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

[Bug-Creator]

I've long supported the idea of the Mac Pro as a Halo product

The B/W G3 might have been a Halo product, if we stretch the definition a bit one might say that about some top end G5 or OG MacPro configs.

The Trashcan and CheeseGrater_MK2 never qualified as they were barely o.k. on launch and turned into a joke after not getting timely upgrade (read, none at all). The M2-Ultra version was the anti-thesis of "Halo" as it was the butt end of a joke from day one (not saying it didn't have a use case) and the general Mac buying public hardly knew it existed.

Heck if it wasn't for this thread I would have thought it had been discontinued 2 years ago (well one could say it has).

The closest thing to an Halo product in recent Apple history would be the VisionPro, just didn't work out that way.

 

[rafark]

I started to type a thread on the new Spark workstations, but fell asleep. Yeah, I'm right there with you. The 7,1 was all this back in 2019 if you had the pockets. Crazy to me how they axed it. A studio cannot be your halo product, Apple - it just can't - it's laughable.

And now Nvidia + Windows covering laptops, desktops, and actual workstations... they're going to continue eating everyone's lunch. AMD better keep introducing new features to legacy hardware.

Apples flagship Mac is the MacBook Pro. The Mac Pro was basically vanity at this point. I prefer non MacBook macs but to be honest a super powerful (and expensive) desktop computer wouldn’t be profitable for a company the size of apple because that market is too niche for them

 

[dmr727]

Please tell that to everyone running discrete GPUs with performance that crushes AS.

So you're no longer talking consumer GPUs. Okay, and how many of those folks would install their $10K+ graphics card into a Mac Pro if given the chance?

 

[leman]

Nvidia and their partners just announced a 3 tier lineup:

- laptops
- desktops (mini and studio sized boxes)
- workstations (MP sized towers)

Just to make it clear, the upcoming Nvidia workstations (DGX Station) cost over $90,000 for basic configuration, and that is without a GPU that can actually perform graphical tasks.

The Spark laptops and desktops compete with M5 Max at best, and come with much weaker CPU. The Spark GPU is an equivalent of RTX 5070, so a bit slower than M5 Max for most applications (and of course, Nvidia has an edge for quantized inference).


Which again leads to the question: what exactly are you suggesting here? That Apple's Mac Pro is a $100000 tower? Would you buy one?

 

[leman]

So you're no longer talking consumer GPUs. Okay, and how many of those folks would install their $10K+ graphics card into a Mac Pro if given the chance?

The answer is zero because no MP model can supply enough power to run these behemoths.

 

[dmr727]

The answer is zero because no MP model can supply enough power to run these behemoths.

Heh - I didn't even consider that. 🙂

 

[mcnallym]

Nvidia and their partners just announced a 3 tier lineup:

- laptops
- desktops (mini and studio sized boxes)
- workstations (MP sized towers)

My response was specifically to a person about the Nvidia solution where selling a complete ARM based standalone solution with unified memory and comparing it to the Mac Studio and claimging that left the Mac Studio outmanoeuvred. As I pointed out then a Mac Pro and Mac Studio in terms of SoC are the same so Mac Pro in its current incarnation from 2023 would make no difference and the real difference is in effect the CUDA Support on the Nvidia solution.
But hey if you want to then take my reply and then move to to different add in solutions that added to Laptops, Desktops and Workstation as opposed to the system that the person was comparing to the Studio then more power to you but is disingenuous at best.
As stated then Nividia are not making a Mac Pro but a Mac Studio style machine. These different add-in allows PARTNERS to add Nvidia Solutions other the partners offerings without having to do custom hardware. you can put them into a generic x86 system and has drivers for Linux and Windows which the partners support as OS already.
Benholts post quoted below for reference.

Agree 100%, Apple officially pulling the plug on the Mac Pro in March was spectacularly bad timing. They abandoned their top-tier workstation precisely when local LLM inference made massive unified memory a killer feature.

Nvidia’s RTX Spark announcement this week capitalizes on that exact void. By combining a 20-core ARM CPU, a Blackwell GPU, and up to 128GB of memory (all with native CUDA support) Nvidia is bringing Apple-style memory efficiency to Windows. Apple effectively ceded the local pro AI market just as demand exploded. The Mac Studio is capable, but up against equivalent unified-memory ARM PC backed by Nvidia's dominant AI software stack, it's completely outmaneuvered.

Benholt there is clearly comparing the Mac Studio to the RTX Spark which is an Nvidia specific product consisting of an ARM based SOC timing together Blackwell GPU with ARM CPU with Unified Memory. As I said then the key difference between a sealed Mac Studio (which has the same Soc as the Mac Pro so no different) and the sealed RTX Spark is the CUDA Support which Apple not interested in supporting.

Apple has never been interested in supporting CUDA in Mac OS. Mojave 2018 dropped the support for the Nvidia Web Drivers and CUDA add-in software. 2014 saw the launch of the last Mac with an Nvidia GPU. Apple support since then been AMD GPU in Mac Pro 2013 and 2019 models. Want to install Nvidia then load a non Apple OS and don't call Apple Support.

So no CUDA support from Apple and Apple seem to have been happy with that for the past 8 years in terms of OS.
So this is not a recent thing.

What is distinctly missing in Benholts post is the separate RTX PRO solutions for Laptops/Desktops/Workstations. Benholt is stating that the United Memory is a killer feature for running LLM (not my words but Benholts so take that up with Benholt if have a problem with it) so not surprising not referencing add in solutions that don't do Unified Memory in his post.
These addin solutions would not work in an Apple Silicon Mac OS Mac Pro so you would have to be moving to a non Apple system, like the kind of generic x86 systems that Nvidia's partners make.

Literally since Mojave in 2018 then has been clear that Apple not interested in CUDA and not moving towards it anyone that requires CUDA (and indeed is a very valid requirement for people that work in the arena that CUDA is for) then people should have been planning a migration away from Apple for that work.

Apple told people this 8 years ago through actions. The fact that people got it working in a Mac Pro 2019 and the older 2009-2012 models by booting into Linux and Windows is just a it works, as opposed to deliberate choice by Apple to make it work.

With Apple Silicon so with the launch of the M1 machines in November 2020 then Apple also been clear that not particularly interested in other OS booting on its hardware. Whilst they permit other OS to be loaded they are not actively working on anyone to port their OS to Apple Silicon and is down to the OS people to reverse engineer what need to do and make it work.

Not really sure why people find it so difficult a concept to follow,
If you want to run CUDA you don't start with a Mac.
If you want an Nvidia Solution then you don't start with a Mac.

 

[Omega Mac]

@mcnallym Very interesting.

Do you think Apple actually tactically backed away from Nvidia, under cover of acrimony (used as a plausible off ramp) to avoid a sustained industrial supply chain relational dependency would have gifting Nvidia heads-up on their ground-breaking industry leading Apple Silicon plans at a very early technical level, thus strategically removing their reliance on a potential competitor, since Apple was now moving into Silicon, where Nvidia already was and didn't want to give Nvidia good reason or insight to counter and move into Apple territory sooner?.

Yes Jobs talked about owning the whole package, no secret there, and with him gone maybe some thought they won't follow through, nor is that a new idea either since the Amiga was very much this model, i.e. Apple is the Amiga that never was.

My thinking here is, making such serious industry leading pivot, they would naturally want to maximise their entrance, i.e. learned from iPhone Eric Schmidt Android equation right?

Giving Apple a 5 or more year lead on stiff and capable competition, and now watch everyone Apple clone to within an inch of the court room. :shrug

 

[leman]

Do you think Apple actually tactically backed away from Nvidia, under cover of acrimony (used as a plausible off ramp) to avoid a sustained industrial supply chain relational dependency would have gifting Nvidia heads-up on their ground-breaking industry leading Apple Silicon plans at a very early technical level, thus strategically removing their reliance on a potential competitor, since Apple was now moving into Silicon, where Nvidia already was and didn't want to give Nvidia good reason or insight to counter and move into Apple territory sooner?.

Yes Jobs talked about owning the whole package, no secret there, and with him gone maybe some thought they won't follow through, nor is that a new idea either since the Amiga was very much this model, i.e. Apple is the Amiga that never was.

My thinking here is, making such serious industry leading pivot, they would naturally want to maximise their entrance, i.e. learned from iPhone Eric Schmidt Android equation right?

Does it have to be this intricate? I'd say the simpler explanation is that Nvidia did not have any interesting technology to offer. Apple needed a low-power, high density, efficient GPU architecture that would scale from mobile to desktop and from simple parallel image processing to highly complex asymmetric shaders. And this becomes very clear when we look at the trajectory taken by these two GPU designers recently — Nvidia simplified their execution, focusing on highly parallel architectures and high-throughput matrix accelerators to focus on ML applications. Apple instead went the opposite way — complex execution, sophisticated memory hierarchies, and adaptive resource partitions — and as the result, they outperform Nvidia almost 2-to-1 on complex GPU code, simply because they can achieve much higher shader occupancy under adverse conditions.

So overall, I doubt that there were additional strategic factor in winding down cooperation with Nvidia. Another consideration is that back in the day Nvidia was becoming an unreliable partner (systematic issues with Nvidia GPUs in MacBooks) and sometimes even an aggressive adversary (e.g. Nvidia actively sabotaging Apple's OpenCL initiative to promote their own CUDA).

 

[Flint Ironstag]

So you're no longer talking consumer GPUs. Okay, and how many of those folks would install their $10K+ graphics card into a Mac Pro if given the chance?

Consumer, Pro, doesn't matter the branding of the GPU - but I'm 100% talking about full powered desktop GPUs vs crippled mobile variants.

The point is to throw a bunch of them in a box and leave AS in the dust. And given the amount of MPs for sale with dual pro duo GPUs, which was a $10k option back then, people bought them and would today. There's folks on other platforms (reddit etc) installing mega GPUs in their MPs right now, especially for AI. Take a look for yourself.

If a single $10k GPU could do everything I want, I'd absolutely consider it. But with depreciation and the power of Proxmox, and let's not forget our friend PCIe - I like the flexibility of multiple less-powerful GPUs. I've never been after peak performance in any single area. I don't mind providing input and waiting for output. It's like having a pen pal. As long as my user interface remains responsive, I'm happy.

And there's also the fact that I've been enjoying 100+ TFLOPs in a single MacOS box for... almost 10 years now. Way before RTX 6000 Blackwell in 2024. I understand the the RTX is gonna clean my clock in MANY MANY situations. But the point is that a proper Mac Pro or HP Z840 proxmox will allow you to do this kind of stuff WAY before you can get it on a single card.

Yes it gets hot. Yes it gets loud. Yes, these things should make you laugh and want to add more.

Back to the $10k thing, let's compare my old obsolete rusty rig to Nvidia RTX Pro 6000 Blackwell.

Both ~120 TFLOPs FP32. Somebody please chime in and tell me what the fastest Apple Silicon FP32 TFLOPs is. And also the FP64. You can go ahead and cluster them. thanks. Whatever it is, if it's even close - I can add more or faster GPUs. 😐

Nvidia 96GB VRAM / me 192GB (I have not bought IF links for the external GPUs yet) and some traffic will have to traverse PCIe - effect on performance is highly dependent on the app.

Nvidia 350W / me ~2000W (both figures just for GPUs, not supporting hardware)

Nvidia new 2024 $10k / me new all GPUs and PCIe chassis in 2017 would've been ~ $25k

So you see how it really lines up with market values. I didn't buy my stuff new (well, the Pro VIIs were new but old stock), but you see it cost twice as much in 2017 to approximate the performance of a monolithic card in 2024. And the old stuff keeps getting cheaper for now, but as @Regulus67 said, instead of buying it up, folks chase the next dream with inferior specs. Amazing in the worst way.

Just to make it clear, the upcoming Nvidia workstations (DGX Station) cost over $90,000 for basic configuration, and that is without a GPU that can actually perform graphical tasks.

The Spark laptops and desktops compete with M5 Max at best, and come with much weaker CPU. The Spark GPU is an equivalent of RTX 5070, so a bit slower than M5 Max for most applications (and of course, Nvidia has an edge for quantized inference).


Which again leads to the question: what exactly are you suggesting here? That Apple's Mac Pro is a $100000 tower? Would you buy one?

OOF!!! 🤣 I did not know the Nvidia workstations are starting at $90k - that's crazy! And that's because they can get away with it. They need some competition. A $50k maxed out MP wouldn't be looking so bad right now...

- whatever AMD processor for full AMD pipeline. You KNOW they have a build customized for this internally.
- dual AMD whatever duos with combined 512GB+ VRAM and IF link
- PCIe expansion to add more GPUs as you desire (or whatever else)

I'm a fan of not being locked in to CUDA. I think we need to embrace open standards / open source. I'm willing to accept some rough edges and workarounds in exchange for being able to run my stuff on whatever I want, but really that shouldn't exist - the open source should be just as polished. AMD & Apple should have remained teamed up against Nvidia IMHO.

 

[mcnallym]

@mcnallym Very interesting.

Do you think Apple actually tactically backed away from Nvidia, under cover of acrimony (used as a plausible off ramp) to avoid a sustained industrial supply chain relational dependency would have gifting Nvidia heads-up on their ground-breaking industry leading Apple Silicon plans at a very early technical level, thus strategically removing their reliance on a potential competitor, since Apple was now moving into Silicon, where Nvidia already was and didn't want to give Nvidia good reason or insight to counter and move into Apple territory sooner?.

Yes Jobs talked about owning the whole package, no secret there, and with him gone maybe some thought they won't follow through, nor is that a new idea either since the Amiga was very much this model, i.e. Apple is the Amiga that never was.

My thinking here is, making such serious industry leading pivot, they would naturally want to maximise their entrance, i.e. learned from iPhone Eric Schmidt Android equation right?

Giving Apple a 5 or more year lead on stiff and capable competition, and now watch everyone Apple clone to within an inch of the court room. :shrug

Well the Spat with Nvidia started late 2000's with the 2008 Nvidiagate issues with the 8600m chips in MacBook.
Apple acquired PA Semi and an ARM Architecture License in 2008. MacBook Air was launched 2008.
MacBook Air and the later MacBook 12" both kind of held back by Intels inability to move to cool and lower power CPU which probably also helped with the decision to move off x86

I suspect just a coincidence as the talks with PA Semi and ARM would have gone on for some time before being announced so the decision to move don't think a direct response to the problems but did highlight the problems Apple starting to face,

iPhone launched 2007 the year before so suspect talks with PA Semi started then with an idea to move to ARM and then move upwards with it after cutting teeth on the SoC design, so before the Spat with Nvidia began. AMD seemed more open to OpenCL and then Metal then Nvidia so whilst transitioning a move to AMD which Apple had used before Nvidia as GPU partner just moved Nvidia out quicker then would have done.