplet design consisting of the typical M-series chip with additional GPU-only chiplets added on.

For training, yes. But they also have growing demands for private server-side LLM execution. Can’t really outsource that.

Plugging in Mac studios together is for the user (if you want to run a large LLM locally or build a render farm). Server-side tech will obviously look very different. What’s relevant is that they are building the software foundation to do both.

Sure, it's more profitable to force end-users to buy entire machines than it is to sell a cost-effective solution.

But I was only referring to Apple's use case, as it appeared to be what @1d1otic was talking about:

they need workstation grade chips to make their own servers for AI.

 

[leman]

True, and I hope they do, as someone who would rather game on a mac than have to deal with windows. And I did state:

It's debatable as to if they'd ever go for a separate card, my guess is that they will at least move to a chiplet design consisting of the typical M-series chip with additional GPU-only chiplets added on.

I was under impression that we are discussing high-end ML applications? How would that relate to gaming? You wouldn’t game on a H200 either, right?

We all want faster GPUs, that for sure, and there are different ways to make that happen. M5 is considerably faster than M4 for example without increasing the GPU die area or power consumption much.

Sure, it's more profitable to force end-users to buy entire machines than it is to sell a cost-effective solution.

I don’t quite follow. What would a cost-effective solution be? We are talking about the ability to distribute demanding compute workflows across multiple machines. It’s not always feasible or cost-effective to build a larger single machine. Cluster interconnect technology like Infiniband exists for a reason - and now it also comes to Apple platforms.

 

[1d1otic]

- What makes you think that Apple wants or needs to train models on their own hardware (inference is a different matter)?

Ai is the future and a lot of companies are training their OWN AI and models. Besides, since Apple already made Apple Intelligence, it's inevitable that Apple needs to create their own. Dont forget that Apple is also researching toward AI itself.

- M3 Ultra is slow for ML because it still lacks ML acceleration. M5 Ultra might be less slow.

It's more important to have dedicated chips such as NPU.

 

[1d1otic]

The Mac Studio is essentially the replacement for the Mac Pro, there's no difference aside from upgradability/extendability.

The problem with Apple's chips in terms of AI is that training requires high-precision GPU-like cores, which apple struggles at (and NVIDIA does not). Apple does have the NE-cores, but they are geared towards running the model, and thus have worse precision and don't have all of the necessary functions for training.

I don't see how the Mac Pro would help, unless Apple gives in and allows NVIDIA/AMD GPUs, or Apple goes on to make their own dedicated GPUs. The SOC model just does not work when you need high performance GPUs, you'll run into thermal limits.

Mac Studio or M3 Ultra is extremely slow compared to others.

It's not even close to high-end desktop based on its performance. Mac Pro on the other hand, it used to have 90 series grade GPU with up to 4 slots. Both specs and performance are not even close to Mac Pro or workstation. Dont forget that Apple made their own servers with Mac Pro parts before.

 

[mcnallym]

Mac Studio or M3 Ultra is extremely slow compared to others.

It's not even close to high-end desktop based on its performance. Mac Pro on the other hand, it used to have 90 series grade GPU with up to 4 slots. Both specs and performance are not even close to Mac Pro or workstation. Dont forget that Apple made their own servers with Mac Pro parts before.

Apple used to make their own servers with the xserve range, They also used to do own printers and Xsan as well. They made Time Capsule and Airport Wireless products.

All of those products are gone. Apple predominantly runs Linux on generic hardware now for its own server use. They moved to selling Mac OS server on Mac Pro and even Mini as not cost effective to continue with Xserve. Instead focussed more on getting Mac to work with Active Directory and Linux environments.

Just because Apple used to do something in the past doesn’t mean they have to do it again now. They used to do Mac Clones and look how well that turned out.

Apple clearly don’t seem put out by the fact that don’t offer HP Z8 class machines.

Apple do Cloud but they are not trying to compete with Azure and AWS.

I don’t see Apple Intelligence trying to compete with AI like ChatGPT or OpenAI instead is more for user centric tasks and making those tasks easier or even done for the user as opposed a more general AI.

Where apple bringing in the hardware is more around this then AI model learning from what I see.

If going to start doing hardware for model learning, training of models then a Mac Pro would not be where want to start with but a specialised kit for the purpose.

 

[leman]

Ai is the future and a lot of companies are training their OWN AI and models. Besides, since Apple already made Apple Intelligence, it's inevitable that Apple needs to create their own. Dont forget that Apple is also researching toward AI itself.

You don’t need your own hardware to train a cutting-edge model. Most market leaders use Nvidia, and there is nothing wrong with that.

Besides, Apple pretty much gave up training their own state of the LLM for the current cycle - they are buying Gemini weights from Google.

It's more important to have dedicated chips such as NPU.

Apple’s NPU tech is state of the art and leading in its class already. But you need different solutions for different types of problems. Which is why Apple currently offers three hardware ML accelerators within Apple Silicon, all optimized for different use cases.

 

[Abydos]

with AI as an issue, where is the unique revenue stream for apple (ok marketing and steering users toward a specific solutions)? I'm not entirely sure being an 'early adapter' or 'me too' corporation rather using open ai or google ai as a tool to explore where apple might progress (think profit). but then again I could be totally wrong

 

[1d1otic]

Apple used to make their own servers with the xserve range, They also used to do own printers and Xsan as well. They made Time Capsule and Airport Wireless products.

All of those products are gone. Apple predominantly runs Linux on generic hardware now for its own server use. They moved to selling Mac OS server on Mac Pro and even Mini as not cost effective to continue with Xserve. Instead focussed more on getting Mac to work with Active Directory and Linux environments.

Just because Apple used to do something in the past doesn’t mean they have to do it again now. They used to do Mac Clones and look how well that turned out.

Apple clearly don’t seem put out by the fact that don’t offer HP Z8 class machines.

Apple do Cloud but they are not trying to compete with Azure and AWS.

I don’t see Apple Intelligence trying to compete with AI like ChatGPT or OpenAI instead is more for user centric tasks and making those tasks easier or even done for the user as opposed a more general AI.

Where apple bringing in the hardware is more around this then AI model learning from what I see.

If going to start doing hardware for model learning, training of models then a Mac Pro would not be where want to start with but a specialised kit for the purpose.

You said it to yourself. To make a specialized kit, you'll need hardware but now, Apple cant make it cause a middle range PC is all they can make. Do you see the problem now?

 

[1d1otic]

You don’t need your own hardware to train a cutting-edge model. Most market leaders use Nvidia, and there is nothing wrong with that.

Besides, Apple pretty much gave up training their own state of the LLM for the current cycle - they are buying Gemini weights from Google.

Not too late to start their own and they already have their own AI servers made with M2 Ultras.

Apple’s NPU tech is state of the art and leading in its class already. But you need different solutions for different types of problems. Which is why Apple currently offers three hardware ML accelerators within Apple Silicon, all optimized for different use cases.

Not meant for AI service itself.

 

[jakey rolling]

You are living in the very distant past. The Mac Pro has been a neglected stepchild in Apple's product line-up since 2013. The intel-based Mac Pros sold between the intro of the trashcan and the intro of the M-series were relatively weak as far as workstation-class machines go (built with the lowest-end intel Xeon WS CPUs, for example*). They at least offered some expandability and useful options for the few users who could make proper use of them, but for most Mac users (even high-end pros), the Post-2013 Mac Pros offered little benefit over other Mac alternatives. The current Mac Pro is little more than an older-generation Mac Studio tarted up in a bigger case with a few PCIe slots (and almost nothing to actually plug into them). I honestly can't think of a single good reason why anyone would opt for a Mac Pro over a Mac Studio outside of wanting a nice looking case that sits on the floor.

Mac Studio or M3 Ultra is extremely slow compared to others.

It's not even close to high-end desktop based on its performance. Mac Pro on the other hand, it used to have 90 series grade GPU with up to 4 slots.

The Post-Trashcan Mac Pro had the ability to use the Afterburner card which is basically now rolled into the M-Series processors. Even when they were using "90 Series" GPUs, those were the so-called "pro" cards that were actually slower than "consumer" cards of the time, but were marketed as "pro" cards due to both NVIDIA and AMD using artificial software and firmware locks to prevent "consumer" cards from being used on certain professional workloads.

Dont forget that Apple made their own servers with Mac Pro parts before.

Yes. And they sucked.

The Xserve was the closest thing Apple ever came to building an enterprise-class server and it was terrible. Apple cheaped out in some areas where you definitely should never cheap out on a server (such as the hard drive controller), and even their "Enterprise Support" for the thing was laughably bad. Mac OS X Server was a decent NAS OS but little more than that. The only two Apple-branded enterprise server products that really worked well were Xsan and WebObjects, both products acquired by Apple with some fanfare and then more or less left to slowly rot on the vine (WebObjects, by the way, not only powered the iTunes store for many years, but also powered Dell's online store back when Dell was the top-cheese of the Windows PC market).

*I'll take a quick sidebar on this one as well. Apple used Xeon CPUs in the Mac Pro from the beginning, but the way they did so and marketed the machines made it very clear that the only reason they did so was to add a dubious selling point. Xeon processors and ECC RAM offered no advantage to anyone running workloads outside of high-precision scientific or engineering realms, such as modelling or CAD applications. Those buying Mac Pros for audio or video production, or for photo editing or desktop publishing (or basically anything else), would have been better served with a higher-end Core i7/i9 based system with faster RAM and a better GPU. Of course, putting a "consumer-grade" CPU into the Mac Pro would mean Apple could no longer point to the $10k workstations from HP and Dell to justify the insane pricing of the thing. The biggest problem with the Intel Mac Pro was that it was built like a Kenworth when most of its target market just needed a Ford F-350.

 

[!!!]

I was under impression that we are discussing high-end ML applications? How would that relate to gaming? You wouldn’t game on a H200 either, right?

I know, but high-end GPUs often translate to both gaming and ML. If Apple builds better GPUs for the benefit of ML, then we may also see improved game performance.

I wouldn't game on an H200 in the same way I wouldn't buy a Mac Studio instead of an H200.

I don’t quite follow. What would a cost-effective solution be? We are talking about the ability to distribute demanding compute workflows across multiple machines.

I'm not sure how else to put this, when you buy a Mac Studio, you are paying for the CPU, media engine, SSD, ports, case, apple tax, etc. It's an entire computer. When you buy a dozen NVIDIA GPUs and stick them into a big workstation, you are only paying for each GPU. In that same light, theoretically, Apple offering some sort of compute node with ML-optimized chips would be a far more cost effective option, rather than buying an entire computer where most of it is unneeded. Anyways, Apple making something like that isn't going to happen, so you're left with paying Apple's tax.

Mac Studio or M3 Ultra is extremely slow compared to others.

It's not even close to high-end desktop based on its performance. Mac Pro on the other hand, it used to have 90 series grade GPU with up to 4 slots. Both specs and performance are not even close to Mac Pro or workstation. Dont forget that Apple made their own servers with Mac Pro parts before.

I think what you're trying to say is that the GPUs in Apple's M-series is underpowered. I agree. But that doesn't mean that the studio is underpowered compared to the current Mac Pro, because they both share the same chips. And since Apple is not interested in GPU expandability, any Mac Pro is essentially useless compared to the Studio.

Apple's long-ago history of NVIDIA and AMD GPUs and servers and all that is long gone. Apple is far more interested in the consumer market, including the so-called pro-sumer market. Why sell a Mac Pro for $2000 and let users buy their own GPUs when you can sell the same user Mac Studios for $5k that don't approach the performance of any NVIDIA chip? Apple makes more money, investors are happy, Apple can continue to sell $1500 iPhones, the end.

 

[cjsuk]

Mac Pro is only good if you need a crap load of PCI slots. Otherwise it's functionally useless.

As for LLM workloads, it's a statistically insignificant number of people give a crap or have enough money to give a crap.

If you want that, go buy a PC.

 

[senttoschool]

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

I was quite early with my post in 2021. I was right that it made no financial sense for Apple to make an "Extreme" SoC without another market. Heck, Apple barely updates the Ultra since it's so niche.

I felt like the only way Apple could justify the expense and engineering for an "Extreme" chip was if they use it in the cloud as well.

The most likely scenario is that Apple develops an inference chip for themselves that is not 2x Ultras glued together. It doesn't really make much sense to have useless silicon like many P/E cores, display controllers, etc. in an inference chip. Transistors are getting very expensive. It'll be optimized purely for cloud inference. I'm guessing a giant NPU.

Next, I could actually see Apple make a 2x Ultra chip if local LLMs blow up and become must haves. In that scenario, there might be a market big enough to justify the cost.

 

[1d1otic]

I think what you're trying to say is that the GPUs in Apple's M-series is underpowered. I agree. But that doesn't mean that the studio is underpowered compared to the current Mac Pro, because they both share the same chips. And since Apple is not interested in GPU expandability, any Mac Pro is essentially useless compared to the Studio.

Apple's long-ago history of NVIDIA and AMD GPUs and servers and all that is long gone. Apple is far more interested in the consumer market, including the so-called pro-sumer market. Why sell a Mac Pro for $2000 and let users buy their own GPUs when you can sell the same user Mac Studios for $5k that don't approach the performance of any NVIDIA chip? Apple makes more money, investors are happy, Apple can continue to sell $1500 iPhones, the end.

Why do you compare with a current Mac Pro which is a crap? I'm not taking about upgradability, I'm talking about aka Extreme chips which is a lot powerful than Ultra. Using a same chip is just a stupid choice.

 

[1d1otic]

You are living in the very distant past. The Mac Pro has been a neglected stepchild in Apple's product line-up since 2013. The intel-based Mac Pros sold between the intro of the trashcan and the intro of the M-series were relatively weak as far as workstation-class machines go (built with the lowest-end intel Xeon WS CPUs, for example*). They at least offered some expandability and useful options for the few users who could make proper use of them, but for most Mac users (even high-end pros), the Post-2013 Mac Pros offered little benefit over other Mac alternatives. The current Mac Pro is little more than an older-generation Mac Studio tarted up in a bigger case with a few PCIe slots (and almost nothing to actually plug into them). I honestly can't think of a single good reason why anyone would opt for a Mac Pro over a Mac Studio outside of wanting a nice looking case that sits on the floor.


The Post-Trashcan Mac Pro had the ability to use the Afterburner card which is basically now rolled into the M-Series processors. Even when they were using "90 Series" GPUs, those were the so-called "pro" cards that were actually slower than "consumer" cards of the time, but were marketed as "pro" cards due to both NVIDIA and AMD using artificial software and firmware locks to prevent "consumer" cards from being used on certain professional workloads.


Yes. And they sucked.

The Xserve was the closest thing Apple ever came to building an enterprise-class server and it was terrible. Apple cheaped out in some areas where you definitely should never cheap out on a server (such as the hard drive controller), and even their "Enterprise Support" for the thing was laughably bad. Mac OS X Server was a decent NAS OS but little more than that. The only two Apple-branded enterprise server products that really worked well were Xsan and WebObjects, both products acquired by Apple with some fanfare and then more or less left to slowly rot on the vine (WebObjects, by the way, not only powered the iTunes store for many years, but also powered Dell's online store back when Dell was the top-cheese of the Windows PC market).

*I'll take a quick sidebar on this one as well. Apple used Xeon CPUs in the Mac Pro from the beginning, but the way they did so and marketed the machines made it very clear that the only reason they did so was to add a dubious selling point. Xeon processors and ECC RAM offered no advantage to anyone running workloads outside of high-precision scientific or engineering realms, such as modelling or CAD applications. Those buying Mac Pros for audio or video production, or for photo editing or desktop publishing (or basically anything else), would have been better served with a higher-end Core i7/i9 based system with faster RAM and a better GPU. Of course, putting a "consumer-grade" CPU into the Mac Pro would mean Apple could no longer point to the $10k workstations from HP and Dell to justify the insane pricing of the thing. The biggest problem with the Intel Mac Pro was that it was built like a Kenworth when most of its target market just needed a Ford F-350.

Why do you justify degeneration of Mac specs for? Apple used to have a workstation for professional uses but now, Mac only offers up to middle range specs which is a joke.

 

[iPadified]

Why do all companies need to have the full hardware/software stack? An excessive waste of resources. A Z8 is an excellent machine and there are plenty other. Also, I suspect NVIDIA is getting rich by providing for the data centers, not the workstation market.

Anyone calling themself a "pro" would also be OS independent and choose the optimal hardware and software combination for a job. Apple has never ever been perfomance leader. Performance per watt (since a few years) but not raw perfromance.

Apple is doing what Apple is doing best: providing an appealing package (although a little on the expensive side) for the end user of technology.

 

[leman]

Why do you justify degeneration of Mac specs for? Apple used to have a workstation for professional uses but now, Mac only offers up to middle range specs which is a joke.

This sounds like an overly dramatic take to me. The upcoming M5 Max Studio should match or outperform the Nvidia Spark for FP16/BF16 processing, while offering a much better CPU, overall better specs, and much better versatility, and that at a similar price. The market is not the same as what it was ten years ago.

 

[1d1otic]

This sounds like an overly dramatic take to me. The upcoming M5 Max Studio should match or outperform the Nvidia Spark for FP16/BF16 processing, while offering a much better CPU, overall better specs, and much better versatility, and that at a similar price. The market is not the same as what it was ten years ago.

Doesn't change the fact M5 Ultra wont gonna replace 4x GPU that workstations can have or even Extreme chip. You are comparing a single desktop with one GPU to workstation with four GPU.

 

[1d1otic]

Why do all companies need to have the full hardware/software stack? An excessive waste of resources. A Z8 is an excellent machine and there are plenty other. Also, I suspect NVIDIA is getting rich by providing for the data centers, not the workstation market.

Anyone calling themself a "pro" would also be OS independent and choose the optimal hardware and software combination for a job. Apple has never ever been perfomance leader. Performance per watt (since a few years) but not raw perfromance.

Apple is doing what Apple is doing best: providing an appealing package (although a little on the expensive side) for the end user of technology.

And yet, they degraded their own Mac markets for a while and lose many markets such as 3D graphics.

 

[leman]

Doesn't change the fact M5 Ultra wont gonna replace 4x GPU that workstations can have or even Extreme chip. You are comparing a single desktop with one GPU to workstation with four GPU.

You are correct, it won’t. Then again, does it have to? How large is the market for the Blackwell 6000 Pro class GPUs and how many folks are running four of them in a single box?

I’d say they Apple can serve a significant portion of a market with a Studio that is somewhere around the 5000 Blackwell Pro levels, and it looks like they are on a good way. And Infinibsnd over TB gives options for those folks who need more compute, while remaining cost-effective compared to the market.

 

[1d1otic]

You are correct, it won’t. Then again, does it have to? How large is the market for the Blackwell 6000 Pro class GPUs and how many folks are running four of them in a single box?

I’d say they Apple can serve a significant portion of a market with a Studio that is somewhere around the 5000 Blackwell Pro levels, and it looks like they are on a good way. And Infinibsnd over TB gives options for those folks who need more compute, while remaining cost-effective compared to the market.

The market for Blackwell 6000 Pro or something similar is a LOT bigger. Dont forget that servers and workstations are connected each other with similar technology and yet, Apple is killing their own market.

Besides, Apple's GPU performance isn't great so what's the problem with that? Apple dont even have RTX 5090 grade GPU and workstations.