Mac Pro Reportedly on 'Back Burner' and 'Largely Written Off' at Apple

[Alan Wynn]

It makes no sense for a “pro”, say scientific researcher, to waste dollars on that kind of hardware. Just go to the cloud.

Using Spot instances on AWS would typically payoff an equivalent machine in about 3 months. Reserved instances would usually take 6 months. Storage and egress are even more expensive.

In addition, what is the price for a GPU with 256GB of VRAM on AWS?

LucidLink (simulated block storage) is $80 a terabyte about 10x the purchase price of that storage. Even Wasabi/B2 s3 storage payoff actual storage in under a year.

I’m literally running my stuff on AWS from MacBook Air because I don’t need to pay for the hardware in hand.

What instances are you using? How much are you paying?

AWS and its competitors are almost never cheaper than owning one's own hardware. Sometimes they make sense from a business perspective if it is easier to pay for OpEx than CapEx.

I get movie studios and other use cases may have certain needs. I don’t personally have any experience in that space.

I have to say that the number of people who have said the equivalent of "I do not use this and so it should not exist" to those of us that actually need these tools.

 

[DrJR]

Using Spot instances on AWS would typically payoff an equivalent machine in about 3 months. Reserved instances would usually take 6 months. Storage and egress are even more expensive.

In addition, what is the price for a GPU with 256GB of VRAM on AWS?

LucidLink (simulated block storage) is $80 a terabyte about 10x the purchase price of that storage. Even Wasabi/B2 s3 storage payoff actual storage in under a year.



What instances are you using? How much are you paying?

AWS and its competitors are almost never cheaper than owning one's own hardware. Sometimes they make sense from a business perspective if it is easier to pay for OpEx than CapEx.



I have to say that the number of people who have said the equivalent of "I do not use this and so it should not exist" to those of us that actually need these tools.

What kind of “spot instances”? The ones where I’d leave my MBA grinding for an hour on a data crunch pipeline while I went to lunch costs nothing with my student account, but maybe $1 per run on AWS Sagmaker. Depends on what I’m doing. PhDs used to buy a Mac Pro and run it a month straight to do one pipeline. No more. AWS is great! Owning your own hardware is very 2015.

 

[avro707]

would be so happy if they decided to admit they were wrong to remove those features.

Could you imagine the uproar from the Apple supporters who invested so much time arguing the Apple line who would now have to do mental gymnastics to now explain why Apple has decided to switch course?

I’m happy with the two 2019 Mac Pros I have. They have been absolutely faultless. They are powered by solar power and my battery system (as is the rest of my place).

My only gripe is that memory prices are escalating to stupid levels, but that’s a problem everyone is facing.

 

[Alan Wynn]

What kind of “spot instances”? The ones where I’d leave my MBA grinding for an hour on a data crunch pipeline while I went to lunch costs nothing with my student account, but maybe $1 per run on AWS Sagmaker. Depends on what I’m doing. PhDs used to buy a Mac Pro and run it a month straight to do one pipeline. No more. AWS is great! Owning your own hardware is very 2015.

What are the configurations of the machines you are running? Do you know anything about AWS’s actual cost structure, or are you only in the free tier?

 

[mozumder]

ECC RAM for LLM still matters due to reduced accuracy, training instability, wasted compute resources, file corruption, and more. Without it, it will only limit the size of LLM you can run or otherwise, quite risky. Whenever you have a lot of memory, ECC is necessary.

A bit error isn't going to cause reduced accuracy or training instability.

 

[lusty]

Huh?
Just make the main drive socketed on the Studios, like it is right now on the Mini & Mini Pro.

I’m not sure you understand the problem. PCIe lanes can only be used once in an architecture. You can use them for PCIe expansion or Thunderbolt, or graphics, or sound. The M series chips onlyhave enough for the graphics, sound, networking and USB/Thunderbolt.
This means that Apple couldn’t make a proper Mac Pro even if they wanted to, there simply isn’t enough bandwidth on the chip to offer a bunch of PCIe expansion for storage or anything else. Yes, they can add a storage slot, but no they can’t let you add 8 PCIe storage cards.
Thunderbolt is a PCIe x4 connector so not the same as an x16 slot in terms of bandwidth, not even close.

For most, that doesn’t matter and the Studio is absolutely capable. For anyone that needs more bandwidth in the system though, the M series is simply not as capable as a Xeon. That’s not a bad thing, they’re designed for different workloads.

I wasn’t bashing the chip in my post, I was explaining why the limitation is there. It’s a technical limitation of the architecture and partly why the M chips run efficient and cool as they do less stuff.

 

[Ethosik]

I have no idea on that one, I’m genuinely surprised, about the only thing I can think of is that maybe yield on the M3 Ultra is low so they prioritized the studio

That’s not a “mess up”, that’s an intentional architecture decision. They made the determination that the M4 didnt fit with making an ultra version. It’s possible they wanted the die space the ultrafusion interconnect takes up for added components in the much much much higher volume max chips that maybe they’ve managed to minituarize enough to fit the interconnect back in on the M5, it’s possible it was a node yield issue where they prioritized a max chip, it’s possible they’re trying to milk the last out of the M3 process node and tape out, etc. there are a lot of completely valid reasons to make that choice, and again it’s actually rather common for higher end enterprise chips to lag a generation behind more consumer oriented desktop chips.

It is a mess up even if it was intentional. It speaks to them not adding enough focus on the high end components.

 

[seek3r]

What kind of “spot instances”? The ones where I’d leave my MBA grinding for an hour on a data crunch pipeline while I went to lunch costs nothing with my student account, but maybe $1 per run on AWS Sagmaker. Depends on what I’m doing. PhDs used to buy a Mac Pro and run it a month straight to do one pipeline. No more. AWS is great! Owning your own hardware is very 2015.

In my experience doctoral students doing really big calculations usually use local lab resources for staging jobs for tests and then run big workloads on either university compute cluster or national lab/etc compute cluster resources. If you have a valid reason it’s not all *that* hard to get time on NCAR’s machines etc

That said my experience on that end is mostly on the HPC side, maybe for things like AI that’s different

 

[jafd]

But as many have already stated, there are breakout boxes for that.

I said „not easily yankable”

 

[jafd]

Seriously, I’d like something like „Mac Pro Ultra” which is like your Mac Pro, but you can fill PCIe slots with cards that are equivalent to Mac Mini or Studio in computing power, basically DPUs with their own Apple Silicon SoC, RAM, local storage, maybe 10/100 GbE and Thunderbolt at the back, and they have internal interconnect over PCIe and some shared memory. Imagine clustering and computation possibilities one of those things would give.

 

[DrJR]

In my experience doctoral students doing really big calculations usually use local lab resources for staging jobs for tests and then run big workloads on either university compute cluster or national lab/etc compute cluster resources. If you have a valid reason it’s not all *that* hard to get time on NCAR’s machines etc

That said my experience on that end is mostly on the HPC side, maybe for things like AI that’s different

today that is correct. 2015 was different

 

[platinumaqua]

there’s such a missed opportunity there for the studio to have been called the “Apple Seed” and a breakout box the “Apple Branch” with each PCIE slot a “leaf”

AppleSeed is one of Apple's beta testing programs

 

[chucker23n1]

AppleSeed is one of Apple's beta testing programs

Yep.

They also often have John Appleseed as a demo user.

 

[Darth Tulhu]

When has Apple EVER covered all bases?

They only cover what they deem most profitable AND necessary.

I could be wrong but I believe that with the Mac Studio Apple has covered the pro bases they want to, because those are the ones they've deemed most profitable.

The puck has been moving away from the Mac Pro for a while now, so the report of the Mac Pro being written off seems credible as Apple is always skating to where they think the puck is going to be.

And the Mac pro is in the opposite direction.

 

[CWallace]

They could and should have continued using Intel for machines they didn’t want to compete in the CPU arena. And it would’ve remained upgradable too.

Could have: Yes.
Should have: No.

Apple struggles to maintain macOS as is, and it would only get worse with two dedicated forks (one for Apple Silicon and one for x86 Xeon because x86 will not support more and more features Apple Silicon does). They also would need to maintain two R&D teams for the Apple Silicon Mac Pro and the Intel x86 Xeon Mac Pro. And we have Intel's struggles with design and fabrication of said x86 Xeons.

It's just not worth it to Apple and probably would end up not being worth it to Apple customers.

So much of this is downstream of having Cook in charge.
I strongly believe there's an alternate reality he was hoping we'd be living in now where it's iPhones & iPads only.

I recall it was Steve Jobs who was pushing iPads as the future of general-purpose computing and "traditional PCs" like the Mac would fade away into a smaller and smaller niche for people doing "edge-case" work.

 

[jktwice]

I think with the advent of artificial intelligence workloads there is still a place for this kind of computer, but it is very low volume.

To get the most out of it, Apple should either sell SOCs on a card that one can plug in and accelerate their workloads with, or open up GPUs on the platform for CUDA or similar workloads. But instead they won't do this because then they can't get you to buy a whole new machine to cluster it.

Maybe if Apple makes a breakthrough with AI, they will roll out a new Mac Pro to facilitate this kind of development.

 

[seek3r]

AppleSeed is one of Apple's beta testing programs

In that case “Appleseed+”

 

[seek3r]

today that is correct. 2015 was different

My experience on that end stretches back from now back to ~2005, it was true then too, it’s been true on the HPC end for a long time - in fact my job back then was one of a couple folks maintaining a university compute cluster. What’s funny is that a single mac studio probably has more horsepower now than that entire datacenter did (though less RAM)

 

[seek3r]

It is a mess up even if it was intentional. It speaks to them not adding enough focus on the high end components.

Or it speaks to them making technical and business decisions based on vastly better info than you have. Apple’s chip engineers arent stupid, they didnt leave off the uf interconnect on the M4 Max for funsies, there definitely were valid reasons

 

[Abazigal]

When has Apple EVER covered all bases?

They only cover what they deem most profitable AND necessary.

In a sense, Apple attempts to cover all the bases, in a manner that's profitable and convenient for them.

For example, no modular Mac back on 2015, because the iMac already exists, and was deemed good enough for the majority of professional workflows, even though it didn't directly address power users' demand for expandability and the freedom to connect their own displays.

 

[x86dude]

Beautiful design wasted.

Bring in new people to run the desktop division.
People with ideas who want to do it.

 

[Bistroengine]

Ridiculous. Instating throwing up their arms and giving up, maybe try going back to the drawing board and give pro users what they really wanted in a Pro machine. Affordable memory upgrades, Swappable CPU upgrades, at least 2 to 4 internal M.2 SSD ports and/or internal SSD drive bays, fast PCI expansion slots and lower the price. When a faster M series chip is released then sell a plug-n-play board to instantly boost your performance. They did it with the original 'Cheese Grater', they can do it again if they stop being money-hungry greedy bastards

 

[Ethosik]

Or it speaks to them making technical and business decisions based on vastly better info than you have. Apple’s chip engineers arent stupid, they didnt leave off the uf interconnect on the M4 Max for funsies, there definitely were valid reasons

Then why didn't they update the Mac Studio when M3 was the latest? Why wait until M4 Max to release M3 Ultra?

It just speaks volumes that Apple does not care about the pro line of the Macs at all ever since the trash can Mac Pro and they BARELY care about the laptop line.

If what you are saying is true that there was ZERO possibility of making an M4 Ultra, then maybe go back to Intel or AMD if you cannot properly make a high end processor - huh? If Apple is so incompetent at it, there is Intel and AMD that blow away anything M3 Ultra is capable of and now I have ZERO confidence Apple will EVER match the desktop power of the x86 alternatives.

EDIT: Okay even if I grant you that there were extreme limitations on the M4 that could NOT have the Ultra variant. Why isn't the Mac Pro updated to M3 Ultra? Why is the Mac Pro STILL $7,000+?

MAYBE......just MAYBE......an over engineered $7,000 tower that can barely compete with a $2,500 x86 system is the reason it is not selling well? Why the heck would ANYONE purchase a Mac Pro today while it is still stuck on M2 Ultra? Why the heck would ANYONE purchase the Mac Pro back when M2 Ultra was the latest when it had a $3,000 premium.

Oh right, I know only those that REQUIRE PCIe. Nobody in their right mind would get one for any other reason. And everyone wonders why it isn't selling well?

 

[BNBMS]

Or it speaks to them making technical and business decisions based on vastly better info than you have. Apple’s chip engineers arent stupid, they didnt leave off the uf interconnect on the M4 Max for funsies, there definitely were valid reasons

Apple themselves made quite a lot of stupid choices.

 

[Ethosik]

Apple themselves made quite a lot of stupid choices.

They have, 2010 and MAYBE 2012 Mac Pro was the last good one. I seem to recall 2012 Mac Pro having some controversial decisions too but maybe I am remembering wrong.

Again, it's not rocket science. Get Dell to help you build a desktop computer if you cannot do it yourself. Apple is one of the biggest IT companies and they haven't been able to make a proper desktop since 2012! The 2019 was severely over engineered. That was also over priced as well.

I am not the type to argue about Apple's pricing, I literally have 8 Macs. Including TWO almost maxed out M2 Ultra Mac Studios. But let's be real. Even the 2019 Mac Pro was way too expensive for its offering. I was thinking about getting one, but the 2019 i9 iMac was a FAR better option with FAR better specs with several thousand dollars less.