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

[Charlie Bonesx]

In day to day usage, the modern Mac is less capable than it used to be, because modern software has gone so far down the "simple is easier" fallacy, that it's past the inflection point of "less is less".

That’s an entirely false premise.

What could you do in macOS five or ten years ago that you can’t do today?

suspect KDE, for example, can get good at being user-friendly and covering more people's use cases on existing "obsolete" hardware faster than Apple can become anti-fragile to circumstances.

First of all, as you know, is just a desktop environment and some apps, not an operating system

Even so they’ve only been at it 30 years. I’m sure they’ll “get good” at being user friendly any day now

 

[leman]

Man I feel you. My next Macs will definitely be more MP 2019s. Mostly rack mount.

Didn’t you say you need performance? What are you collecting this obsolete hardware for?

 

[leman]

IMO, Apple hasn't been truly interested in Pro level computing since the Mac Pro 4,1/5,1. That was the real end. They were in the game with the G3, G4, G5 Powermacs and then up to the MP5,1. That was really the inflection point where iPhones, iPads and to a lesser extent Mac laptops became their focus. The 6,1 was Apple just taking everything on the inside of a midtower and putting it on the outside, proprietary GPUs, to push...Thunderbolt? The 7,1...they were forced into it, kicking and screaming. The engineers, God Bless them, hit it out of the park on that system (I don't think there has been a better designed and put together system since or likely even before) and it was the last day in the sun for the Mac Pros.

High end pro systems? They have abandoned that space to Dell Precisions, HP Zs, Lenovo ThinkCentres all of which can be expanded to incredible levels. Heck, you can buy an old dual CPU Cascade Lake workstation today, throw in tons of DDR4, a few cheap Xeon Golds/Platinums and a few high end GPUs and do a good number of things even maxxed M5 Macs would struggle with. Go modern Xeon/Threadripper and well, it isn't pretty at all for Apple Silicon. (lack of high end GPU support and huge RAM count is what kills them in this comparison).

This is true. At the same time, the computing landscape has changed significantly since these early days. High-end desktops were indispensable for running certain workloads. We had a bunch of Mac Pros to run probabilistic simulations and data aggregation pipelines. These days ,however, my M5 Air fits a complex Stan model faster than the university supercomputer...

With moving to in-house chips, Apple essentially shifted the compute landscape. They give you a very strong baseline, and more than adequate prosumer options, at the cost of the ultimate high-end performance. Still, the Max series offer the same throughput as the 16-core Threadrippers, while being significantly better for general productivity and mixed usage. I'd wager that covers quite a large chunk of use cases out there.

 

[theluggage]

This is true. At the same time, the computing landscape has changed significantly since these early days. High-end desktops were indispensable for running certain workloads. We had a bunch of Mac Pros to run probabilistic simulations and data aggregation pipelines. These days ,however, my M5 Air fits a complex Stan model faster than the university supercomputer...

...and even if it didn't you now have a decent broadband link to the supercomputer - more than capable of sending back gigabytes of results and high-def rendered graphics - while the supercomputer can be sitting on a really fast network backbone and a nice fat internet pipe with access to the source data.

That's why the market for high-end, monolithic personal workstations is being eaten away at both ends, and isn't going to attract much investment from the likes of Apple.

 

[BSDnostalgia]

Nobody running towers of this caliber cares about bang per watt. We can afford the electricity. And our power crushes AS. Come to grips with that.

I wouldn't say "crushes". A little slower single-threaded, a little faster total throughput. But, the M3 Ultra has a lot more main memory bandwidth.

Man I feel you. My next Macs will definitely be more MP 2019s. Mostly rack mount.

The maxed 2019 was a speedy machine. With the Xeon W-3575M? A very hot CPU in both senses.

Compared to the maxed M3 Ultra (28 cores), you have a machine with the 3575 single-threaded about 73%, but, total throughput/FP64 123%-129% on 44/88 cores.

And, you did say you didn't care about watts. Because the CPU alone is 340-408W. The M3 Ultra SOC is what, about 80W, including GPUs?

Memory bandwidth is, I believe, about 300 GB/sec for the Xeon, and, over 800 GB/sec for the M3 Ultra.

So, overall, somewhat comparable, actually, with different strengths for each. Except power consumption, where the Xeon has a huge disadvantage, unless you live in Fairbanks, AK, and need to heat your house.

You haven't stated what your actual applications are, so, it is difficult to be any more precise than the above. These were over $13K starting price with 32 GB. With 1.5 TB, $38K.

The rackmount server configuration is a definite plus.

 

[travsG5]

This is true. At the same time, the computing landscape has changed significantly since these early days. High-end desktops were indispensable for running certain workloads. We had a bunch of Mac Pros to run probabilistic simulations and data aggregation pipelines. These days ,however, my M5 Air fits a complex Stan model faster than the university supercomputer...

With moving to in-house chips, Apple essentially shifted the compute landscape. They give you a very strong baseline, and more than adequate prosumer options, at the cost of the ultimate high-end performance. Still, the Max series offer the same throughput as the 16-core Threadrippers, while being significantly better for general productivity and mixed usage. I'd wager that covers quite a large chunk of use cases out there.

I can't disagree with anything you say here, you are 100% correct. Apple Silicon is very impressive and yep, covers most user's needs, probably even most pros. Like I said, I don't necessarily blame Apple, the computing landscape has indeed changed. We all miss expandable storage, high end upgradeable GPUs and RAM though. Probably a lot of why we are in this thread 🙂

 

[jinnyman]

And, you did say you didn't care about watts. Because the CPU alone is 340-408W. The M3 Ultra SOC is what, about 80W, including GPUs?

I'm just passing by, but M3 Ultra is no way only 80W. Isn't it like almost 400W? Heck. My M1 MAX is higher than that (although unable to sustain)

 

[teh_hunterer]

They give you a very strong baseline, and more than adequate prosumer options, at the cost of the ultimate high-end performance.

I suspect this cost to prosumer options may even be temporary anyway. Wouldn't it have been mostly yield issues making a bigger chip not financially viable to build?

With their fusion architecture allowing them to make bigger chips now for less cost, I think we'll eventually see them build an "extreme" chip. But of course the rampocalypse might delay this for some time.

 

[BSDnostalgia]

I'm just passing by, but M3 Ultra is no way only 80W. Isn't it like almost 400W?

I wasn't sure where to look up how much the SOC uses? Anybody know? That is, I found it somewhere, but, it may not be reliable. The M3 Ultra in total would include the SSDs as well as port-related stuff. Trying to compare apples and Apples. In the big Mac Pro, you include the GPUs and all the other motherboard stuff, and storage How much is that total? 1200 Watts?

 

[teh_hunterer]

I wasn't sure where to look up how much the SOC uses? Anybody know? That is, I found it somewhere, but, it may not be reliable. The M3 Ultra in total would include the SSDs as well as port-related stuff. Trying to compare apples and Apples. In the big Mac Pro, you include the GPUs and all the other motherboard stuff, and storage How much is that total? 1200 Watts?

Easiest is to just Google the question and expand the AI answer which dives into the details. 30ish watts idle, 80w heavy CPU use, 300w heavy CPU/GPU use.

 

[leman]

I suspect this cost to prosumer options may even be temporary anyway. Wouldn't it have been mostly yield issues making a bigger chip not financially viable to build?

With their fusion architecture allowing them to make bigger chips now for less cost, I think we'll eventually see them build an "extreme" chip. But of course the rampocalypse might delay this for some time.

I doubt that die stacking will change the equation in any meaningful way. Building large scalable systems still requires a lot of resources and unique technical solutions (you need to move a lot of data around the dies, for example), and so far we have plenty of evidence that Apple does not consider it worthwhile to pursue this.

 

[Flint Ironstag]

I wouldn't say "crushes". A little slower single-threaded, a little faster total throughput. But, the M3 Ultra has a lot more main memory bandwidth.

You're 100% right. They're all fast, and anyone can cherry pick benchmarks. I favor the brute force and ultimate flexibility of massive GPU power.

Somebody again please tell me the maximum FP32 of AS. Go ahead and cluster 4 of them.

 

[Flint Ironstag]

Didn’t you say you need performance? What are you collecting this obsolete hardware for?

@leman Yes. I need performance. Please tell me how much FP32 the best AS has. Go ahead and cluster them. We're all waiting for your answer. Come back and tell us in public. We are waiting.

My hardware is not obsolete, it's fully supported by Apple as of this posting. AS cannot even come close yet in raw compute, which is the performance I'm after.

Please, please come back and tell me how much FP32 you have. And I'm dying to know how much FP64. 🤣

 

[leman]

@leman Yes. I need performance. Please tell me how much FP32 the best AS has. Go ahead and cluster them. We're all waiting for your answer. Come back and tell us in public. We are waiting.

My hardware is not obsolete, it's fully supported by Apple as of this posting. AS cannot even come close yet in raw compute, which is the performance I'm after.

Please, please come back and tell me how much FP32 you have. And I'm dying to know how much FP64. 🤣

FP32 and FP64 in which context? What applications are you talking about? I remember you mentioning something about hash cracking — what does that have to do with floating point?

 

[teh_hunterer]

@leman Yes. I need performance.

You still haven't explained what you need this performance to do. I don't think you do need it. I don't think you do anything in the real world with your computer except game, set up old systems, and post on the internet about it.

Please tell me how much FP32 the best AS has. Go ahead and cluster them. We're all waiting for your answer. Come back and tell us in public. We are waiting.

For someone who is openly dodging inconvenient questions, the lack of self awareness and pure temerity to say "we're all waiting for your answer" to someone else is staggering. This is truly getting into 🤡 territory.

 

[teh_hunterer]

I doubt that die stacking will change the equation in any meaningful way. Building large scalable systems still requires a lot of resources and unique technical solutions (you need to move a lot of data around the dies, for example), and so far we have plenty of evidence that Apple does not consider it worthwhile to pursue this.

I thought it wasn't die stacking, but rather them being able to make a bunch of smaller chips and then connect them to each other through an interposer that was essentially fast enough for them to function as if they were all on the same chip. Hence bypassing the yield issues of having to make a massive perfect chip - you can combine a bunch of CPU and GPU chips with no defects into one massive chip.

It's not exactly my wheelhouse, but the way it was explained to me, it seemed that this would allow Apple to make the bigger chips that they were rumoured to be looking at making but ultimately decided not to because it was too hard and expensive.

Not only that, but it would probably allow them to update these chips faster than they do now (where the ultra chips are a generation behind).

But maybe there are other factors at play as you say.

 

[Flint Ironstag]

You still haven't explained what you need this performance to do. I don't think you do need it. I don't think you do anything in the real world with your computer except game, set up old systems, and post on the internet about it.



For someone who is openly dodging inconvenient questions, the lack of self awareness and pure temerity to say "we're all waiting for your answer" to someone else is staggering. This is truly getting into 🤡 territory.

I don't have to explain diddly squat to you, and it doesn't matter what you think I need or not. It's great! 🤣👍🏿

 

[Flint Ironstag]

@teh_hunterer how much FP32 do you have? How much FP64?

Please tell us the specs of the Mac Pro you own - I didn't see it in your signature. @leman you too please. Context would be helpful.

We're all waiting to see your response.

Otherwise...

 

[teh_hunterer]

I don't have to explain diddly squat to you, and it doesn't matter what you think I need or not. It's great! 🤣👍🏿

Actually you do have to explain it - that is if you want to keep posing as someone who is doing all this complicated pro level work that we're all too ignorant to understand. If you want to keep projecting this identity, you do in fact have to back it up and explain when you keep saying you use all this FP32 and FP64 and hash cracking stuff, what are you actually using it for?

You're throwing out buzzwords in the hopes that we'll all accept that you are just too advanced for the rest of us.

I can't force you to explain yourself - but I can damn well sit here and appreciate that everyone can see through what you're trying to do.

Context would be helpful.

Indeed 🤣

 

[blackadde]

@teh_hunterer how much FP32 do you have? How much FP64?

Please tell us the specs of the Mac Pro you own - I didn't see it in your signature. @leman you too please. Context would be helpful.

We're all waiting to see your response.

Otherwise...

Otherwise what?

 

[mcnallym]

Actually you do have to explain it - that is if you want to keep posing as someone who is doing all this complicated pro level work that we're all too ignorant to understand. If you want to keep projecting this identity, you do in fact have to back it up and explain when you keep saying you use all this FP32 and FP64 and hash cracking stuff, what are you actually using it for?

You're throwing out buzzwords in the hopes that we'll all accept that you are just too advanced for the rest of us.

I can't force you to explain yourself - but I can damn well sit here and appreciate that everyone can see through what you're trying to do.



Indeed 🤣

From his answers to me and also his thread sanity checking new build Mac Pro 7,1 in 2025 then the main feature of it is as a Proxmox Environment. Even has a PCIe extender for more GPU.

From what can tell has previously used Lenovo and HP Z but prefers the way Apple put the Mac Pro together.

As a virtualisation host then yep, lots of CPU and RAM required which really isn’t the way Apple gone, and AS not built for that sort of thing,

Proxmox if running on Apple Silicon requires Rosetta 2 to run, so no native AS version, but still x86

Bought in 2025 so not going to be new.

Having used Lenovo/HP previously then the workload doesn’t require a Mac Pro as in a Mac OS machine, but a system for Virtualisation so an x86 tower with lots of CPU, Storage and GPU.

Video editing in his thread says could use a mini or possibly a studio, so I suspect that the heavy lifting is not done in Mac OS but Linux/Windows Guests, and nothing wrong with that at all.

So nothing in here that if I was Apple would have me screaming at the engineers to develop an Extreme SoC or even keep an Intel Mac Pro about in the range that fits outside that crafted model of the Ecosystem.

Apple have left the “Server” arena years ago. Dropping the Xserve hardware, then dropping server product to make it an App on Mac OS before even dropping that.

 

[BSDnostalgia]

Apple have left the “Server” arena years ago. Dropping the Xserve hardware, then dropping server product to make it an App on Mac OS before even dropping that.

And, I regret that Apple made those decisions. It would cost virtually nothing to repackage the M3 Ultra (soon to be an M5 Ultra I hope) in a 2U rackmount server with a bunch of NVMe slots.

 

[Audentia]

That isn't feasible at all. The whole point is to have the CPU and GPU and memory on the same chip, because the reduced physical distance between everything is what makes it more efficient and powerful.

Introducing PCIe boards throws all that away for nothing. Nobody is ever going to build that. Physics doesn't agree with it.


No? It doesn't work like what you described at all. Nor would it, because that would be a very ineffective and wasteful way to build that tech. As has been pointed out that stuff costs something in the $100k range.

Look I get that the idea of having this desktop tower that you can keep adding stuff to modularly seems neat, but computing is permanently moving away from it.

uhhh, no. https://www.nvidia.com/en-us/data-center/nvlink-c2c/

Physics is not the problem here. cost may well be.

If you want to talk about physics and waste, what's wasteful is a bunch of thunderbolt connected mac studios when there's clearly better options, this isn't about neat modularity nostalgia, it IS about bandwidth and physics. It is about density of compute in a workstation package, not a server package, not a tabletop pc package.

Computing permanently moving away from it? are you from the past? NVLink and these servers are doing the opposite, and should be impossible by your logic. Computing platforms are always about increasing bandwidth at different points of expansion, unified memory on a SoC is no different, but is not end game nor one size fits all.

 

[BSDnostalgia]

uhhh, no. https://www.nvidia.com/en-us/data-center/nvlink-c2c/

Nvidia was rather boring when it seemed to be focused only on ray-tracing hair at real-time speeds. Then, it was AI.

Now, it turns out that Nvidia has a real strategy to eat everyone's lunch. Everyone better pay attention to what Nvidia is doing besides dGPUs.

 

[leman]

uhhh, no. https://www.nvidia.com/en-us/data-center/nvlink-c2c/

Physics is not the problem here. cost may well be.

If you want to talk about physics and waste, what's wasteful is a bunch of thunderbolt connected mac studios when there's clearly better options, this isn't about neat modularity nostalgia, it IS about bandwidth and physics. It is about density of compute in a workstation package, not a server package, not a tabletop pc package.

Yep. Thunderbolt is not a particularly suitable vessel for this, which is why it would be great if the next Studio came with a faster interconnect port + external switch support. I am not familiar with this market segment, and I wonder what options currently exist. I have read something about the upcoming UltraEthernet for example?

At the same time, there are two topics here that I think need to be distinguished clearly. One is high-bandwidth interconnects for work orchestration. You have a tasks that you can divide into parallel blocks and execute on different devices, and while the high-speed connection allows you to scale these workloads better, it is still an order of magnitude or more slower than the within-device data bus. That's the standard datacenter approach and it requires a good understanding of the problem and building the software with the asymmetrical memory architecture in mind.

What many Mac Pro enthusiasts here have been calling for however is the transparent scalability — the idea that you add in another compute board to your chassis and it "just works". To be truly seamless, that would require much faster between-blade connectivity, and that's where things quickly become unrealistic.

https://www.nvidia.com/en-us/data-center/nvlink-c2c/

Now, it turns out that Nvidia has a real strategy to eat everyone's lunch. Everyone better pay attention to what Nvidia is doing besides dGPUs.

Nvidia is one sly company, but most of the stuff they are doing nowadays is ML-focused. They have rearchitected their GPUs with ML in mind, for example. They are certainly taking lessons from Apple, which makes sense given where the industry and consumer sentiment are going, but beyond ML I don't see anything truly disruptive yet. The big limitation, as always, is the software. This is the third big attempt to bring Windows to ARM, let's see how it will work out.