Goodbye, Mac Pro

> The mystery is why Apple seemed to repeatedly come up with designs that were not what customers were asking for and why it couldn’t manage to do basic speed bump updates.

For the same reason they are unable to deliver a coherent language, or stable and working developer tools; a bunch of out-of-touch, overpaid, bored, assholes design features and products to satisfy their own boredom, rather than what the market needs, while a bunch of tin men sit on top, with zero taste and no understanding of the market or audience.

Crazy how we’re all supposed to just move on from expandability and repairability because we’re supposed to just be used to things being worse now. Feels like the same thing people will say about Apple’s software quality a few years from now.

I don’t understand how MSI, for instance, can throw out so many random tower motherboards over the course of a year in numbers that must, for some, expect to sell as many as the Mac Pro — if not fewer — but Apple can’t release one annually.

A tower is an engineering softball! Fans galore. No constraint from space! Throw extra liquid cooling around and see how hard it runs! You don’t even have to make anything for the slots; let others do it.

Just the Studio on a board with slots would work. How much engineering could that take?

Simply having the sports car in your lineup has to be worth the cost.

Such an own goal, but not just one, an own goal over and over, every game for years.

> The mystery is why Apple seemed to repeatedly come up with designs that were not what customers were asking for and why it couldn’t manage to do basic speed bump updates.

The same could be said about mouses. Apple hasn't been able to ship a decent mouse in the last 28 years.

The Mac Pro's last decade and a half is a story of misplaced paranoia, incompetence, and arrogance. There is no reason it had to go this way, but Apple continually committed own goals before finally putting customers out of their misery.

The trashcan Mac Pro on release was obviously the wrong product for workstation customers. Say what you will about bad bets on dual GPUs, but even if it just ran a single GPU, it was a bad idea. Workstation customers just want a big box. Apple took a tool and made a consumer product out of it. The marketing, starting from "Can't innovate anymore..." at the reveal was coming from an executive team that saw the media crowning Samsung the forever king of mobile with a washed post-Jobs Apple. They were clearly taking it personally to an almost paranoid degree. I remember seeing pre-roll ads in movie theaters for the thing which is how off target the marketing was. They were treating it like an iPhone!

When they finally seemed to collectively realize years too late that they had made a bit of a gigantic blunder with this thing, they did the roundtable. The sensible, smart thing would be to release exactly the product that workstation customers wanted with no BS. Hell, even reuse the cheese grater design if you still have the molds and want to save money. But because modern Apple is at times overly concerned with things "Only Apple" can make, they took way too long to release a too expensive and overly engineered workstation and display. You release the Mac Pro and a standalone 5K display in early to mid 2018 - and in this instance where they are simply putting the 5K iMac display in a new enclosure plus making a Xeon desktop Mac, such a thing was absolutely doable in that timeframe - maybe the Pro at least has a bit more time in the sun. Maybe it can even assemble a decent customer base before Apple Silicon that can convince higher ups it should stick around! Instead, it withered on the vine again with a "We Tried" update that pleased almost nobody. Then they killed it.

The 2006 era Cheesegrater Mac Pro was the last good Mac Pro. I miss it. That was also the last era of "good" macOS releases (before Apple became The iPhone Company).

Everything after the cheesegrater was a downgrade.

The Trash Can was a dumpy design exercise with no strategy or understanding of market. Heating issues and other problems galore. It was like the defective butterfly keyboard of Macs.

The Wheel Mac Pro was pure hubris. A product years too late designed in a vacuum from the Apple Silicon team. All this "design" for almost no available components, but Apple will sell you the wheels separately! Apple milking the little pro market it had left.

I'll never understand how Siracusa fell for the trap. The 2006 era was good hardware, the enclosure made it easy to expand, and it had Bootcamp to run Windows with no penalty. Every subsequent model fell off hard in some aspect and was so infrequently updated you were buying overpriced obsolete hardware most of the time.

This marks the end of an era. Apple now only sells disposable computers you can't repair or expand. Apple applied the iPhone Enshittification model to Desktop Computing.

It's amazing the cylinder and the updated cheesegrater were like Thunderbolt monsters and PCIe monsters. But they were like 3 years too late!

@Leo - Talking about taste… You know the funny thing is that they made the Studio with a completely uninteresting design, but went hog wild on the Mac Pro with some avant garde shit. Totally opposite of what made sense.

The Studio would have been the perfect excuse to do one of their special designs. Could have been a perfect cube, like the G4 Cube. Or it could have been a downsized mini version of the 2019 Mac Pro enclosure, with the fancy 3D grate. Instead it’s like 2.5 Mac minis stacked up. Just no fanfare in the thing.

———

As for the 2013 Mac Pro, it’s actually a great machine. I have one in my Mac collection, and I admire it all the time. At the time, I derided it, and I apologize.

However it was incredibly stupid on the part of Apple to market that as a replacement for the tower. Had they called the 2013 Mac Pro the “Mac Studio,” it would have made more sense. They could have left a sole tower model in the old enclosure and just updated the Xeon chip once a year along side it, and no one would have said a thing.

It's curious to blame future CEO John Ternus because wasn't it him who introduced the 2019 Mac Pro?

Tower computers are easier to repair, easier to upgrade. Apple has no interest in making these kind of products. Buy that Mac and if you need an upgrade, well buy another one. Isn’t that really the reason they slowly killed this thing?

The 2013 Mac Pro was the same mistake made with the G4 Cube - a belief that people place a value on compactness in desktop computing.

Like the G4 Cube, it cooked its internal components, because Apple is actually not good at thermal design.

Fundamentally, what Apple is about is ensuring that computers can't be upgraded, no more decade-lasting workstations; that is the company's entire trajectory. Every Mac is an iPad now. Disposable, you buy it in its only ever configuration, all the money goes to Apple, and if you need more than you have, be it memory, or storage, or ability to support a certain number of displays, you eWaste the existing machine, and buy another.

You also never get comfortable with a software setup, because the forced upgrade train is coming, you'll be forever trying to walk on shifting sands, until the Grand Strategy is revealed, and everything, is just a subscription service, hardware, software, and content.

There is one thing I'm definitely sure it's time to move on from, and that's providing financial support to anything Snell is involved in.

It’s clear that a segment of the market does value compactness. Mini ITX, STX, etc. None of that, however, serves as top performance machines. Nobody builds Threadripper, Epyc proper or Xeon machines in tiny hard to reach and operate cases, 1000$ self-proclaimed “piece of art” tower cases.

Like I said, a bunch of highfalutin, out of touch assholes designing impractical shit, rather than answer what the market has been asking for decades.

The ARM Mac “Pro” is the biggest joke of all. A phone CPU with soldered memory and SSD, and no GPU support, in a time where GPU is king. Out. Of. Touch.

I'm not a hardware person, my thinking is that it would not be able to get the current level of performance from something that is not a SOC like the AS line of chips - and as a result a lack of upgradability.

I do miss the time of cards like those from Sonnet, though they offered a relatively increase in performance.

@CowMonkey: This is really out of touch. If you are including thermal performance/effeciency sure AS beats the pants off of everyone else on a per watt basis (though Intel is working to close that gap and maintaining compatability with the majority of software), but if you just care about pure performance, AMD and Intel both clobbers the AS chips and NVIDIA clobbers AS on GPU performance.

And that is just talking about single chips. If you need more CPU or GPU on Windows/Linux machines you can get multi-CPU motherboards and run multiple GPUs. Of course that all generates enourmous amounts of heat and consumes massive amounts of power, but if you need it, you can have it.

https://www.cpubenchmark.net/multithread/

>There is one thing I'm definitely sure it's time to move on from, and that's providing financial support to anything Snell is involved in.

Okay, you gonna have to explain this.

"it would not be able to get the current level of performance from something that is not a SOC like the AS line of chips"

This is a matter of priorities, not technology. CAMM exists, Apple just chooses not to use it. And that's ignoring what gildarts said, which is that Apple's chips are extremely fast compared to similar systems on the Intel/AMD/Nvidia side, but not compared to ultra-high-end professional options.

Apple could provide a high-end, upgradeable, super-fast tower system. The fact that they don't is a choice.

This is a sign the Mac is on its way out. Time to move on. Time to make Linux better than Apple ever made MacOS

@Kristoffer

> Okay, you gonna have to explain this.

If Snell thinks it's "time to move on", as in "nothing to see here folks, just accept it, Apple knows best", then I'm thinking it's time to not buy from any advertiser on Six Colours, or pay for memberships on The Incomparable.

That's the only power we have as consumers, to cease putting air in the slowly deflating life raft of Apple-specific media people.

Unless I'm wildly mistaken, and his point was it's time to move on from Apple?

@Kristoffer (I thought I replied to this, but it seems to be missing - I may have forgotten to submit, and I can't remember exactly what I wrote)

> Okay, you gonna have to explain this.

"Jason Snell: RIP to a real one, but it’s time for us all to move on."

This is real insidious "smoothing the way", where an influencer engineers a group position that it's reasonable to accept a curtailing of the things you want. The invader always has a collaborator to smooth the way. When bad things happen, there are always "reasonable people" with "reasonable" voices; you can't do anything about it, it's fait accompli, you may as well just get used to the new reality and you can become happier sooner.

Well no, if I have to move on from something, it will be moving on from supporting Snell's ability to draw a paycheck as a member of the Apple propaganda community.

It's sad, but also, TBH I can't see how it could have been continued. Slots for cards are of no use without software support, and it's been a long time since Apple cared about optical drives or HDDs. (Even with external storage, the RAID support is shot; they could at least fix that!)

So really, what would be the point of an AS Mac Pro?

The machine was never for me, but I kind of wonder which of no GPU or RAM upgrade upset its customers more?

@Paul McGrane

> I kind of wonder which of no GPU or RAM upgrade upset its customers more?

It's both; the fundamental problem was Apple trying to sell a machine which cost just as much, if not more than an equivalent workstation from Lenovo etc (which could be configured MUCH more highly), but which had the post-purchase reconfigurability of an iPad, when reconfigurability is a primary duty of a machine of this type.

Add to that, the AS 2023 machine came out with Sonoma, whose entire release cycle is marked by broken PCI NVME SSD drivers (because iPads don't have PCI slots), that cause drives to spontaneously unmount under load, and are unable to be seen by the system until a full plug-out power cycle.

Which is problematic when the primary use case for that machine was NVME PCI storage.

I feel like Apple fell behind in ML in the 2010s because their flagship trashcan Mac only supported some external GPUs but was baked into a "thermal corner".

By 2019, they came around and did a new cheese grater MBP that supported graphics cards with PCI expansion because it was still Intel.

And then they switched to custom silicon and dropped custom GPU support AGAIN, and now they’re  behind in AI/ML. I know it’s not directly related, but like how could they think it’s wise to not sell a machine where you can develop LLMs on it hardware?

"So really, what would be the point of an AS Mac Pro?"

Apple has the choice to make a configurable professional computer. Obviously, if they just put an iPad into a tower case, there's no point, but that's not what anyone is asking for.

@Plume Well, yes, OK, so what's the plan? Perhaps they could choose to maintain Intel support for such a hypothetical machine? Both Intel and AMD will outperform in CPU, and Nvidia in GPU, at energy cost. So maybe that's the answer?

@Julian

> I feel like Apple fell behind in ML in the 2010s because their flagship trashcan Mac only supported some external GPUs but was baked into a "thermal corner".

It didn't support *any* external GPUs. External graphics was a Thunderbolt 3 feature, and the 2012 was a TB2 machine. It only ever had the one generation of GPUs, which it treated as a consumable, like printer toner.

"Perhaps they could choose to maintain Intel support for such a hypothetical machine?"

Yes. they could keep Intel support. But also, nothing stops Apple Silicon Macs from being expandable, except Apple's decisions. For example, fricken Raspberry Pis can use external GPUs. Why in the world can't Apple's devices?

It always felt Apple's hand was forced with the 2019 Mac Pro.

They had already launched the iMac Pro a few years earlier, and it seemed like this was their preferred solution to supporting "pro" customers going forward.

But online chatter and commentary insisted a big-box-machine was vital and so they blinked and suddenly there were sit-down sessions promising a Mac Pro, and then barely a year later the cheese-grater was unveiled.

To be almost immediately rendered redundant by the M1 chip.

A similar thing happened with the trashcan. Apple was telling everyone this was the beset machine to do deep learning on, thanks to the OpenCL standard they and AMD/ATI were committed to as an open alternative to Nvidia's CUDA.

Except documents released in one of their trials revealed that just as they were promoting OpenCL, they were already well along to implementing a competing, proprietary technology themselves: Metal.

Metal made sense for phones -- just as converting the phone chips into M1 chips did -- but it meant they weren't investing any money into any technology that could compete with CUDA, so they ceased to be a seller of computers for deep-learning professionals.

@Plume

> nothing stops Apple Silicon Macs from being expandable, except Apple's decisions.

Yup, System 76 has an ARM workstation, with standard *ECC* chipped ram, standard PCI graphics, and standard storage. Apple could easily offer that sort of machine, but it would derail the gravytrain by showing the things Apple says are great advantages of "unified", are not actually desired by the target market, given the flexibility cost they bring.

https://system76.com/desktops/thelio-astra-a1.1-n1/configure

@Someone, @Plume

I generally think that unified memory is a good thing and not worth giving up. I was playing around on a PC recently and the limited VRAM on the dedicated GPU was terribly constraining, sure it can swap with main memory but it isn't pleasant or performant. The problem on the PC side is the chips with big unified memory (AMD Strix Halo 128GB, NVIDIA DX) have relatively limited memory bandwidth. Apple at least is using pretty fast unified memory (614 GB/s) not as fast as dedicated GPUs but faster than big unified memory pools on the PC side.

To my mind "super-chips" and SoC like architectures make far more sense than discrete GPUs and CPUs. I also think there is no real technical reason why you couldn't build a CPU-IO-GPU three or more chip architecture/module with a large pool of fast upgradable memory. There is no reason something like LPCAMM2 modules couldn't be designed to scale to 200+ GB/s per module and then 2-8 channel memory achieving GPU like bandwidth.

It might not be feasible to separate the GPU, CPU, and IO chips onto separate sockets and maintain good latency but even that I think should be possible, put them close enough together and you could in theory have 3 upgradable components.

On the PC side I hope someone at least starts to think about going this route, on the Mac side the main reason I doubt Apple will ever go back to upgradable memory and storage is the cynical cost reason. If we can buy storage or memory from anyone what reason do we have to pay Apple's outrageous prices?

You can have unified memory *and* replaceable memory. You can have a system that has unified memory *and* supports an external GPU. All of these things are decisions Apple made, not intrinsic limits Apple is bound by. These decisions make sense for a MacBook Neo, they just don't make sense for a professional tower computer.

The point is that there is absolutely nothing preventing Apple from making a professional computer that is fast *and* upgradeable. They just decided that this is not a market they care about. It's a decision, not a constraint.

@Someone

>This is real insidious "smoothing the way", where an influencer engineers a group position that it's reasonable to accept a curtailing of the things you want. The invader always has a collaborator to smooth the way. When bad things happen, there are always "reasonable people" with "reasonable" voices; you can't do anything about it, it's fait accompli, you may as well just get used to the new reality and you can become happier sooner.

So true.

@ You can have unified memory *and* replaceable memory. You can have a system that has unified memory *and* supports an external GPU

It's interesting how many people have fallen for Apple's marketing and don't know this. In the 90s there were Macs with soldered *and* replaceable memory. I really hoped Apple would go that route, but they really are just the iPhone company now. Or that's their goal.

Really sad to see the Mac wither away one piece at a time when it was such an amazing tool, designed with as much care as you'd want to design the things you're making with the tool. It was a great inspiration, but that company doesn't exist anymore

As usual, I'm an outlier here. :-)

I had a 'trash can' Mac Pro in my last job. I ran it hard for 5+ years until I left that job. I didn't do anything that would ask much of a GPU - just the occasional reprocessing of a video from Podcast Producer. But I ran a lot of database work with it, some VMs, and a lot of software builds. No complaints.

My current Mac at work is an M1 Max Mac Studio, with about the same kinds of workloads (minus Podcast Producer, of course). I probably could have talked my employer into a Mac Pro, but for what I do with it, I really couldn't justify the price difference. I have a 27" iMac in my home office. Being an Intel model, it's obsolete now, so I'll replace it with a Mac Studio when the new models come out this year.

IMO, I don't think Snell is trying to act as a collaborator to lay groundwork for Apple. It struck me more as realism: Here's where we are. And, for most people, that's not such a bad place to be.

A new Mac Pro that you all are wishing for (but notably, not buying) would play to Apple’s weakness.

Why would they do that?

Why invest all this money and build capacity for amazing SoCs that do what no one else can… and then hack them apart to sell in generic pieces and compete in the commodity parts business. That’s stupid.

So you can upgrade slow RAM? Go get a PC then. Jeez.

Mmm. Perhaps I am thinking of Apple's partnership with GPU vendors. They would need them for driver support.

But, sure. Technically possible, even with an integrated SoC, to *add* GPU support to the OS and hardware. That also benefits eGPUs connected over Thunderbolt.

@Bryan

>They had already launched the iMac Pro a few years earlier, and it seemed like this was their preferred solution to supporting "pro" customers going forward.
>
>But online chatter and commentary insisted a big-box-machine was vital and so they blinked and suddenly there were sit-down sessions promising a Mac Pro, and then barely a year later the cheese-grater was unveiled.
>
>To be almost immediately rendered redundant by the M1 chip.

I think this was entirely Apple’s fault. When they had the roundtable in 2017 with their hand-selected members of the press to share their vision for a Mac Pro (a project that hasn’t started yet, and wouldn’t ship for 2.5 years), that was them trying to do damage control and address the overall negative reputation Apple was getting on laptops and desktops. The iMac Pro was a stop-gap that was chosen to try to stem the bleeding from the lack of trashcan Mac Pro sales and it was a product they were equally uncommitted to.

The timing for the 2019 Mac Pro *was* inauspicious. Internally, Apple knows they are about to roll out the new chip, meanwhile, they’ve spent all this time on this very niche product.

And that sucks.

But the fault goes back to Apple, who had to know by mid-2014 that the trash can Pro was a dud. There were SIX years between the trash can Pro and the 2019 Pro. Six.

Apple should have either stuck to its guns and killed the Pro back then, or they should’ve iterated more quickly. I don’t fault Apple for putting the redesign out 6 months before the Apple silicon announcement — sometimes shit just happens. I do fault them for waiting 3.5 years to acknowledge that there was a problem in the Pro space to begin with.

But while I can defend the 2019 Mac Pro, I can’t defend the 2023 Pro on any level. Not only did it not offer any upgradability (aside from some PCI-e slots that were going to be useful to such a small number of people, most of whom wouldn’t be spending five figures on a Mac Pro anyway), it was so ridiculously priced not just against the Studio but every comparable AMD or Intel workstation (where you could pair it with a much better GPU) might not have the energy efficiency, but that’s besides the point.

I really think the 2023 Pro was Apple trying to find the the limit of how much it could sell a computer for and yeah, we found it.

> A new Mac Pro that you all are wishing for (but notably, not buying)

I found it amusing when you claimed that Apple needed to scam its customers in order to ensure that their privacy was preserved, but complaining that people don't buy a product that literally does not exist is even funnier.

@ChristinaWarren

The only quibble I have with your timeline is this bit:

> The iMac Pro was a stop-gap that was chosen to try to stem the bleeding from the lack of trashcan Mac Pro sales and it was a product they were equally uncommitted to.

I don't think you'll find many people within the Mac Pro community who believe the official narrative that the iMac Pro was *made as a stopgap*, after the 2019 Mac Pro had been decided as a strategy.

The strategy, and timeline is far more likely to be that the iMac Pro was entirely conceived as a wholesale replacement for the Mac Pro as a product line. Had the pro community not been specific in its critique of the weak, and non-upgradabe graphics of the 2013 Mac Pro, the iMac Pro would have been the only high end desktop from Apple.

The product matrix would have been laptops, Mac Mini, iMac, iMac Pro.

When Apple lost control of the narrative, and even the rented mouths like Gruber were speaking out to how bad the situation was, Apple realised their already planned, likely completed and released to manufacturing product; the iMac Pro was going to be a disaster in that atmosphere, so they decided on announcing a new tower, and then offered up the iMac Pro as an "interim" product.

It was such a joke of a product that when Apple used it as their headline VR-capable computer, one of the few apps demoed at the product launch that actually saw release (Gravity Sketch), didn't support the Vega56 variant out of performance concerns, until some time later, presumably after pressure from Apple.

Something else to remember: Even though the Apple Silicon systems started coming out the next year, the 2019 Mac Pro was still fully supported. So if you were one of the (relatively few) buyers of that system in 2019 (or later), it didn't suddenly stop working when the AS systems started coming out. You can even run Tahoe on it, as long as you don't need Apple Intelligence.

And when macOS 27 comes out, the 2019 Mac Pro will *still* be able to run Tahoe (or anything all the way back to... Catalina?). So whatever someone bought a 2019 Mac Pro for back then, there's a good chance they can still do the same things on it today. If you want to say that it will now be obsolete because it won't run 27, that's still 7 years of use. That's the same that I'm getting out of my 2019 27" iMac, and I'm not really sad about that.

We have a couple of Windows XP systems here that are connected to expensive instruments (scientific, not musical). We keep them around because they still server their purpose, and upgrading them to something later would require replacing the $150K instruments. We just don't allow them on the network. :-)

@ChristinaWarren

> But the fault goes back to Apple, who had to know by mid-2014 that the trash can Pro was a dud. There were SIX years between the trash can Pro and the 2019 Pro. Six.

Interesting how that timeline aligns with that of Project Titan. They had bigger things on their minds.

@ Someone

> When Apple lost control of the narrative, and even the rented mouths like Gruber were speaking out to how bad the situation was, Apple realised their already planned, likely completed and released to manufacturing product; the iMac Pro was going to be a disaster in that atmosphere, so they decided on announcing a new tower, and then offered up the iMac Pro as an "interim" product.

I think this is all fair — and yes, the tooling and design work on the iMac Pro was almost certainly done before that April 2017 briefing (which is why I will always believe that work on the 2019 Mac Pro hadn’t started in earnest until after that roundtable). But whether the iMac Pro was originally seen as the replacement for the Pro desktop and then positioned as the stopgap or quickly thrown together (let’s say in 2016) as a stopgap until Apple could reassess market needs, the end result is still Apple very clearly telegraphing to its Mac Pro user base (a user-base already angered by the fact that Adobe software now could run better and cheaper on PCs — and this was after Apple had already fumbled Final Cut Pro and sent that audience over to Premiere) that an interim solution is coming, but don’t worry, we’ve got a great Pro machine you’re going to love.

And credit where credit is due: Apple did deliver. Now whether it should’ve taken 2.5 years to get that product to market, I can’t say. To @unsignedchar’s point, it’s possible Apple’s hardware engineers could’ve been resource constrained working on the ill-fated car project.

The problem, of course, was it delivered a product that was going to be abandoned almost the second it was released (although I guess Apple and AMD kept up GPU rebadges for a few years, although that stopped by the time the 7000 series AMD GPUs were out).

I still think Apple could’ve done something better with the Apple silicon Mac Pro. As Sherief Farouk said, Apple could’ve absolutely supported external graphics cards if it wanted to. I have to assume Apple didn’t want to point out the gulf in raw capabilities between its integrated GPUs and the stuff available from AMD and Nvidia. Ironically, the unified memory has made Mac Studio really strong inference machines, but Nvidia (and to a much lesser extent, AMD) are faster and have much better 3D performance, and historically, Mac Pro owners are doing graphical work, not AI engineering.

I think most of us could’ve forgiven the lack of upgradable memory. What was impossible to really square was the lack of any other upgradability aside from additional storage. So you paid thousands more for a few PCI-e lanes that while faster than TB4 (and even TB5), aren’t going to be utilized by anyone except for the users with the most minute edge cases.

And that’s what makes the final death of the Mac Pro so sad: it ultimately became a machine that was useful for almost no one, whereas it was historically a machine that few needed, but many aspired to owning.

Only a few of us are crazy enough to really make use of more than one or two 16x PCIe slots. Personally, I'd love to have a motherboard like that for use in my server rack, but I'd absolutely not run macOS on it.

Honestly, I'm super impressed Marco was able to get a mini-fleet of Mac Minis working well in a server rack situation. Don't know if it would have been cheaper to deploy a few multinode linux servers with GPU cards and whisper.c++, but it certainly would have been easier.

@ChristinaWarren

For me, working right now on a 2019 Mac Pro, there's nothing in terms of upgradable that I would forgive. The moment stuff's soldered, it has to be priced as disposable. iPad Pro money is the most I would spend on an Apple Silicon Mac (and my iPad Pro is just about the most disappointing Apple product I've ever owned, due to its garbagefire OS).

I honestly don't believe anyone could tell the difference in use in a blind test between SOC Vs traditional architecture, provided the software was written competently for both.

People crowing about how the M2 Ultra was like having 120GB of memory available to a GPU... well no, because the memory bandwidth meant in practice it only had 32GB *usable* memory, and the GPU itself was *radically* lower performance than the AMD options at the time.

@gildarts

> Only a few of us are crazy enough to really make use of more than one or two 16x PCIe slots.

I have 2x W5700X GPUs on x16 MPX slots, and a 4x4TB NVME SSD on my remaining x16 slot (that's all my x16s used), Afterburner (boy that was was a kick in the guts to customers) relegated to x8 slot, and I go over 100% lane allocation by plugging a USB drive into a thunderbolt port. Still, I can drive 12 x 4k displays, or 6 x 6k displays.

For all Thunderbolt has progressed, it's still just 4 lanes of PCI 4, whereas PCI 5 x16 is normal for motherboards - so 8x the bandwidth per slot, compared to each TB5 *bus*.

And PCI 6 announced a few years back doubles that bandwidth again.

I'd be fine with a GPU on die, *if* I could still plug in an eGPU to run displays. That's the real annoyance to me, that I could have to replace my entire computer, because of a limit in any one component - but of course, that's the intended business model.