Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
The engineers who worked on PowerPC are long retired.
I wasn’t commenting on how hard it is, just that they don’t have any expertise (or apparent desire) to do it.

They don't have the desire indeed, because it comes with various downsides. I see no reason why they wouldn't have the expertise. They can design their own CPU, various satellite chips, a package, logic board, etc., but they can't make a CPU that isn't in an SoC?

Even their messaging is heavily SOC focussed. How could they row back from memory and storage speed and latency figures that aren’t possible with bus architectures?

Well, storage speed wouldn't change that much. With enough PCIe lanes, you can reach 14 GB/s. Latency would be worse on paper, but I'm unsure this would have much effect in reality. It might be more of a security concern, however. They do like the controller to be on the SoC.

Memory, indeed. There, the main argument would be much higher capacity — but at the cost of being slower.

 
You're close to "getting it" with regards to what apple is doing now with the studio.

Don't think in terms of RAM slots or whatever, just think in terms of small/med/large lego bricks.

The studio is a lego brick
Pick the size of brick, pick the quantity and link them together.

We've been doing this with enterprise VM workloads for decades. I.e., you're not working with blade arrays or whatever. Just using lego bricks of self-contained compute/ram/storage in one module.

Depending on your workload the type of bricks you use may change, once you figure out your workload requirements, just scale with more of the same modules that contain cpu/gpu/ram/storage in the same brick.

In server land we call this hyper convergence.


This is where I see the "Mac Pro" going... being replaced by an array of studios or similar. I thought they should do it with a mainboard and blades as well (basically Mac Studio on a blade), but this way they can sell the studios individually and not change anything if thunderbolt is fast enough.

Are there compromises vs. a backplane with a bunch of blades? Sure. But its simpler and probably "good enough" to work and then speeding it up is just a question of making the thunderbolt ports faster with later generations.

The computing landscape is filled with previous bespoke solutions to problems that died because "simple and good enough" won.

multiple studios is just a network, a MAC PRO is 1 powerful system that combines RAM, STORAGE, and GPU power. Am I missing something? or can you connect mac studios to work together as one computer?
 
  • Like
Reactions: peter2
multiple studios is just a network, a MAC PRO is 1 powerful system that combines RAM, STORAGE, and GPU power. Am I missing something? or can you connect mac studios to work together as one computer?
RDMA over thunderbolt recently got enabled, so yeah for high end workloads you can cluster them fairly effectively up to say 4 nodes with 2TB combined ram and 4x the spec of a single ultra in cpu/gpu cores to work with, getting somewhere close to 3-4x the performance of a single box.

Its not quite the same as one computer but for extremely high end workloads that take a long time to run, clustering is pretty effective.

That can be achieved for say 40-60k USD.

For context that's around the cost of 1x H100 nvidia GPU without the system to put it in and provides an order of magnitude more VRAM for large data-sets (like extremely large LLMs). To get something with 1-2 TB of VRAM any other way you're talking multiple 30-50k GPUs (with 80GB or so each - so you need like 12 of them to get to 1TB VRAM), plus the servers to put them in, kilowatts of power, cooling, networking, server rack(s), switching, etc. so.... you do the math...

The performance will be a lot higher, sure - but you're talking hundreds of thousands of dollars / millions of dollars vs. under 100k. The jump in cost from clustered mac studios right now to anything better for high end GPU compute is HUGE.

If your workload is bigger than that, you're better off renting the compute or spending 10-20x the money (or more) on building your own PC hardware based clusters.
 
Last edited:
They don't have the desire indeed, because it comes with various downsides. I see no reason why they wouldn't have the expertise. They can design their own CPU, various satellite chips, a package, logic board, etc., but they can't make a CPU that isn't in an SoC?



Well, storage speed wouldn't change that much. With enough PCIe lanes, you can reach 14 GB/s. Latency would be worse on paper, but I'm unsure this would have much effect in reality. It might be more of a security concern, however. They do like the controller to be on the SoC.

Memory, indeed. There, the main argument would be much higher capacity — but at the cost of being slower.

A socketed separate CPU does not make sense for Apple. The current M chips are already very difficult to manufacture and the tight integration with the RAM next to the CPU makes sense for them.

They're not about to spend millions or billions of dollars to manufacture a unique SKU for a CPU just for the Mac Pro. Mac Pro's aren't sold as much as their other Macs. It doesn't make sense for them. And they are never ever going to hand off the RAM and GPU as separate modules.
 
Re: just a network....

With RDMA its sort of halfway between being a bunch of Studios on a network and having a bunch of studios with a shared memory bus. Halfway is maybe selling it short too... way further from network and closer to talking to another CPU socket in a multi-socket PC.

There's far less overhead than a regular network - its not as fast as local memory but each machine can access the other machine's memory in a far more direct manner than going over the network the normal way.

SO its faster, not just because its 80/120 gigabit - but because there's no overhead of encapsulating data into IP, then into ethernet frames, etc. and then decoding it via cpu on the other end (all of which takes time! and additional data which reduces performance),

It's much more lightweight - there's not a lot of documentation on it yet, i'm sure an infiniband expert can elaborate more, but its closer to the CPU talking directly to the other machine over PCIe (e.g., like your CPU talking to your local GPU) than sharing data over a traditional ethernet/ip network.
 
Last edited:
I don't think it has anything to do with expertise because the 2019 MP is an engineering marvel and that's a "recent" machine. .
They engineered a nice case to store the third party chipsets in, sure. They didn’t engineer the actual computer though
 
  • Like
Reactions: throAU
They engineered a nice case to store the third party chipsets in, sure. They didn’t engineer the actual computer though
You sure about that?
Take a closer look at the logic board. It has Thunderbolt routed to the two MPX slots. In addition to extra power delivery, with no need for additional power cables.
The proprietary IO card has two Thunderbolt ports.
There are also mini 8pin power connectors, for standard PCIe graphic cards.
The PSU connects directly to the logic board. Again without any cabling.
 
Last edited:
I wasn’t commenting on how hard it is, just that they don’t have any expertise (or apparent desire) to do it. Even their messaging is heavily SOC focussed.
I think it's down to "desire" rather than anything else. Making a more traditional multi-chip chipset is almost certainly simpler than a SoC, I don't think Apple Silicon designers would strain their brains making a discrete chipset. If not, expertise can be purchased and Apple have plenty of cash. But it wouldn't be Apple Silicon - being a SoC, having unified RAM shared by an on-die GPU/NPU/Media Engine is the Unique Selling Point of Apple Silicon.

Even the later Intel Macs were concentrating more and more functionality into the T1/T2 chips which was a stepping stone to going full SoC.

so where’s the motivation to create a more expensive system for a small niche where Apple will be the underdog anyway?
Exactly. You can even get ARM-based server/workstation systems using CPUs from Ampere etc. (Not so much in the consumer market yet, apart from low-powered "maker boards" like the Pi).... and if you put third-party GPUs/TPUs in a workstation/server, that's largely what will determine it's performance, not whether it's running MacOS, Windows or Linux.

What a non-SoC Mac Pro wouldn't do is offer trickle-down benefits to the SoC-based Studios and MacBook Pros, because it wouldn't add any incentive for developers to optimise software for Apple Silicon GPUs/NPUs/Media engines.

(Clearly some firewall between the Mac Pro team and the M1 team. The 2019 Mac Pro design makes some strange choices when the ARM transition is announced mere months later.)

The Mac Pro design process presumably started in 2017 after the famous press conference where they admitted that the Trashcan wasn't wonderful. My guess is that the 2017 iMac Pro was going to be the new Mac Pro but when they showed the plans to key customers/developers it was laughed off-stage. So they were committed to releasing a Mac PCIe tower sometime after 2017. Even if they knew Apple Silicon was coming, the priority for that was SoCs for laptops where the performance-per-battery-hour would be most advantageous, not PCIe towers for which it was really not suited. The M1 Ultra was the first chip that could have done the job (badly) and that didn't arrive until 2022. So the 2019 Mac Pro did have a gap to fill - and it was probably doomed to be a one-and-done. The 2023 MP has a niche, but it isn't really a comparable system - Apple Silicon just isn't the right tool for that job.

Also, the 2019 Mac Pro was very dependent on the (then) new generation of Xeon-W chips with much increased RAM and PCIe bandwidth and Intel aren't known for delivering chips, in quantity, on time. Who knows how delayed it was? Being an "early adopter" of those chips was the only thing that gave the Mac Pro impressive specs compared to bog-standard Xeon-W systems.

One thing that I do wonder about with the 2019 MP is that it's other Unique Selling Point - the MPX slots distributing power and Thunderbolt to PCIe cards - looks ideal for making a system that could be expanded with multiple Apple Silicon "compute modules" on MPX modules, running RDMA or something over PCIe. That might have been workable even with an Intel mainboard. Plus, if the Mx "Extreme" had ever materialised it would presumably haven had twice the PCIe lanes & twice the RAM of an Ultra & might have been a contender for a Mac Pro mainboard. So maybe there were grander plans for Apple Silicon Mac Pros that fell by the wayside.
 
They engineered a nice case to store the third party chipsets in, sure. They didn’t engineer the actual computer though

You don't own one so you don't get to use it every day for years. It's an amazing piece of technical achievement (compared to what's out there in this same category)
 
Why on earth would I want to own one? All my big toys have been paid for by others!
 
You sure about that?
Take a closer look at the logic board. It has Thunderbolt routed to the two MPX slots. In addition to extra power delivery, with no need for additional power cables.

It has PCI-e and Display Port routed. to (from in latter case) the MPX slots. The Thunderbolt is all provisioned from the edges of MP. Top/front daughter board, the I/O card, and the MPX card itself. but yes, it does it without free-flowing wires.

So yes there is a substantive DIsplay Port switch and PCI-e switches there. But mostly off-the-shelf stuff. Needs to be logic board routed, but the chips themselves are bought. (including Thunderbolt which came from Intel/TI/etc. )


The proprietary IO card has two Thunderbolt ports.
There are also mini 8pin power connectors, for standard PCIe graphic cards.
The PSU connects directly to the logic board. Again without any cabling.

There were functional standard connectors but Apple skipped them so they designed the 'computer' ( although those connectors don't directly compute) ? A modern refrigerator has a logic board that needs to be laid out also ... is that designing a 'computer'? Yes, there is design work there, but a 'computer' in the personal computer sense of the word not as much.

Apple added the T2 chipset. That is a compute element that was not 3rd party (off the shelf). That would be a better example.
 
  • Like
Reactions: wyliej
Nobody makes apple only PCI cards, so forcing a form factor that makes them unique would be a non-starter, the external chassis as cringy as it is would be the way to go. Of course TB5 suffers from the same connector issue that USB-C has and when it is your system bus essentially there has to be a locking mechanism to hold it in place. But that's a $0.10 clip to Make sure that the chassis and CPU can't disconnect from each other.. Then you can slap a M2 drive card to have massive on-bus storage, a GPU card, your RED card, etc... But not sure if enough people will do it to justify NVIDIA or AMD to produce drivers for their GPUs (if even for just AI compute rather than actual video) since MacOS doesn't natively ship with NVIDIA drivers (and getting PyTorch to play nice with the CUDA drivers is hard enough on Ubuntu let alone on a system that they don't support natively)
From Intel to Silicon - a different way of handling main processing and video.
Use of USB C / TB meant for standardized interface with external devices and the same with PCI for internal an industry standard. Apple doesn't have to play with those standards and ISOs, it can elect to interface in another fashion. Consider - ISA bus gave way to VLB Vesa Local Bus, that also gave way to PCI and its incarnations. If some "module" is going to exist internally, Apple can design what it wants as long as it interfaces properly. There is nothing new about this notion just as on board memory and storage and video has been done in different ways. As for NVIDIA and AMD, they may or may not be players initially. I am not discounting what you are saying if Apple were to embrace historic means to an end but they don't have to and this was shown with the jump to silicon and the previous jump to Intel. In short, Apple can elect to define its own internal interface and specs and not reflect industry standards as they come and go.
 
Not sure why Apple has shown deliberate neglect for the Pro,

largely that is a 'two way street' . the Mac Pro forums/thread regular have folks chime on on how they bought a 2-3 year old Mac Pro and made it work just fine with some commodity parts replacements. They don't buy new Mac Pros. That pragmatically makes them non Apple customers (they don't hand money to Apple). There are a few that might buy direct refurbished, but not in large numbers.

Likewise even the direct buyers have a fairly substantial subset of only buying every 6-10 years. So Apple is selling in those intermediate years to who? Say collective pool of Mac Pro users is 800K and they are spread evebkt out over that 8-ish year timespan. That only gives them 100K per year buying new.

Supposedly the Apple Vision Pro is a 'giant flop' because it only sold about 300-400K a year. Three (or four) times that number.

Longer cycle buys of new equipment lead to longer cycle delivery of new systems.

like at least throw processor bumps in, a Mac Pro with M5 Ultra would be a bone at least to creatives to use,

Apple SoCs typically have 'hand me down' products to continue to sell them into. When the MBP 14"/16" dumped the M3 Max then Studio got the M3 Utlra (***) . the M2 Ultra got dumped relatively quickly from the Mac Studio... where does that get the 'hand me down' to? Pragmatically the Mac Pro.

' Folks hand wave that the iPhone churning every year and not fully pay attention to that chip NOT being dumped every year. The iPhone continues to be sold for at least two more years. Apple then passes it along to other products like entry iPad or AppleTV.

Similarly now that Apple has gone to making different sized dies for. Axx and Axx Pro the Pro chip is now going to fall into low costs Macs after spends just a single year in iPhone Pro model (that 'dies' every year).


The Ultra class chips have far more higher upfront R&D costs and yet Apple is going to toss them in the garabage can every 12 months. Probably not. The higher R&D costs (coupled to relatively low volumes ) is much more likely going to lead to longer lifecycles ; not shorter ones. Hence Apple dropping hints that "Ultra" many not show up every generation. There is large economics behind that. Intel, AMD, Nvidia don't dump large chips in the trash can every 12 months. There might be a new hype train chip every 12 months but the replacement isn't dumped.

there is a dual edge sword with Apple silicon only having one 'client' (Apple products). The number of Apple products is small. The general trend of Apple dumping sales of pervious gen with the newest gen pragmatically makes that list of eligible systems for deployment much smaller.

The old Xeon W chips in the Mac Pro lived off a subset of the larger. Xeon server market. That chip didn't pay for itself in the Intel land. ( Neither does Threadripper in AMD Epyc land.). the Intel Mac Pro largely lived off of a very large group of non Mac Pro users paying for the R&D. Once decoupled from those non-users then have very real issues.

Hand waving continuing Intel Mac Pro also ignores the huge shared R&D costs being paid on software (and hardware ) side by the rest of the Mac ecosystem which as moved on. Nobody paying for Intel code development/maintenance R&D means it doesn't get done eventually.


My concern about apple truly ditching the creatives openly is the halo effect, yes the middle tier creatives are served just fine by current consumer level Macs (I edit video just fine on my M4Max MBP with 128GB of RAM) but the heavy weight guys (like say movie SFX studios)

heavyweight SFX studios don't edit video. Once rendered it is not any particularly harder to do than video captured by a camera.

AI SFX slop is being produced at a way higher volume than the narrow heavyweight SFX houses could every do. That isn't where the edit volume is. And that is a much bigger halo , hype train.


are going to switch over to PC workstations (and switching isn't as hard as it used to be), where all their PCIe cards work, they can upgrade their machines ad nauseam and that halo effect will lose. It's similar to why Chevy and Ford produce halo editions of their Corvette and Mustang that are way up there in price, in tiny volumes too. Sure they are a side hustle almost, but they are a halo to the brand.

How many of these halo cars have saved GM, Chrysler , Ford. from Government bailouts when they got in trouble? The Mustang got canned before for being brought back to help push a new EV vehicle.

The 'halo' effect is grossly overblown on these boards. It has no where near the effect folks hand wave at it say it does.

Intel selling $6K server CPUs doesn't not significantly improve placement in the general PC laptop market if do not have a very good laptop SoCS for those systems. Intel had deficient laptop SoCs and Apple completely dropped them. What products you have in the appropriate product segments matters more. None of those USA muscle car vendors sell more cars than Toyota does. How many bailouts done for Toyota?



*** There are indications that the M3 Max that the laptops used is not the same base Max used for the Ultra. Millions of laptops have zero need for a UltraFusion connector. If the M3 Max die was already excessively large and difficult to make... incrementally bigger would not particularly make any economic sense. 'Halos' don't pay fab bills.

There is a major economic difficultly with try to slap together two large chips as very chunky chiplets. There is also probably tension in making the GPU clusters large and allocating the additional PCI-e headers that a Mac Pro needs for creditable/competitive PCI-e backhaul. Unless Apple makes major change to how they construct something like the Ultra it won't be viable to chuck them into trash at the same pace as the rest of the SoC dies they are doing.
 
Last edited:
  • Like
Reactions: throAU
RDMA over thunderbolt recently got enabled, so yeah for high end workloads you can cluster them fairly effectively up to say 4 nodes with 2TB combined ram and 4x the spec of a single ultra in cpu/gpu cores to work with, getting somewhere close to 3-4x the performance of a single box.

Its not quite the same as one computer but for extremely high end workloads that take a long time to run, clustering is pretty effective.

That can be achieved for say 40-60k USD.

For context that's around the cost of 1x H100 nvidia GPU without the system to put it in and provides an order of magnitude more VRAM for large data-sets (like extremely large LLMs). To get something with 1-2 TB of VRAM any other way you're talking multiple 30-50k GPUs (with 80GB or so each - so you need like 12 of them to get to 1TB VRAM), plus the servers to put them in, kilowatts of power, cooling, networking, server rack(s), switching, etc. so.... you do the math...

The performance will be a lot higher, sure - but you're talking hundreds of thousands of dollars / millions of dollars vs. under 100k. The jump in cost from clustered mac studios right now to anything better for high end GPU compute is HUGE.

If your workload is bigger than that, you're better off renting the compute or spending 10-20x the money (or more) on building your own PC hardware based clusters.
Thunderbolt is way too slow for this. The nVidia AI development systems have dual 200gbit network ports and a high end would be more like 800gbit/s these days…
 
Thunderbolt is way too slow for this. The nVidia AI development systems have dual 200gbit network ports and a high end would be more like 800gbit/s these days…

The Nvidia systems are 50k for a single H100 card - nobody is going to be running that either at home or in a small business.

You say it is "too slow", but people are literally doing with better performance than devices costing far more money.
 
Last edited:
Nobody said they would. For small requirements Apple make some great systems. Those systems have limitations and that’s what is being discussed here. A fully specced out Mac Pro (Intel) was tens of thousands too, and the wheels were famously expensive.
 
Coming up with increasingly absurd specs isn't going to make the "Apple should really improve the Mac Pro" pitch any more likely.
 
Thunderbolt is way too slow for this. The nVidia AI development systems have dual 200gbit network ports and a high end would be more like 800gbit/s these days…

and here I was thinking 40Gbps cat8 ethernet is the fastest networking cables on earth
 
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.