I bet there’s gonna be a banger of a Mac Studio announced in June.
Apple really stumbled into making the perfect hardware for home inference machines. Does any hardware company come close to Apple in terms of unified memory and single machines for high throughput inference workloads? Or even any DIY build?
When it comes to the previous “pro workloads,” like video rendering or software compilation, you’ve always been able to build a PC that outperforms any Apple machine at the same price point. But inference is unique because its performance scales with high memory throughput, and you can’t assemble that by wiring together off the shelf parts in a consumer form factor.
It’s simply not possible to DIY a homelab inference server better than the M3+ for inference workloads, at anywhere close to its price point.
They are perfectly positioned to capitalize on the next few years of model architecture developments. No wonder they haven’t bothered working on their own foundation models… they can let the rest of the industry do their work for them, and by the time their Gemini licensing deal expires, they’ll have their pick of the best models to embed with their hardware.
Jeff Geerling doing that 1.5TB cluster using 4 Mac Studios was pretty much all the proof needed to demo how the Mac Pro is struggling to find any place any more.
But those Thunderbolt links are slower than modern PCIe. If there's actually a M5-based Mac Studio with the same Thunderbolt support, you'll be better off e.g. for LLM inference, streaming read-only model weights from storage as we've seen with recent experiments than pushing the same amount of data via Thunderbolt. It's only if you want to go beyond local memory constraints (e.g. larger contexts) that the Thunderbolt link becomes useful.
Wasn't streaming models from storage into limited memory a case where it was impressive that you could make the elephant dance at all?
If you want to get usable speeds from very large models that haven't been quantitized to death on local machines, RDMA over Thunderbolt enables that use case.
Consumer PC GPUs don't have enough RAM, enterprise GPUs that can handle the load very well are obscenely expensive, Strix Halo tops out at 128 Gigs of RAM and is limited on Thunderbolt ports.
Why everyone wants to live in dongle/external cabling/dock hell is beyond me. PCIe cards are powered internally with no extra cables. They are secure. They do not move or fall off of shit. They do not require cable management or external power supplies. They do not have to talk to the CPU through a stupid USB hub or a Thunderbolt dock. Crappy USB HDMI capture on my Mac led me to running a fucking PC with slots to capture video off of a 50 foot HDMI cable, that then streamed the feed to my Mac from NDI, because it was more reliable than the elgarbo capture dongle I was using. This shit is bad. It sucks. It's twice the price and half the quality of a Blackmagic Design capture card. But, no slots, so I guess I can go get fucked.
For anything that's even somewhat in the consumer space rather than pure workstation/professional, the main reason is that dongles can be used with a laptop but add-in cards can't. When ordinary consumer PCs (or even office PCs) are in the picture, laptops are a huge chunk of the target audience.
The market segments that can afford to ignore laptops and only target permanently-installed desktops are mostly those niches where the desktop is installed alongside some other piece of equipment that is much more expensive.
I don't think Apple just stumbled into it, and while I totally agree that Apple is killing it with their unified memory, I think we're going to see a pivot from NVidia and AMD. The biggest reason, I think, is: OpenAI has committed to enormous amount capex it simply cannot afford. It does not have the lead it once did, and most end-users simply do not care. There are no network effects. Anthropic at this point has completely consumed, as far as I can tell, the developer market. The one market that is actually passionate about AI. That's largely due to huge advantage of the developer space being, end users cannot tell if an "AI" coded it or a human did. That's not true for almost every other application of AI at this point.
If the OpenAI domino falls, and I'd be happy to admit if I'm wrong, we're going to see a near catastrophic drop in prices for RAM and demand by the hyperscalers to well... scale. That massive drop will be completely and utterly OpenAI's fault for attempting to bite off more than it can chew. In order to shore up demand, we'll see NVidia and AMD start selling directly to consumers. We, developers, are consumers and drive demand at the enterprises we work for based on what keeps us both engaged and productive... the end result being: the ol' profit flywheel spinning.
Both NVidia and AMD are capable of building GPUs that absolutely wreck Apple's best. A huge reason for this is Apple needs unified memory to keep their money maker (laptops) profitable and performant; and while, it helps their profitability it also forces them into less performant solutions. If NVidia dropped a 128GB GPU with GDDR7 at $4k-- absolutely no one would be looking for a Mac for inference. My 5090 is unbelievably fast at inference even if it can't load gigantic models, and quite frankly the 6-bit quantized versions of Qwen 3.5 are fantastic, but if it could load larger open weight models I wouldn't even bother checking Apple's pricing page.
tldr; competition is as stiff as it is vicious-- Apple's "lead" in inference is only because NVidia and AMD are raking in cash selling to hyperscalers. If that cash cow goes tits up, there's no reason to assume NVidia and AMD won't definitively pull the the rug out from Apple.
As someone who came from the SGI O2/Octane era when high-end workstations were compact, distinctive, and sexy, I’ve never really understood the allure of the Mac Pro, with the exception of the 2013 Mac Pro tube, which I owned (small footprint, quiet, and powerful).
For me, aesthetics and size are important. That workstation on your desk should justify its presence, not just exist as some hulking box.
When Apple released the Mac Studio, it made perfect sense from a form-factor point-of-view. The internal expansion slots in the M2 Mac Pro didn't make any sense. It was like a bag of potato chips - mostly air. And far too big and ugly to be part of my work area! I'm surprised that Apple didn't discontinue it sooner.
As someone who worked on the M2 Mac Pro and has a real soft spot for it, I get it. It’s horrendously expensive and doesn’t offer much benefit over a Mac Studio and a thunderbolt pci chassis. My personal dream is that vms would support pci pass through and so you can just spin up a Linux vm and let it drive the gpus. But at that point, why are you buying a Mac?
They're trying to make it very clear they're not speaking on behalf of Apple Inc, despite having worked (or working) there.
Big companies like to give employees some minimal "media training", which mostly amounts to "do not speak for the company, do not say anything that might even slightly sound like you're speaking for the company".
SR-IOV and VFIO passthrough are different things. SR-IOV partitions a PCIe device across multiple VMs simultaneously (common for NICs and NVMe). VFIO passthrough gives one VM exclusive ownership of a physical device. For GPU compute you almost always want full passthrough, not SR-IOV partitioning.
The harder problem on Apple Silicon is that the M2 Ultra's GPU is integrated into the SoC -- it's not a discrete PCIe device you can isolate with an IOMMU group. Apple's Virtualization framework doesn't expose VFIO-equivalent hooks, so even if you add a discrete AMD Radeon to the Mac Pro's PCIe slots, there's no supported path to pass it through to a Linux guest right now.
On Intel Macs this actually worked via VFIO with the right IOMMU config. Apple Silicon VMs can do metal translation layers but that's not the same as bare-metal GPU access. It's a real limitation and I doubt Apple will prioritize solving it since it would undercut the "just use macOS" pitch.
As much as I love alluring designs such as the NeXT Cube (which I have), the Power Mac G4 Cube (which I wish I had), and the 2013 Mac Pro (which I also have), sometimes a person needs a big, hulking box of computational power with room for internal expansion, and from the first Quadra tower in the early 1990s until the 2012 Mac Pro was discontinued, and again from 2019 until today, Apple delivered this.
Even so, the ARM Mac Pro felt more like a halo car rather than a workhorse. The ARM Mac Pro may have been more compelling had it supported GPUs. Without this support, the price premium of the Mac Pro over the Mac Studio was too great to justify purchasing the Pro for many people, unless they absolutely needed internal expansion.
I’d love a user-upgradable Mac like my 2013 Mac Pro, but it’s clear that Apple has long moved on with its ARM Macs. I’ve moved on to the PC ecosystem. On one hand ARM Macs are quite powerful and energy-efficient, but on the other hand they’re very expensive for non-base RAM and storage configurations, though with today’s crazy prices for DDR5 RAM and NVMe SSDs, Apple’s prices for upgrades don’t look that bad by comparison.
It's dumb from a practical perspective. But I keep hoping they'll vertically compress their trashcan design so it looks like their Cupertino headquarters.
I'm surprised they even tried selling an Apple Silicon Mac Pro - I expected that product to die the moment they announced the transition. Everything that makes Apple Silicon great also makes it garbage for high-performance workstations.
The allure of the Mac Pro is that you could dodge the Apple Tax by loading it up with RAM and compute accelerators Apple couldn't mark up. Well, Apple Silicon works against all of that. The hardware fabric and PCIe controller specifically prohibit mapping PCIe device memory as memory[0], which means no GPU driver ever will work with it. Not even in Asahi Linux. And the RAM is soldered in for performance. An Ultra class chip has like 16 memory channels, which even in a 1-DIMM per channel routing would have trace lengths long enough to bottleneck operating frequency.
The only thing the socketed RAM Mac Pros could legitimately do that wasn't a way to circumvent Apple's pricing structure was take terabytes of memory - something that requires special memory types that Apple's memory controller IP likely does not support. Intel put in the engineering for it in Xeon and Apple got it for free before jumping ship.
Even then, all of this has gone completely backwards. Commodity DRAM is insanely expensive now and Apple's royalty-bearing RAM prices are actually reasonable in comparison. So there's no benefit to modularity anymore. Actually, it's a detriment, because price-discovery-enforcing scalpers can rip RAM out of perfectly working computers and resell the RAM. It's way harder to scalp RAM that's soldered on the board.
i wish i'd never traded in my 2016 mac pro (aluminum polished tube) as it was beefy, it was silent, clever thermo design (like the powerpc cube 20 years earlier or so), and i'd upgraded the living crap out of it for cheap.
Apple really dropped the ball here. They had every ability to make something competitive with Nvidia for AI training as well as inference, by selling high end multi GPU Mac Pro workstations as well as servers, but for some reason chose not to. They had the infrastructure and custom SoCs and everything. What a waste.
It really could have been a bigger market for them than even the iPhone.
Just about everybody who isn't Nvidia dropped the ball, bigtime.
Intel should have shipped their GPUs with much more VRAM from day one. If they had done this, they'd have carved out a massive niche and much more market share, and it would have been trivially simple to do.
AMD should have improved their tools and software, etc.
Apple should have done as you say.
Google had nigh on a decade to boost TPU production, and they're still somehow behind the curve.
Such a lack of vision. And thus Nvidia is, now quite durably, the most valuable company in the world. Imagine telling that to a time traveler from 2018.
I think for AMD, they were focused on competing against Intel. Remember AMD was almost bankrupt about 15 years ago because of competing against Intel. But the very first GPU use for AI was actually with an ATI/AMD GPU, not an Nvidia one. Everyone thinks Nvidia kicked off the GPU AI craze when Ilya Sutskever cleaned up on AlexNet with an Nvidia GPU back in 2012, or when Andrew Ng and team at Stanford published their "Large Scale Deep Unsupervised Learning using Graphics Processors" in 2009, but in 2004, a couple of Korean researchers were the first to implement neural networks on a GPU, using ATI Radeons: https://www.sciencedirect.com/science/article/abs/pii/S00313...
And as of now I do believe AMD is in the second strongest position in the datacenter space after Nvidia, ahead of even Google.
> And thus Nvidia is, now quite durably, the most valuable company in the world.
Nvidia is the most valuable company in the world right up until the AI bubble pops. Which, while it's hard to nail down when, is going to happen. I wouldn't call their position durable at all.
The crashing and burning of Nvidia stock has been predicted for a while now and keeps not really happening. It’s gone pretty flat and volatile up there around $180 but they keep delivering the results to back it up. I was thinking this week that Apple is really primed to make a killing from people who want to run their LLM on-device coupled with an agent in the next couple of years. We’re a long way off being able to train the models – this is going to need an Nvidia-powered datacentre for the foreseeable future, but the local inference seems absolutely like a market that Apple could capture, gutting all the most premium revenue from Anthropic and OpenAI by selling Macs with a large amount of integrated memory to anyone who wants to give them the money to run their native OpenClaw/agent instead of paying ever-growing monthly bills for tokens.
It is definitely a case that they will fall a long way but Nvidia will not fail as a whole. They have a way of maximizing their position relentlessly. CUDA turns out to endlessly put them in amazing positions on things like image recognition, AR, Crypto and now AI.
For all the faults of them leaning in hard on these things for stock market and personal gains, Nvidia still has some of the best quality products around. That is their saving grace.
They will not be the world most valuable company once the bubble pops, will probably never get back there again, but they will continue to be a decent enough business. I just want them going back to talking about graphics more than AI again, that will be nice.
Don’t mistake stock market performance for revenue. NVIDIA makes ~200B annually, same as what Apple makes from iPhones. It’s a big market but GPUs aren’t just AI.
I'm purely talking in terms of revenue. There's a huge demand for AI systems from personal workstations to datacenter servers, and Apple was one of the few companies in the world in a position to build complete systems for it.
But for some reason Apple thought the sound recording engineer or the video editor market was more important... like, WTF dude? Have some vision at least!
It is more important. Both for the customer base that actually buys Apple machines as well as the cache and mindshare of being used by the people that create American culture.
Even if Apple had an amazing GPU for AI it wouldn’t matter hugely - local inference hasn’t taken off yet and cloud inference and training all uses servers where Apple has no market share and wasn’t going to get it since people had already built all the stacks around CUDA before Apple could even have awoken to that.
Their rule of only releasing major software updates once a year in June is holding them back IMO. Their local LLM apis were dated before macOS/iOS 26 was even released. Just because something worked 20 years ago doesn’t mean it works today, but I’m sure it’s hard to argue against a historically successful strategy internally.
Those back log orders are wild! One does wonder that if the bubble collapses or more global upsets happen in that time, how many of those will ever be fulfilled? Reality might be not so impressive, but considering if it fell even 80%, that is still $200 B in revenue and that is huge.
Remember when a $1 billion valuation used to be a big thing? That is nothing compared with nowadays.
The Apple I was a pretty poor predictor of what mainstream mass-market computing was going to end up looking like. I don't think anybody has yet come up with the Apple II of local LLMs, let alone the VisiCalc or Windows 95.
How is this dropping the ball? I think they dropped the ball a long time ago by waiting until M5 to do integrated tensor cores instead of the separate ANE only which was present before.
For multi-gpu you can network multiple Macs at high speed now. Their biggest disadvantage to Nvidia right now is that no one wants to do kernel authoring in Metal. AMD learned that the hard way when they gave up on OpenCL and built HIP.
> something competitive with Nvidia for AI training
Apple is counting on something else: model shrink. Every one is now looking at "how do we make these smaller".
At some point a beefy Mac Studio and the "right sized" model is going to be what people want. Apple dumped a 4 pack of them in the hands of a lot of tech influencers a few months back and they were fairly interesting (expensive tho).
> Apple is counting on something else: model shrink
The most powerful AI interactions I've had involved giving a model a task and then fucking off. At that point, I don't actually care if it takes 5 minutes or an hour. I've cued up a list of background tasks it can work on, and that I can circle back to when I have time. In that context, smaller isn't even the virtue at hand–user patience is. Having a machine that works on my bullshit questions and modelling projects at one tenth the speed of a datacentre could still work out to being a good deal even before considering the privacy and lock-in problems.
Cool? And it has nothing to do with what kind of consumer hardware Apple should sell. If your use cases are literally "bigger model better" then the you should always use cloud. No matter how much computing power Apple squeezes into their device it won't be a mighty data center.
For running the model once it’s been trained, all a datacenter does is give you lower latency. Once the devices have a large enough memory to host the model locally, then the need to pay datacenter bills is going to be questioned. I’d rather run OpenClaw on my device plugged into a local LLM rather than rely on OpenAI or Claude.
> At some point a beefy Mac Studio and the "right sized" model is going to be what people want.
It's pretty clear that this isn't going to happen any time soon, if ever. You can't shrink the models without destroying their coherence, and this is a consistently robust observation across the board.
I don’t think it’s about literally shrinking the models via quantization, but rather training smaller/more efficient models from scratch
Smaller models have gotten much more powerful the last 2 years. Qwen 3.5 is one example of this. The cost/compute requirements of running the same level intelligence is going down
I have said for a while that we need a sort of big-little-big model situation.
The inputs are parsed with a large LLM. This gets passed on to a smaller hyper specific model. That outputs to a large LLM to make it readable.
Essentially you can blend two model type. Probabilistic Input > Deterministic function > Probabilistic Output. Have multiple little determainistic models that are choose for specific tasks. Now all of this is VERY easy to say, and VERY difficult to do.
But if it could be done, it would basically shrink all the models needed. Don't need a huge input/output model if it is more of an interpreter.
Yes, but bigger models are still more capable. Models shrinking (iso-performance) just means that people will train and use more capable models with a longer context.
Not all enterprises are the same, I imagine many companies have different departments working with local optimums, so someone who could benefit from it to get more productivity might not have access to it because the department that is doing hardware acquisition is being measured in isolation.
Give every iPhone family a in house Siri that will deal with canceling services and pursuing refunds.
Your customer screw up results in your site getting an agent drive DDOS on its CS department till you give in.
Siri: "Hey User, here's your daily update, I see you haven't been to the gym, would you like me to harass their customer service department till they let you out of their onerous contract?"
The 2019 Mac Pro’s main purpose was to provide much needed reassurance that Apple cared about the Mac. In prior years the quality of the Macs had fallen over all product lines. And the question of does Apple care about the Mac at all was a legitimate one.
This Mac Pro was about resetting and giving a clear signal that Apple was willing to invest in the Mac far more than it was about ‘slots’.
Today, Mac hardware is the best it has ever been, and no one is reasonably questioning apple’s commitment to a Mac hardware.
So it makes sense for the Mac Pro to make a graceful exit.
The Ultra variants of the M series chips had previously consisted of two of the Max chips bonded together.
The M5 generation Pro and Max chips have moved to a chiplet based architecture, with all the CPU cores on one chiplet, and all the GPU cores on another.
Still rocking a 2019 (Intel) Mac Pro here, all slots filled with various Pro Tools and UAD DSP cards, SSD, GPU, etc. I'm planning to get as much mileage out of it as I can. I'm sure a Studio would be more performant, but the Thunderbolt to PCIe chassis are not cheap.
While the trash can generation was somewhat present and around, I don't think I ever saw a cheese grater in the flesh. Did it have any users? Were there any actual useful expansion cards? Did anybody continue buying this at all, after it didn't get the M3 Ultra bump, that the Mac Studio got last year?
It had many hardware upgrades over the years - upgraded CPUs, 128GB RAM, 4TB NVME storage, a modern AMD GPU, USB3/c, thunderbolt, etc
The only reason it got replaced is because it became too much of a PITA to keep modern OSX running on it (via OCLP)
Replaced with an M4 Max Mac Studio, which is a nice and faster machine but with no ability to upgrade anything and much worse hardware resale value on M-series I'll have to replace it in 2-3 years
I'm a former 4,1 user, myself — replaced with an M2Pro mini Jan 2023 (finally retired fully 2025).
Absolutely recommend you purchase the 4-bay Terramaster external enclosure — gives you four SATA slots that are hot-swappable (unlike MacPro's). 10gbps via USB-C.
If you were using a 2009 Mac Pro for work until a year or two ago then you seriously need to think about how much your time is worth and how much of your time you were wasting by "saving money" on not buying a new computer.
If you're self employed, the cost of equipment and depreciation make hanging on to that 2009 system even more of a poor choice.
If you were still using a 2009 system I don't see why you'd "have to replace in 2-3 years."
The cheese grater mac pros were very popular, in that people got them and continued to use them.
The most notable feature was that there were mac-specific graphics cards, and you could also run PC graphics cards (without a nice boot screen). They had a 1.4kw power supply I believe, and there was extra pcie power for higher-end graphics cards. You could upgrade the memory, add up to 6 or more sata hard disks (2 in dvd slot). You could run windows, dual booting if you wanted and apple supported the drivers.
The 2013 was kind of a joke. small and quiet, but expansion was minimal.
2019 looked beefy, but the expansion was more like a cash register for apple, not really democratic. There were 3rd party sata hard disk solutions,
the 2023 model was basically a joke. I think maybe the pcie slots were ok for nvme cards, not a lot else (unless apple made it).
nowadays an apple computer is more like an iphone - apple would prefer if everything was welded shut.
I hate how I can't buy an new apple silicon with upgradeable RAM or SSD. Is there a legit reason why they couldn't make these things upgradeable at all even on a studio machine? 4TB is the smallest SSD I ever want in a new machine, but buying one from Apple is stupidly expensive. Back in the intel days, I'd buy a macbook pro, for example, with less ram and a smaller SSD than the max available and then upgrade to much cheaper aftermarket parts a few years later when prices dropped.
I'm still not going to use windows or linux. Don't want to be an IT guy on the side just to keep linux machines working. This may not be obvious to some unless you try to use printers and scanners that are more than 5 years old and what them to be on the network. And, you don't install virtualization tools like vmware that require compiling and loading kernel drivers which ends up being incompatible with new OS releases...etc.
Windows is just too much of a painful acceptance of mediocrity and apathy in product design for me.
> Is there a legit reason why they couldn't make these things upgradeable at all even on a studio machine?
For the SSD, no. For the memory, yes. The memory lives on the same chip as the CPU and the GPU, it's even more tightly bound than just being soldered on. The memory being there has legitimate technical benefits that make it much easier/cheaper for them to reach the extremely high memory bandwidths that they do.
>it's even more tightly bound than just being soldered on
No. There is a reason for it but no, it's just soldered on the same carrier board as the APU, in order to be really close to it. Apple could have used a form factor like CAMM2 and it would have worked the same, be it at slightly higher cost. The reason is simply to kill upgrade options and cut manufacturing costs - same as for any other soldered ram.
This, although it's not merely "easier/cheaper", it's "impossible" (unless you sacrifice a ton of performance)
Same reason as a) GDDR on dGPUs (I think I read somewhere that GDDR is very much like regular DDR, just with much tighter paths and thus soldered in) and b) Framework Desktop (performance would reportedly halve if RAM were not soldered)
SSD reasons I seem to recall are architectural for security: some parts (controller?) that usually sit on a NVMe SSD are embedded in the SoC next to (or inside?) the secure enclave processor or whatever the equivalent of the T2 thing is in Mx chips, so what you'd swap would be a bank of raw storage chips which don't match the controller.
Not surprising, as the market has broadly moved on from add-in cards in favor of smaller form factors and external devices, absent some notable holdouts in specific verticals.
Gonna miss it, though. If they had reduced the add-in card slots to something more reasonable, lowered the entry price, and given us multi-socket options for the CPU (2x M# Ultras? 4x?), it could have been an interesting HPC or server box - though they’ve long since moved away from that in software land, so that was always but a fantasy.
At least the Mac Studio and Minis are cute little boxes.
They've been trying to kill the Mac Pro for over a decade. I wonder how long before they backtrack again? It seems like they should at least have a migration path for users who needed the expansion cards the Mac Pro supported. Pushing them to the PC seems pretty bad.
Apple's new "Pro" definition seems more like "Prosumer".
The form-factor always felt like a weird fit for Apple Silicon. With the Intel boxes it was understandable; you want a few liters of free space for a couple AMD cards or some transcode hardware. The system was designed to be expandable, and the Mac Pro was the apex of Apple's commitment to that philosophy after bungling the trashcan Mac Pro.
None of the Apple Silicon hardware can seemingly justify this form factor, though. The memory isn't serviceable, PCIe devices aren't really supported, the PSU doesn't need much space, and the cooling can be handled with mobile-tier hardware. Apple's migration path is "my way or the highway" for Mac Pro owners.
Their justification for the form factor, when it was released, was that pro users need various PCI cards to interface with some of their equipment, and this would allow them to do that.
It seemed like the guts of the Mac Pro were essentially shoved inside of a box and stuck in the corner of the tower. It would seem like they could decouple it and sell a box that pro users could load cards into (like other companies do for eGPUs). It wouldn’t feel like a very Apple-like setup, but it would function and allow Apple to focus where they want to focus without simply leaving those users behind.
I suppose the other option would be to dispense with the smoke and mirrors and let people slot a Mac Studio right into the Mac Pro tower, so it could be upgraded independently of the tower.
The alternative is people leave the platform or end up with a bunch of Thunderbolt spaghetti. Neither of which seem ideal.
I suspect we'll start seeing higher-spec Mac Studio options.
One of those with an M* Ultra, and some sort of Thunderbolt storage expansion would probably cover most of the Pro's use cases. And Apple probably doesn't want to deal with anything more exotic than those.
I have three of the trash can ones. They are absolute pieces of art, as useless as they are computationally these days (energy-to-performance wise at least). I will never sell nor give them away.
Sad. I had this pipe dream of an Apple Silicon system made as a PCIe endpoint, so a Mac Pro could be a coordinator and host to like 4 of such systems in a cluster with very fast interconnect. Imagine the possibilities.
This honestly saddens me a little. From the PowerMac's to the MacPro I always loved them when having the opportunity to work with them. Plus I loved the expandability they offered.
I don't find the external GPU houses for Mac Studio as appealing to use.
With the popularity of mac mini (and macbooks for that matter) for doing ML/AI work, I would have thought Apple could make a Mac Pro that could make for a good workstation for doing in-house ML/AI stuff.
I bought a GPU maybe a decade ago for this, and it's not worth the hassle (for me at least), but a nice out-of-the box solution, I would pay for.
The problem is that the M1 chips foretold the doom of the Mac Pro unless they could figure out some way to do something that you couldn't do with a Mac Studio - thunderbolt is so good that it's hard to justify anything else.
If they had done more with NUMA in the M series maybe you could have a Mac Pro with M5 Ultras that can take a number of M5 "daughter cards" that do something useful.
This makes sense, for that kind of money you could always build a beastly workstation in a real ATX case with standard components. Install Linux and the Mac looks like an expensive toy in comparison.
Mac Studio waits for the Ultra chips to ship, which are always last in a generation. Perhaps the M5's chiplet architecture will help them move faster there.
Reading comments, I don’t think people are being completely fair here. For Intel and AMD to approach what Apple has accomplished they’re making many of the same compromises with Panther Lake and Ryzen AI Max. Apple chose to put disk controllers on their SoP rather than having them on the storage module. This shaves a tiny bit of latency. Worth it? No idea. I’m shit at hardware design.
As for not having a Pro or otherwise expandable system? It’s shit. They make several variations of their chips, and I don’t think it would hurt them to make an SoP for a socket, put a giant cooling system in it, and give it 10 or 12 PCIe slots. As for what would go in those slots? Make this beast rack mountable and people would toss better network cards, sound/video output or capture, storage controllers, and all kinds of other things in there. A key here would be to not charge so much just because they can. Make the price reasonable.
They have tried variations of this since time immemorial (we can argue about "price reasonablé") but there's just not much you can do with it that you can't do much cheaper or simpler in other ways.
The Xserve has been dead for 15 years now, and it was never tremendously amazing (though it was nice kit).
Apple apparently has some sort of "in-house" xserve-like thing they don't sell; but turning that into a product would likely be more useful than a Mac Pro, unless they add NUMA or some other way of allowing an M5 to access racks and racks of DIMMs.
The 2013 trash can was the end of the Mac Pro. It was never the same after that. The 2012 and earlier Mac Pros were awesome. I had a 2010 model. Here's what I loved:
• Multiple hard drive bays for easy swapping of disks, with a side panel that the user could open and close
• Expandable RAM
• Lots of ports, including audio
• The tower took up no desktop space
• It was relatively affordable, starting at $2500. Many software developers had one. (The 2019 and later Mac Pros were insanely expensive, starting at $6000.)
The Mac Studio is affordable, but it lacks those other features. It has more ports than other Macs but fewer in number and kind than the old Mac Pro, because the Mac Studio is a pointlessly small desktop instead of floor tower.
That's when they stopped designing computers for the pro market and started selling mid-century Danish furniture that can also edit videos.
I knew it was all over when third party companies had to develop the necessarily-awkward rack mount kits for those contraptions. If Apple actually cared about or understood their pro customers, they would have built a first party solution for their needs. Like sell an actual rack-mount computer again—the horror!
Instead, an editing suite got what looked like my bathroom wastebasket.
When it was introduced, Apple said the trash can was a revolution in cooling design.
Then they said they couldn't upgrade the components because of heat. Everyone knows that wasn't true.
By the time Apple said they had issues with it in 2017, AMD were offering 14nm GCN4 and 5 graphics (Polaris and Vega) compared to the 28nm GCN1 graphics in the FirePro range. Intel had moved from Ivy Bridge to Skylake for Xeons. And if they wanted to be really bold (doubtful, as the move to ARM was coming) then the 1st gen Epyc was on the market too.
Moore's Law didn't stop applying for 6 years. They had options and chose to abandon their flagship product (and most loyal customers) instead.
The biggest issue was actually that the Mac Pro was designed specifically for dual GPUs- in the era of SLI this made some sense, but once that technology was abandoned it was a technological dead-end.
If you take one apart you'll see why, it's not the case that you could have ever swapped around the components to make it dual-CPU instead; it really was "dual GPU or bust".
Somewhat ironically, in todays ML ecosystem, that architecture would probably do great. Though I doubt it could possibly do better than what the M-series is doing by itself using unified memory.
I'll admit that while I've used the trash can but never taken it apart myself. But I can't imagine it would have been impossible to throw 2x Polaris 10 GPUs on the daughterboards in place of the FirePros.
For what is essentially a dead-end technology, I'm somewhat doubtful people would have bought it (since the second GPU is going to be idle and add to the cost massively).
the CPU being upgraded would have been much easier though I think.
Aside from the GPU mess, the 2013 was a nice machine, basically a proto-Mac Studio. Aside from software, the only thing that pushed me off my D300/64GB/12-core as an everyday desktop + front-end machine is the fact that there's no economically sensible way to get 4K video at 120 Hz given that an eGPU enclosure + a decent AMD GPU would cost as much as a Mac mini, so I'm slumming it in Windows for a few months until the smoke clears from the next Mac Studio announcement.
At which point I'll decide whether to replace my Mac Pro with a Mac Studio or a Linux workstation; honestly, I'm about 60/40 leaning towards Linux at this point, in which case I'd also buy a lower-end Mac, probably a MacBook Air.
I'm in the Linux desktop / Mac laptop camp, and it works well for me. Prevents me getting too tied up in any one ecosystem so that I can jump ship if Apple start releasing duds again.
I guess A/V pros are used to getting screwed constantly, but it must be really irritating to face the prospect of eventually having to move PCI add-in cards to TB5 enclosures that cost $1000 per slot.
They replaced it with Mac Neo. Did you notice the wonderful build quality, the accesible price and that everyone is buying it ?
And it has USB: U from universal.
If I remember correctly, the maximum configuration was something like $35k back in the day. I wonder what those people feel like now. On the other hand, if they have $35k to burn, probably they don't even think about it.
If you spend $35k and just idle the machine or just check e-mail you've burnt the money. If it's your work machine and you've got a $100/hr billable rate it's paid for in a little over a month. Three months at a $50/hr rate.
If you bought the $35k Mac Pro in 2023 when it was released and have a $50/hr rate it's been paid off for about 30 months. So as of today those owners probably aren't too broken hearted. They'll likely get at least another three years out of them.
People buying $35k Mac Pros probably paid them off after a single contract. So they've just been making money rather than costing money.
Siracusa—probably best known here for doing fabulous OS X reviews for Ars—is a co-host of ATP. He is also known is such circles for having Mac Pros, and using them for a long time (sometimes by choice, sometimes by circumstance). He thinks Apple should make a Mac Pro, not necessarily because it's a big seller, but because he thinks Apple should make a "best computer," much in the same way car companies might make a car that will never sell but pushes engineers, etc.
Ages ago, when new Mac hardware came out, I'd amuse myself by putting together an "ultimate Mac workstation" in the configurator --- once upon a time, one could hit 6 figures pretty easily --- these days, well I panic bought a duplicate computer because I was worried a chipped/cracked display was going to make it unusable (turns out a screen protector has worked thus far).
I agree with the reasoning, and would like to see Apple continue to make aspirational hardware, but maybe the mainstream stuff is good enough?
Even Siracusa admits that - he's found it hard to articulate what a true "Mac Pro" would do that you can't do with other things.
Back in the heyday of the $100k Mac Pro you could certainly imagine it doing things that wouldn't be easily done by anything under $50k, and it would look good doing it.
I bet there’s gonna be a banger of a Mac Studio announced in June.
Apple really stumbled into making the perfect hardware for home inference machines. Does any hardware company come close to Apple in terms of unified memory and single machines for high throughput inference workloads? Or even any DIY build?
When it comes to the previous “pro workloads,” like video rendering or software compilation, you’ve always been able to build a PC that outperforms any Apple machine at the same price point. But inference is unique because its performance scales with high memory throughput, and you can’t assemble that by wiring together off the shelf parts in a consumer form factor.
It’s simply not possible to DIY a homelab inference server better than the M3+ for inference workloads, at anywhere close to its price point.
They are perfectly positioned to capitalize on the next few years of model architecture developments. No wonder they haven’t bothered working on their own foundation models… they can let the rest of the industry do their work for them, and by the time their Gemini licensing deal expires, they’ll have their pick of the best models to embed with their hardware.
DGX workstations, expensive but allow PCI cards as well.
https://marketplace.nvidia.com/en-us/enterprise/personal-ai-...
Jeff Geerling doing that 1.5TB cluster using 4 Mac Studios was pretty much all the proof needed to demo how the Mac Pro is struggling to find any place any more.
https://www.jeffgeerling.com/blog/2025/15-tb-vram-on-mac-stu...
But those Thunderbolt links are slower than modern PCIe. If there's actually a M5-based Mac Studio with the same Thunderbolt support, you'll be better off e.g. for LLM inference, streaming read-only model weights from storage as we've seen with recent experiments than pushing the same amount of data via Thunderbolt. It's only if you want to go beyond local memory constraints (e.g. larger contexts) that the Thunderbolt link becomes useful.
Wasn't streaming models from storage into limited memory a case where it was impressive that you could make the elephant dance at all?
If you want to get usable speeds from very large models that haven't been quantitized to death on local machines, RDMA over Thunderbolt enables that use case.
Consumer PC GPUs don't have enough RAM, enterprise GPUs that can handle the load very well are obscenely expensive, Strix Halo tops out at 128 Gigs of RAM and is limited on Thunderbolt ports.
Why everyone wants to live in dongle/external cabling/dock hell is beyond me. PCIe cards are powered internally with no extra cables. They are secure. They do not move or fall off of shit. They do not require cable management or external power supplies. They do not have to talk to the CPU through a stupid USB hub or a Thunderbolt dock. Crappy USB HDMI capture on my Mac led me to running a fucking PC with slots to capture video off of a 50 foot HDMI cable, that then streamed the feed to my Mac from NDI, because it was more reliable than the elgarbo capture dongle I was using. This shit is bad. It sucks. It's twice the price and half the quality of a Blackmagic Design capture card. But, no slots, so I guess I can go get fucked.
For anything that's even somewhat in the consumer space rather than pure workstation/professional, the main reason is that dongles can be used with a laptop but add-in cards can't. When ordinary consumer PCs (or even office PCs) are in the picture, laptops are a huge chunk of the target audience.
The market segments that can afford to ignore laptops and only target permanently-installed desktops are mostly those niches where the desktop is installed alongside some other piece of equipment that is much more expensive.
For LLMs and other pure memory-bound workloads, but for eg. diffusion models their FPU SIMD performance is lacking.
I don't think Apple just stumbled into it, and while I totally agree that Apple is killing it with their unified memory, I think we're going to see a pivot from NVidia and AMD. The biggest reason, I think, is: OpenAI has committed to enormous amount capex it simply cannot afford. It does not have the lead it once did, and most end-users simply do not care. There are no network effects. Anthropic at this point has completely consumed, as far as I can tell, the developer market. The one market that is actually passionate about AI. That's largely due to huge advantage of the developer space being, end users cannot tell if an "AI" coded it or a human did. That's not true for almost every other application of AI at this point.
If the OpenAI domino falls, and I'd be happy to admit if I'm wrong, we're going to see a near catastrophic drop in prices for RAM and demand by the hyperscalers to well... scale. That massive drop will be completely and utterly OpenAI's fault for attempting to bite off more than it can chew. In order to shore up demand, we'll see NVidia and AMD start selling directly to consumers. We, developers, are consumers and drive demand at the enterprises we work for based on what keeps us both engaged and productive... the end result being: the ol' profit flywheel spinning.
Both NVidia and AMD are capable of building GPUs that absolutely wreck Apple's best. A huge reason for this is Apple needs unified memory to keep their money maker (laptops) profitable and performant; and while, it helps their profitability it also forces them into less performant solutions. If NVidia dropped a 128GB GPU with GDDR7 at $4k-- absolutely no one would be looking for a Mac for inference. My 5090 is unbelievably fast at inference even if it can't load gigantic models, and quite frankly the 6-bit quantized versions of Qwen 3.5 are fantastic, but if it could load larger open weight models I wouldn't even bother checking Apple's pricing page.
tldr; competition is as stiff as it is vicious-- Apple's "lead" in inference is only because NVidia and AMD are raking in cash selling to hyperscalers. If that cash cow goes tits up, there's no reason to assume NVidia and AMD won't definitively pull the the rug out from Apple.
As someone who came from the SGI O2/Octane era when high-end workstations were compact, distinctive, and sexy, I’ve never really understood the allure of the Mac Pro, with the exception of the 2013 Mac Pro tube, which I owned (small footprint, quiet, and powerful).
For me, aesthetics and size are important. That workstation on your desk should justify its presence, not just exist as some hulking box.
When Apple released the Mac Studio, it made perfect sense from a form-factor point-of-view. The internal expansion slots in the M2 Mac Pro didn't make any sense. It was like a bag of potato chips - mostly air. And far too big and ugly to be part of my work area! I'm surprised that Apple didn't discontinue it sooner.
As someone who worked on the M2 Mac Pro and has a real soft spot for it, I get it. It’s horrendously expensive and doesn’t offer much benefit over a Mac Studio and a thunderbolt pci chassis. My personal dream is that vms would support pci pass through and so you can just spin up a Linux vm and let it drive the gpus. But at that point, why are you buying a Mac?
Opinions are my own obvs.
> Opinions are my own obvs.
Whose else would they be?
> as someone who worked on the m2 mac pro
They're trying to make it very clear they're not speaking on behalf of Apple Inc, despite having worked (or working) there.
Big companies like to give employees some minimal "media training", which mostly amounts to "do not speak for the company, do not say anything that might even slightly sound like you're speaking for the company".
Maybe he was trying to say he isn’t a spokesman for anyone else :-)
> My personal dream is that vms would support pci pass through and so you can just spin up a Linux vm and let it drive the gpus.
SR-IOV is just that? and is well supported by both Windows and Linux.
SR-IOV and VFIO passthrough are different things. SR-IOV partitions a PCIe device across multiple VMs simultaneously (common for NICs and NVMe). VFIO passthrough gives one VM exclusive ownership of a physical device. For GPU compute you almost always want full passthrough, not SR-IOV partitioning.
The harder problem on Apple Silicon is that the M2 Ultra's GPU is integrated into the SoC -- it's not a discrete PCIe device you can isolate with an IOMMU group. Apple's Virtualization framework doesn't expose VFIO-equivalent hooks, so even if you add a discrete AMD Radeon to the Mac Pro's PCIe slots, there's no supported path to pass it through to a Linux guest right now.
On Intel Macs this actually worked via VFIO with the right IOMMU config. Apple Silicon VMs can do metal translation layers but that's not the same as bare-metal GPU access. It's a real limitation and I doubt Apple will prioritize solving it since it would undercut the "just use macOS" pitch.
As much as I love alluring designs such as the NeXT Cube (which I have), the Power Mac G4 Cube (which I wish I had), and the 2013 Mac Pro (which I also have), sometimes a person needs a big, hulking box of computational power with room for internal expansion, and from the first Quadra tower in the early 1990s until the 2012 Mac Pro was discontinued, and again from 2019 until today, Apple delivered this.
Even so, the ARM Mac Pro felt more like a halo car rather than a workhorse. The ARM Mac Pro may have been more compelling had it supported GPUs. Without this support, the price premium of the Mac Pro over the Mac Studio was too great to justify purchasing the Pro for many people, unless they absolutely needed internal expansion.
I’d love a user-upgradable Mac like my 2013 Mac Pro, but it’s clear that Apple has long moved on with its ARM Macs. I’ve moved on to the PC ecosystem. On one hand ARM Macs are quite powerful and energy-efficient, but on the other hand they’re very expensive for non-base RAM and storage configurations, though with today’s crazy prices for DDR5 RAM and NVMe SSDs, Apple’s prices for upgrades don’t look that bad by comparison.
> aesthetics and size are important
It's dumb from a practical perspective. But I keep hoping they'll vertically compress their trashcan design so it looks like their Cupertino headquarters.
> That workstation on your desk should
Under your desk, right? Right?!
I mean... if you spent $7,000 on it, do you really want to hide it away under the desk?
Yes, because you’re buying a tool, not a conversation piece.
I'm surprised they even tried selling an Apple Silicon Mac Pro - I expected that product to die the moment they announced the transition. Everything that makes Apple Silicon great also makes it garbage for high-performance workstations.
The allure of the Mac Pro is that you could dodge the Apple Tax by loading it up with RAM and compute accelerators Apple couldn't mark up. Well, Apple Silicon works against all of that. The hardware fabric and PCIe controller specifically prohibit mapping PCIe device memory as memory[0], which means no GPU driver ever will work with it. Not even in Asahi Linux. And the RAM is soldered in for performance. An Ultra class chip has like 16 memory channels, which even in a 1-DIMM per channel routing would have trace lengths long enough to bottleneck operating frequency.
The only thing the socketed RAM Mac Pros could legitimately do that wasn't a way to circumvent Apple's pricing structure was take terabytes of memory - something that requires special memory types that Apple's memory controller IP likely does not support. Intel put in the engineering for it in Xeon and Apple got it for free before jumping ship.
Even then, all of this has gone completely backwards. Commodity DRAM is insanely expensive now and Apple's royalty-bearing RAM prices are actually reasonable in comparison. So there's no benefit to modularity anymore. Actually, it's a detriment, because price-discovery-enforcing scalpers can rip RAM out of perfectly working computers and resell the RAM. It's way harder to scalp RAM that's soldered on the board.
[0] In violation of ARM spec, even!
>That workstation on your desk should justify its presence
It does the work you want it to do is not enough to justify its presence ?
i wish i'd never traded in my 2016 mac pro (aluminum polished tube) as it was beefy, it was silent, clever thermo design (like the powerpc cube 20 years earlier or so), and i'd upgraded the living crap out of it for cheap.
Apple really dropped the ball here. They had every ability to make something competitive with Nvidia for AI training as well as inference, by selling high end multi GPU Mac Pro workstations as well as servers, but for some reason chose not to. They had the infrastructure and custom SoCs and everything. What a waste.
It really could have been a bigger market for them than even the iPhone.
Just about everybody who isn't Nvidia dropped the ball, bigtime.
Intel should have shipped their GPUs with much more VRAM from day one. If they had done this, they'd have carved out a massive niche and much more market share, and it would have been trivially simple to do.
AMD should have improved their tools and software, etc.
Apple should have done as you say.
Google had nigh on a decade to boost TPU production, and they're still somehow behind the curve.
Such a lack of vision. And thus Nvidia is, now quite durably, the most valuable company in the world. Imagine telling that to a time traveler from 2018.
I think for AMD, they were focused on competing against Intel. Remember AMD was almost bankrupt about 15 years ago because of competing against Intel. But the very first GPU use for AI was actually with an ATI/AMD GPU, not an Nvidia one. Everyone thinks Nvidia kicked off the GPU AI craze when Ilya Sutskever cleaned up on AlexNet with an Nvidia GPU back in 2012, or when Andrew Ng and team at Stanford published their "Large Scale Deep Unsupervised Learning using Graphics Processors" in 2009, but in 2004, a couple of Korean researchers were the first to implement neural networks on a GPU, using ATI Radeons: https://www.sciencedirect.com/science/article/abs/pii/S00313...
And as of now I do believe AMD is in the second strongest position in the datacenter space after Nvidia, ahead of even Google.
> And thus Nvidia is, now quite durably, the most valuable company in the world.
Nvidia is the most valuable company in the world right up until the AI bubble pops. Which, while it's hard to nail down when, is going to happen. I wouldn't call their position durable at all.
The crashing and burning of Nvidia stock has been predicted for a while now and keeps not really happening. It’s gone pretty flat and volatile up there around $180 but they keep delivering the results to back it up. I was thinking this week that Apple is really primed to make a killing from people who want to run their LLM on-device coupled with an agent in the next couple of years. We’re a long way off being able to train the models – this is going to need an Nvidia-powered datacentre for the foreseeable future, but the local inference seems absolutely like a market that Apple could capture, gutting all the most premium revenue from Anthropic and OpenAI by selling Macs with a large amount of integrated memory to anyone who wants to give them the money to run their native OpenClaw/agent instead of paying ever-growing monthly bills for tokens.
It is definitely a case that they will fall a long way but Nvidia will not fail as a whole. They have a way of maximizing their position relentlessly. CUDA turns out to endlessly put them in amazing positions on things like image recognition, AR, Crypto and now AI.
For all the faults of them leaning in hard on these things for stock market and personal gains, Nvidia still has some of the best quality products around. That is their saving grace.
They will not be the world most valuable company once the bubble pops, will probably never get back there again, but they will continue to be a decent enough business. I just want them going back to talking about graphics more than AI again, that will be nice.
Don’t mistake stock market performance for revenue. NVIDIA makes ~200B annually, same as what Apple makes from iPhones. It’s a big market but GPUs aren’t just AI.
I'm purely talking in terms of revenue. There's a huge demand for AI systems from personal workstations to datacenter servers, and Apple was one of the few companies in the world in a position to build complete systems for it.
But for some reason Apple thought the sound recording engineer or the video editor market was more important... like, WTF dude? Have some vision at least!
It is more important. Both for the customer base that actually buys Apple machines as well as the cache and mindshare of being used by the people that create American culture.
Even if Apple had an amazing GPU for AI it wouldn’t matter hugely - local inference hasn’t taken off yet and cloud inference and training all uses servers where Apple has no market share and wasn’t going to get it since people had already built all the stacks around CUDA before Apple could even have awoken to that.
Some people at Apple see it. That’s why they added matmul to M5 GPU and keep mentioning LMStudio in their marketing.
Their rule of only releasing major software updates once a year in June is holding them back IMO. Their local LLM apis were dated before macOS/iOS 26 was even released. Just because something worked 20 years ago doesn’t mean it works today, but I’m sure it’s hard to argue against a historically successful strategy internally.
Huh? What local LLM apis? It uses Metal.
$280b and growing 70% YoY.
$1t backlog in orders in next 2 years.
Those back log orders are wild! One does wonder that if the bubble collapses or more global upsets happen in that time, how many of those will ever be fulfilled? Reality might be not so impressive, but considering if it fell even 80%, that is still $200 B in revenue and that is huge.
Remember when a $1 billion valuation used to be a big thing? That is nothing compared with nowadays.
Just look at the price of H100 cloud rental prices. Demand is increasing.
Nah, Apple made the right choice. Nobody except a niche market of hobbyists is interested in running tiny quantized models.
About the same niche market as the people who bought the Apple I, and we know where that went.
The Apple I was a pretty poor predictor of what mainstream mass-market computing was going to end up looking like. I don't think anybody has yet come up with the Apple II of local LLMs, let alone the VisiCalc or Windows 95.
this is what needs to come back with modern hardware and modern interconnect
https://en.wikipedia.org/wiki/Xserve
> They had the infrastructure and custom SoCs and everything. What a waste.
What are they wasting, exactly?
How is this dropping the ball? I think they dropped the ball a long time ago by waiting until M5 to do integrated tensor cores instead of the separate ANE only which was present before.
For multi-gpu you can network multiple Macs at high speed now. Their biggest disadvantage to Nvidia right now is that no one wants to do kernel authoring in Metal. AMD learned that the hard way when they gave up on OpenCL and built HIP.
> something competitive with Nvidia for AI training
Apple is counting on something else: model shrink. Every one is now looking at "how do we make these smaller".
At some point a beefy Mac Studio and the "right sized" model is going to be what people want. Apple dumped a 4 pack of them in the hands of a lot of tech influencers a few months back and they were fairly interesting (expensive tho).
> Apple is counting on something else: model shrink
The most powerful AI interactions I've had involved giving a model a task and then fucking off. At that point, I don't actually care if it takes 5 minutes or an hour. I've cued up a list of background tasks it can work on, and that I can circle back to when I have time. In that context, smaller isn't even the virtue at hand–user patience is. Having a machine that works on my bullshit questions and modelling projects at one tenth the speed of a datacentre could still work out to being a good deal even before considering the privacy and lock-in problems.
Cool? And it has nothing to do with what kind of consumer hardware Apple should sell. If your use cases are literally "bigger model better" then the you should always use cloud. No matter how much computing power Apple squeezes into their device it won't be a mighty data center.
For running the model once it’s been trained, all a datacenter does is give you lower latency. Once the devices have a large enough memory to host the model locally, then the need to pay datacenter bills is going to be questioned. I’d rather run OpenClaw on my device plugged into a local LLM rather than rely on OpenAI or Claude.
> it has nothing to do with what kind of consumer hardware Apple should sell
Of course it does. What I'm describing is absolutely not "bigger model better." It's a beachhead into that strategy's capital city.
What "tooling" do you use to let AIs work unattended for long periods?
> At some point a beefy Mac Studio and the "right sized" model is going to be what people want.
It's pretty clear that this isn't going to happen any time soon, if ever. You can't shrink the models without destroying their coherence, and this is a consistently robust observation across the board.
I don’t think it’s about literally shrinking the models via quantization, but rather training smaller/more efficient models from scratch
Smaller models have gotten much more powerful the last 2 years. Qwen 3.5 is one example of this. The cost/compute requirements of running the same level intelligence is going down
I have said for a while that we need a sort of big-little-big model situation.
The inputs are parsed with a large LLM. This gets passed on to a smaller hyper specific model. That outputs to a large LLM to make it readable.
Essentially you can blend two model type. Probabilistic Input > Deterministic function > Probabilistic Output. Have multiple little determainistic models that are choose for specific tasks. Now all of this is VERY easy to say, and VERY difficult to do.
But if it could be done, it would basically shrink all the models needed. Don't need a huge input/output model if it is more of an interpreter.
Yes, but bigger models are still more capable. Models shrinking (iso-performance) just means that people will train and use more capable models with a longer context.
Of course they are! Both are important and will be around and used for different reasons
Cheaper than what you’d expect though. You could get a nice setup for $20-40k 6mo ago. As far as enterprise investments go, that’s a rounding error.
Not all enterprises are the same, I imagine many companies have different departments working with local optimums, so someone who could benefit from it to get more productivity might not have access to it because the department that is doing hardware acquisition is being measured in isolation.
Drop that down to 5k, and make it useful.
Give every iPhone family a in house Siri that will deal with canceling services and pursuing refunds.
Your customer screw up results in your site getting an agent drive DDOS on its CS department till you give in.
Siri: "Hey User, here's your daily update, I see you haven't been to the gym, would you like me to harass their customer service department till they let you out of their onerous contract?"
Nothing is a bigger market than the iPhone, let alone expensive niche machines.
The 2019 Mac Pro’s main purpose was to provide much needed reassurance that Apple cared about the Mac. In prior years the quality of the Macs had fallen over all product lines. And the question of does Apple care about the Mac at all was a legitimate one.
This Mac Pro was about resetting and giving a clear signal that Apple was willing to invest in the Mac far more than it was about ‘slots’.
Today, Mac hardware is the best it has ever been, and no one is reasonably questioning apple’s commitment to a Mac hardware.
So it makes sense for the Mac Pro to make a graceful exit.
The Ultra variants of the M series chips had previously consisted of two of the Max chips bonded together.
The M5 generation Pro and Max chips have moved to a chiplet based architecture, with all the CPU cores on one chiplet, and all the GPU cores on another.
https://www.wikipedia.org/wiki/Apple_M5
So what will the M5 Ultra look like?
If you integrate two CPU chiplets and two GPU chiplets, you're looking at 36 CPU cores, 80 GPU cores, and 1228 GB/s of memory bandwidth.
Still rocking a 2019 (Intel) Mac Pro here, all slots filled with various Pro Tools and UAD DSP cards, SSD, GPU, etc. I'm planning to get as much mileage out of it as I can. I'm sure a Studio would be more performant, but the Thunderbolt to PCIe chassis are not cheap.
While the trash can generation was somewhat present and around, I don't think I ever saw a cheese grater in the flesh. Did it have any users? Were there any actual useful expansion cards? Did anybody continue buying this at all, after it didn't get the M3 Ultra bump, that the Mac Studio got last year?
I just replaced a 2009 MacPro
It had many hardware upgrades over the years - upgraded CPUs, 128GB RAM, 4TB NVME storage, a modern AMD GPU, USB3/c, thunderbolt, etc
The only reason it got replaced is because it became too much of a PITA to keep modern OSX running on it (via OCLP)
Replaced with an M4 Max Mac Studio, which is a nice and faster machine but with no ability to upgrade anything and much worse hardware resale value on M-series I'll have to replace it in 2-3 years
I'm a former 4,1 user, myself — replaced with an M2Pro mini Jan 2023 (finally retired fully 2025).
Absolutely recommend you purchase the 4-bay Terramaster external enclosure — gives you four SATA slots that are hot-swappable (unlike MacPro's). 10gbps via USB-C.
Are SMART attributes independently readable on that? (Or any multi-drive USB-C enclosures?)
If you were using a 2009 Mac Pro for work until a year or two ago then you seriously need to think about how much your time is worth and how much of your time you were wasting by "saving money" on not buying a new computer.
If you're self employed, the cost of equipment and depreciation make hanging on to that 2009 system even more of a poor choice.
If you were still using a 2009 system I don't see why you'd "have to replace in 2-3 years."
The cheese grater mac pros were very popular, in that people got them and continued to use them.
The most notable feature was that there were mac-specific graphics cards, and you could also run PC graphics cards (without a nice boot screen). They had a 1.4kw power supply I believe, and there was extra pcie power for higher-end graphics cards. You could upgrade the memory, add up to 6 or more sata hard disks (2 in dvd slot). You could run windows, dual booting if you wanted and apple supported the drivers.
The 2013 was kind of a joke. small and quiet, but expansion was minimal.
2019 looked beefy, but the expansion was more like a cash register for apple, not really democratic. There were 3rd party sata hard disk solutions,
the 2023 model was basically a joke. I think maybe the pcie slots were ok for nvme cards, not a lot else (unless apple made it).
nowadays an apple computer is more like an iphone - apple would prefer if everything was welded shut.
My first non-Linux PC was a cheese grater, way overkill for my needs but served me well for many years.
I feel like Apple is going back to the days of toaster "appliance" Macs. No slots, no upgrades, just buy a new one in 3-7 years.
I hate how I can't buy an new apple silicon with upgradeable RAM or SSD. Is there a legit reason why they couldn't make these things upgradeable at all even on a studio machine? 4TB is the smallest SSD I ever want in a new machine, but buying one from Apple is stupidly expensive. Back in the intel days, I'd buy a macbook pro, for example, with less ram and a smaller SSD than the max available and then upgrade to much cheaper aftermarket parts a few years later when prices dropped.
I'm still not going to use windows or linux. Don't want to be an IT guy on the side just to keep linux machines working. This may not be obvious to some unless you try to use printers and scanners that are more than 5 years old and what them to be on the network. And, you don't install virtualization tools like vmware that require compiling and loading kernel drivers which ends up being incompatible with new OS releases...etc.
Windows is just too much of a painful acceptance of mediocrity and apathy in product design for me.
> Is there a legit reason why they couldn't make these things upgradeable at all even on a studio machine?
For the SSD, no. For the memory, yes. The memory lives on the same chip as the CPU and the GPU, it's even more tightly bound than just being soldered on. The memory being there has legitimate technical benefits that make it much easier/cheaper for them to reach the extremely high memory bandwidths that they do.
>it's even more tightly bound than just being soldered on
No. There is a reason for it but no, it's just soldered on the same carrier board as the APU, in order to be really close to it. Apple could have used a form factor like CAMM2 and it would have worked the same, be it at slightly higher cost. The reason is simply to kill upgrade options and cut manufacturing costs - same as for any other soldered ram.
This, although it's not merely "easier/cheaper", it's "impossible" (unless you sacrifice a ton of performance)
Same reason as a) GDDR on dGPUs (I think I read somewhere that GDDR is very much like regular DDR, just with much tighter paths and thus soldered in) and b) Framework Desktop (performance would reportedly halve if RAM were not soldered)
SSD reasons I seem to recall are architectural for security: some parts (controller?) that usually sit on a NVMe SSD are embedded in the SoC next to (or inside?) the secure enclave processor or whatever the equivalent of the T2 thing is in Mx chips, so what you'd swap would be a bank of raw storage chips which don't match the controller.
Apparently upgrading the SSD can be done, but it's a weird form factor and you need another Mac to restore it.
Even the toaster appliance Mac’s had upgradeable ram and hard drives though. But it does seem like that to me also.
Not surprising, as the market has broadly moved on from add-in cards in favor of smaller form factors and external devices, absent some notable holdouts in specific verticals.
Gonna miss it, though. If they had reduced the add-in card slots to something more reasonable, lowered the entry price, and given us multi-socket options for the CPU (2x M# Ultras? 4x?), it could have been an interesting HPC or server box - though they’ve long since moved away from that in software land, so that was always but a fantasy.
At least the Mac Studio and Minis are cute little boxes.
They've been trying to kill the Mac Pro for over a decade. I wonder how long before they backtrack again? It seems like they should at least have a migration path for users who needed the expansion cards the Mac Pro supported. Pushing them to the PC seems pretty bad.
Apple's new "Pro" definition seems more like "Prosumer".
The migration path is Thunderbolt PCIe enclosures (basically eGPU enclosures but you don't have to use a GPU).
> but you don't have to use a GPU
That's a cute way of saying that GPUs aren't supported.
The form-factor always felt like a weird fit for Apple Silicon. With the Intel boxes it was understandable; you want a few liters of free space for a couple AMD cards or some transcode hardware. The system was designed to be expandable, and the Mac Pro was the apex of Apple's commitment to that philosophy after bungling the trashcan Mac Pro.
None of the Apple Silicon hardware can seemingly justify this form factor, though. The memory isn't serviceable, PCIe devices aren't really supported, the PSU doesn't need much space, and the cooling can be handled with mobile-tier hardware. Apple's migration path is "my way or the highway" for Mac Pro owners.
Their justification for the form factor, when it was released, was that pro users need various PCI cards to interface with some of their equipment, and this would allow them to do that.
It seemed like the guts of the Mac Pro were essentially shoved inside of a box and stuck in the corner of the tower. It would seem like they could decouple it and sell a box that pro users could load cards into (like other companies do for eGPUs). It wouldn’t feel like a very Apple-like setup, but it would function and allow Apple to focus where they want to focus without simply leaving those users behind.
I suppose the other option would be to dispense with the smoke and mirrors and let people slot a Mac Studio right into the Mac Pro tower, so it could be upgraded independently of the tower.
The alternative is people leave the platform or end up with a bunch of Thunderbolt spaghetti. Neither of which seem ideal.
It was always strange the Apple Silicon kept the 1.3kw power supply which was massive overkill.
I always hoped we’d get a consumer version of what they have internally - 10 or 20 or more Apple Silicon chips for 1000 cores or so.
A bunch of the Mac Pro decisions seem to have been more driven by "we have a warehouse of these parts" than "this is what the system needs".
A lot of Apples offerings feel a bit like that actually.
To be expected when lord of the supply chain Tim Cook is running the show.
I suspect we'll start seeing higher-spec Mac Studio options.
One of those with an M* Ultra, and some sort of Thunderbolt storage expansion would probably cover most of the Pro's use cases. And Apple probably doesn't want to deal with anything more exotic than those.
Can't you already stack 4 Studios? Building a $1,000 case to hold them makes some sort of Apple sense.
https://www.youtube.com/watch?v=x4_RsUxRjKU or something
I have three of the trash can ones. They are absolute pieces of art, as useless as they are computationally these days (energy-to-performance wise at least). I will never sell nor give them away.
Sad. I had this pipe dream of an Apple Silicon system made as a PCIe endpoint, so a Mac Pro could be a coordinator and host to like 4 of such systems in a cluster with very fast interconnect. Imagine the possibilities.
This honestly saddens me a little. From the PowerMac's to the MacPro I always loved them when having the opportunity to work with them. Plus I loved the expandability they offered.
I don't find the external GPU houses for Mac Studio as appealing to use.
The only thing that is missing from the current Mac line up is a one rack unit machine.
They had that with the previous gen mini, but the new ones are too tall now.
With the popularity of mac mini (and macbooks for that matter) for doing ML/AI work, I would have thought Apple could make a Mac Pro that could make for a good workstation for doing in-house ML/AI stuff.
I bought a GPU maybe a decade ago for this, and it's not worth the hassle (for me at least), but a nice out-of-the box solution, I would pay for.
The problem is that the M1 chips foretold the doom of the Mac Pro unless they could figure out some way to do something that you couldn't do with a Mac Studio - thunderbolt is so good that it's hard to justify anything else.
If they had done more with NUMA in the M series maybe you could have a Mac Pro with M5 Ultras that can take a number of M5 "daughter cards" that do something useful.
This makes sense, for that kind of money you could always build a beastly workstation in a real ATX case with standard components. Install Linux and the Mac looks like an expensive toy in comparison.
I kinda would have loved a new Mac Pro, same case, but just stick 4 Mac Studios in there and connect them all via MLX.
Would be a killer local AI setup...for $40k.
Honestly the Mac Studio is the new Mac Pro, this makes more sense to me.
but even that one looks kinda outdated when looking at latest M5 Max laptops.
Mac Studio waits for the Ultra chips to ship, which are always last in a generation. Perhaps the M5's chiplet architecture will help them move faster there.
It is kind of a bummer they never supported Cuda/HIP GPGPU using those slots.
For AI Apple internally has rack mounted M chips but they wont sell them to the public.
Reading comments, I don’t think people are being completely fair here. For Intel and AMD to approach what Apple has accomplished they’re making many of the same compromises with Panther Lake and Ryzen AI Max. Apple chose to put disk controllers on their SoP rather than having them on the storage module. This shaves a tiny bit of latency. Worth it? No idea. I’m shit at hardware design.
As for not having a Pro or otherwise expandable system? It’s shit. They make several variations of their chips, and I don’t think it would hurt them to make an SoP for a socket, put a giant cooling system in it, and give it 10 or 12 PCIe slots. As for what would go in those slots? Make this beast rack mountable and people would toss better network cards, sound/video output or capture, storage controllers, and all kinds of other things in there. A key here would be to not charge so much just because they can. Make the price reasonable.
They have tried variations of this since time immemorial (we can argue about "price reasonablé") but there's just not much you can do with it that you can't do much cheaper or simpler in other ways.
The Xserve has been dead for 15 years now, and it was never tremendously amazing (though it was nice kit).
Apple apparently has some sort of "in-house" xserve-like thing they don't sell; but turning that into a product would likely be more useful than a Mac Pro, unless they add NUMA or some other way of allowing an M5 to access racks and racks of DIMMs.
WSJ recently did a thing about it but details were rather light
The 2013 trash can was the end of the Mac Pro. It was never the same after that. The 2012 and earlier Mac Pros were awesome. I had a 2010 model. Here's what I loved:
• Multiple hard drive bays for easy swapping of disks, with a side panel that the user could open and close
• Expandable RAM
• Lots of ports, including audio
• The tower took up no desktop space
• It was relatively affordable, starting at $2500. Many software developers had one. (The 2019 and later Mac Pros were insanely expensive, starting at $6000.)
The Mac Studio is affordable, but it lacks those other features. It has more ports than other Macs but fewer in number and kind than the old Mac Pro, because the Mac Studio is a pointlessly small desktop instead of floor tower.
That's when they stopped designing computers for the pro market and started selling mid-century Danish furniture that can also edit videos.
I knew it was all over when third party companies had to develop the necessarily-awkward rack mount kits for those contraptions. If Apple actually cared about or understood their pro customers, they would have built a first party solution for their needs. Like sell an actual rack-mount computer again—the horror!
Instead, an editing suite got what looked like my bathroom wastebasket.
When it was introduced, Apple said the trash can was a revolution in cooling design.
Then they said they couldn't upgrade the components because of heat. Everyone knows that wasn't true.
By the time Apple said they had issues with it in 2017, AMD were offering 14nm GCN4 and 5 graphics (Polaris and Vega) compared to the 28nm GCN1 graphics in the FirePro range. Intel had moved from Ivy Bridge to Skylake for Xeons. And if they wanted to be really bold (doubtful, as the move to ARM was coming) then the 1st gen Epyc was on the market too.
Moore's Law didn't stop applying for 6 years. They had options and chose to abandon their flagship product (and most loyal customers) instead.
The biggest issue was actually that the Mac Pro was designed specifically for dual GPUs- in the era of SLI this made some sense, but once that technology was abandoned it was a technological dead-end.
If you take one apart you'll see why, it's not the case that you could have ever swapped around the components to make it dual-CPU instead; it really was "dual GPU or bust".
Somewhat ironically, in todays ML ecosystem, that architecture would probably do great. Though I doubt it could possibly do better than what the M-series is doing by itself using unified memory.
I'll admit that while I've used the trash can but never taken it apart myself. But I can't imagine it would have been impossible to throw 2x Polaris 10 GPUs on the daughterboards in place of the FirePros.
I think on a technical level you're right, but you need to run two of them and they'd need a custom design like so:
https://i.ebayimg.com/images/g/RQIAAOSwxKFoTHe3/s-l1200.jpg
For what is essentially a dead-end technology, I'm somewhat doubtful people would have bought it (since the second GPU is going to be idle and add to the cost massively).
the CPU being upgraded would have been much easier though I think.
Aside from the GPU mess, the 2013 was a nice machine, basically a proto-Mac Studio. Aside from software, the only thing that pushed me off my D300/64GB/12-core as an everyday desktop + front-end machine is the fact that there's no economically sensible way to get 4K video at 120 Hz given that an eGPU enclosure + a decent AMD GPU would cost as much as a Mac mini, so I'm slumming it in Windows for a few months until the smoke clears from the next Mac Studio announcement.
At which point I'll decide whether to replace my Mac Pro with a Mac Studio or a Linux workstation; honestly, I'm about 60/40 leaning towards Linux at this point, in which case I'd also buy a lower-end Mac, probably a MacBook Air.
I'm in the Linux desktop / Mac laptop camp, and it works well for me. Prevents me getting too tied up in any one ecosystem so that I can jump ship if Apple start releasing duds again.
The studio is also like 5x as fast as those machines.
What's your point? Of course processors have gotten a lot faster between 2012 and 2025.
I was talking about the form factor of the machine.
I’m much less concerned about the Mac Pro and more concerned we won’t see an XDR Ultra Display to replace the 32”.
They did come out with a new XDR. It’s just not 32”.
They also discontinued the 1000 dollars Pro Stand, apparently
Good, Mac Pro was fugly.
I like Apple when they make pretty stuff. Especially small, shiny, and quiet.
I guess A/V pros are used to getting screwed constantly, but it must be really irritating to face the prospect of eventually having to move PCI add-in cards to TB5 enclosures that cost $1000 per slot.
> Apple discontinues the Mac Pro
They replaced it with Mac Neo. Did you notice the wonderful build quality, the accesible price and that everyone is buying it ? And it has USB: U from universal.
In other news, the Mac Pro Wheels Kit was also discontinued.
https://www.macrumors.com/2026/03/26/mac-pro-wheels-kit-disc...
I mean for the cost of the Mac Pro wheels you can get a Macbook Neo these days!
The price to value for the Mac Pro kept going down. Mac Studio made more sense.
If I remember correctly, the maximum configuration was something like $35k back in the day. I wonder what those people feel like now. On the other hand, if they have $35k to burn, probably they don't even think about it.
If you spend $35k and just idle the machine or just check e-mail you've burnt the money. If it's your work machine and you've got a $100/hr billable rate it's paid for in a little over a month. Three months at a $50/hr rate.
If you bought the $35k Mac Pro in 2023 when it was released and have a $50/hr rate it's been paid off for about 30 months. So as of today those owners probably aren't too broken hearted. They'll likely get at least another three years out of them.
People buying $35k Mac Pros probably paid them off after a single contract. So they've just been making money rather than costing money.
Another company would have killed it a long time ago. It lasted this long because it’s Apple.
Now everyone that needs classical workstations can finally move on into Linux or Windows workloads.
Believe t-shirts at WWDC were not enough.
Thus the workstation market joins OS X Server.
For those who don't know what the t-shirt reference is, it's a creation by John Siracusa/The Accidental Tech Podcast: <https://cottonbureau.com/p/4RUVDA/shirt/mac-pro-believe-dark>.
And I still don't get it.
Siracusa—probably best known here for doing fabulous OS X reviews for Ars—is a co-host of ATP. He is also known is such circles for having Mac Pros, and using them for a long time (sometimes by choice, sometimes by circumstance). He thinks Apple should make a Mac Pro, not necessarily because it's a big seller, but because he thinks Apple should make a "best computer," much in the same way car companies might make a car that will never sell but pushes engineers, etc.
They made a shirt. It was fun.
Ages ago, when new Mac hardware came out, I'd amuse myself by putting together an "ultimate Mac workstation" in the configurator --- once upon a time, one could hit 6 figures pretty easily --- these days, well I panic bought a duplicate computer because I was worried a chipped/cracked display was going to make it unusable (turns out a screen protector has worked thus far).
I agree with the reasoning, and would like to see Apple continue to make aspirational hardware, but maybe the mainstream stuff is good enough?
> maybe the mainstream stuff is good enough?
Even Siracusa admits that - he's found it hard to articulate what a true "Mac Pro" would do that you can't do with other things.
Back in the heyday of the $100k Mac Pro you could certainly imagine it doing things that wouldn't be easily done by anything under $50k, and it would look good doing it.
Apple still sells a workstation-type machine: the Mac Studio.
No it isn't, it is a mini where you can add audio cards, which is basically the only extensions it has available.
Hardly workstation class.
It's certainly beefier than a Mini - 6 TB5 ports (which can drive 6 PCIe 5.0 x4 slots in an enclosure if you want), M3 Ultra, up to 256GB RAM.
It's not at all a workstation type machine. It's a Mac Mini with bigger SoCs and better cooling.
What is this, a workstation for ants?
It's a pizza box, for a 6" pizza.