Apple has been building their own graphics cards for years. Works out just fine. Not really sure what you mean?
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Mac Pro should use NVIDIA cards
- Thread starter Zest28
- Start date
- Sort by reaction score
Apple has been building their own graphics cards for years. Works out just fine. Not really sure what you mean?
Don’t be too quick to look at yesterday’s marketing decisions as defining performance requirements tomorrow.
iOS 18: AI Server Industry Aiming to Win Business From Apple
AI server makers are hoping to obtain orders from Apple ahead of its highly anticipated unveiling of new AI features later this year, according to...www.macrumors.com
BUT... Apple needs to make their own server, sooner or later, especially for AI. Currently, only Nvidia has the power of AI on both hardware and software and therefore, Apple has to buy x86 and Nvidia based servers which is quite ironic. Apple has neither hardware and software and yet how can they even enter the AI market?
Don’t bother arguing with sunny. They are well known on these forums for applying reductionist pseudo-arguments and cherry picking into the absurd. Not much point into entering a debate with a bad faith counteragent who operates on an axiom “everything that Apple does sucks by definition”
Apple obsoletes a product 7 years after a product is discontinued*, not after launch - so 2030 I think since the Intel Mac Pro was discontinued just last year right? OS updates though are more variable, but on average they give a product about 7 years of OS updates after launch with generally another 3 years of bug fixes/security updates as Apple maintains the last 3 years of OSes. So again, depending on the level of support a user is looking for, buying a used 2019 Intel Mac Pro now they'd have official support until about 2030 give or take.
EDIT: *as mentioned in the linked article laptop batteries may be serviced for longer but that's not really relevant to this discussion about the Mac Pro, just an FYI to anyone reading
Last edited:
Why would Apple bother buying servers with Nvidia hardware when they could just rent them and not have to worry about having hardware to take care of?
They still need dedicated servers. Beside, performance matters and that's why cloud and rendering farm cant replace local workstations.
Gosh, this must be one of the most unuseful or pointless thread in the foruns.
First, are we talking servers or desktop? Bc nvidia performance was due to its server chips, h100, a100 etc. These are not the markets apple will mess with… why? Bc there are several new actors coming up that will run inference way faster than nvidias.
Second, desktop apple silicon is the cheapest way to effciently run most of the beefy LLMs models right now because of unified memory architecture. If you want to runa mixtral model, you need a a6000, which is more expensive than a Studio.
So, whats the point?
Also, @sunny5 music/sound does not rely on GPU… I’m not sure where did you get this info from. Also, video market is dominated by apple, specially because of AS architecture. We dont rely on GPU, more on the video engines and bus bandwith + unified memory.
First, are we talking servers or desktop? Bc nvidia performance was due to its server chips, h100, a100 etc. These are not the markets apple will mess with… why? Bc there are several new actors coming up that will run inference way faster than nvidias.
Second, desktop apple silicon is the cheapest way to effciently run most of the beefy LLMs models right now because of unified memory architecture. If you want to runa mixtral model, you need a a6000, which is more expensive than a Studio.
So, whats the point?
Also, @sunny5 music/sound does not rely on GPU… I’m not sure where did you get this info from. Also, video market is dominated by apple, specially because of AS architecture. We dont rely on GPU, more on the video engines and bus bandwith + unified memory.
1. It's both and Nevada makes both. Beside, even Apple used Mac Pro for sever uses and made a rack version. Also, M2 Ultra is not even close to RTX 3080 or 3090. Apple is limiting themselves since Apple Silicon and as a result, their revenue did not improve.
2. UMA is meaningless when GPU performance and software/ecosystem sucks. Beside all AI software and developers are heavily based on Nvidia's ecosystem and hardware. How does UMA has to do with it? As I have been using AI generated art, Nvidia is overwhelmingly faster despite having low RAM or VRAM while Apple Silicon Mac used all UMA and yet much slower than Nvidia GPU. How is that? Memory is fast, not GPU. If it was not an issue, how come Apple is not doing anything in AI so far while Nvidia is taking 4th place and yet still growing? Nvidia had been investing AI for several decades.
3. And yet, Mac Pro 2023 only has 192GB of RAM. Neil Parfitt mentioned that he used more than 300GB of RAM while working on the music project. Are you gonna argue that UMA is efficient so it will works as 192GB UMA vs 300GB RAM? UMA VS regular RAM theory is already proven wrong by many people. LOL... Also, video on Mac is powerful only because of its dedicated video chips and it still requires GPU for export AND working on graphic on videos. And you just admitted yourself, not relying on GPU is a main problem as they can NOT expand to 3D, game, AI, research, and more where GPU power is needed. That's why Mac is only good for 2D such as video, photo, music, illustrate, and more. How long do we have to hear that Mac sucks for 3D or GPU intensive software?
Apple clearly abandoned Pro market starting with Mac Pro and even Video and Music that Mac is proud of, are also affected. If Apple doesn't care, that's fine as they are limiting and restricting their own just like other Apple devices and fell behind others with AI.
Last edited:
Hi, Audio Engineer here.
That's an absolutely pathological case for an audio project (using that much ram). You'll be using swap memory anyway for projects like that, as well as freezing tracks. 99.9% of audio jobs will be using a fraction of that memory.
2/3rds of the most commonly used professional audio software only work on macOS (Logic and Luna). Even pro tools has functions are that are macOS only (mainly to with video and audio workflows).
I would not be surprised if the next Mac Pro doubles its memory anyway.... But even 192GB is not a limitation for music work.
What I'm saying is that Apple themselves literally shrink their own markets starting with Apple Silicon. Mac Pro 2019 provided up to 1.5TB of RAM while Mac Pro 2023 can not. It only limits what Mac can do especially since GPU sucks and software pools are limited. How come some people justifying poor hardware just because it's good for video and music? That's a poor excuses when it's Apple's fault to shrink the market.
For those who are wondering why Apple has not created a translation layer
www.tomshardware.com
Nvidia bans using translation layers for CUDA software — previously the prohibition was only listed in the online EULA, now included in installed files [Updated]
Translators in the crosshairs.
How would they enforce something like this?For those who are wondering why Apple has not created a translation layer
Nvidia bans using translation layers for CUDA software — previously the prohibition was only listed in the online EULA, now included in installed files [Updated]
Translators in the crosshairs.
Im wondering about the same thing. How would EULA play a role here? I don’t need to agree to an EULA to develop a CUDA compatible library.
iOS 18: AI Server Industry Aiming to Win Business From Apple
AI server makers are hoping to obtain orders from Apple ahead of its highly anticipated unveiling of new AI features later this year, according to...
BUT... Apple needs to make their own server, sooner or later, especially for AI. Currently, only Nvidia has the power of AI on both hardware and software and therefore
Some stuff is a hypetrain struggling to justify why SuperMicro and Nvidia stock is hyper inflated. Apple doesn't have to chase their stock's hype.
Apple doesn't "need to". The major features of the Vision Pro are AI based and require zero connection to an Nvidia 'powered' server at all. The presumption here is that Apple is going to counter the 100+ B element GPT4 and Gemini/Bard with another 100+ Billion element large language model.
The primarily thing Apple has to do is deploy something better than Siri, not some OpenAI 'killer' chatbot.
There is very real chance what Apple is going to deploy is going to run local; not on servers. Several reasons.
i. They are already done that with Siri. Siri is relatively 'poor' primarily for other reasons than the size of the model. Siri gets confused and just doesn't a Google search and tells you to sort it out "here's what I found on the web".
ii. Deploying a 'acre' of Nvidia GPUs would suck up gobs of 'green power' that they don't have access to. Nevermind they don't have the data centers enclosures for that anyway.
iii. Apple doesn't have to do as much data siloing for inference in their data centers. Decent chance here that the objective isn't to 'hoover up' the maximum number of inference queries and associated data.
If Apple runs the AI on end users electricity then it has no impact on Apple's power draw requirements for inference. Nor does it has much impact on capital equipment spend. Again the end users funded all of the inference compute resource requirements.
But, but, but Apple's chapbot won't kill GPT4/Gemini/etc. So what. Are those bots going to be banned from Apple devices. Nope. So the end users lots access how. All of the 'bots' that are boat anchored to hyper expensive servers are going to have to try to pull revenue someone from Apple devices. Apple doesn't have to duplicate everything that everyone else does.
"no decent AI" can run on a phone. Not really:
This guy is is running some reasonable sized models on a Pi 5. Are the results better than GPT4? incrementally. Better than Siri in many cases? Much bigger margin. The latter is the 'problem' that Apple needs to solve sooner.
It is about inference, not training. Apple will need to buy more training hardware , but don't necessarily need to deploy symmetrically on the exact same equipment for inference.
[ Similar with the DLSR versus iPhone cameras. The "good enough" camera you have with you all the time has distinct advantages over the bigger , better camera you have to go get. ]
I suspect the 'hallucination' problem gets more tractable also when not trying to build a "does everything for everybody" AI. Build an expert portrait photography touch up versus build an AI that creates photos from scratch. Grounding the AI in the context of a specific problem means doesn't have blend major concepts that don't belong together ( or get confused trying to blend two conflicting narratives.) . for example the advance editing that Google Photos does on Pixel phones versus what Apple does. Which one is further behind?
A long time ago Apple bought a Cray to support computer design work ( while Seymour Cray bought a mac to do some Cray work. LOL. ). It doesn't matter. It may have changed but Apple was running their supply chain on SAP software on a large Sun box running solaris.
Apple doesn't "has to" buy x86 to run AI cards.
" ... NVIDIA unveiled the NVIDIA Arm HPC Developer Kit to support scientific computing amid the growing need for energy-efficient supercomputers and data centers. It includes an Ampere® Altra® CPU, with 80 Arm Neoverse cores running up to 3.3GHz; dual NVIDIA A100 GPUs, each delivering 312 teraflops of FP16 deep learning performance, as well as two NVIDIA BlueField-2® DPUs, which accelerate networking, storage and security. ..."
NVIDIA and Partners Collaborate on Arm Computing for Cloud, HPC, Edge, PC
GTC -- NVIDIA today announced a series of collaborations that combine NVIDIA GPUs and software with Arm®-based CPUs — extending the benefits of Arm’s...
[ I think AMD is slacking a bit for Linux on Arm support for MI300. Given the MI300A has x86 cores embedded into the package , that isn't too surprising. ]
Microsoft is running a fair amount of the OpenAI workload on Cobalt100 ( Arm) and Maia ( custom Inference )
hardware. No Nvidia , SuperMicro boards , or x86 there at all.
With a systems approach to chips, Microsoft aims to tailor everything ‘from silicon to service’ to meet AI demand - Source
Microsoft unveils two custom chips, new industry partnerships and a systems approach to Azure hardware optimized for internal and customer workloads
news.microsoft.com
Where Apple is going to suffer is where "big inference" comes down to the workstation level. If something like "Sora on a workstation" takes off on local hardware Apple would be caught with their pants down with their "All 3rd party inference is 'evil'" stance. Reports are that the Blackwell B200 'card' is pushing Dell to deal with 1000W modules. If that is the track 'big inference' is on, then Apple probably made the right move.
The wild card is on what "reasonable 'large' sized" inference will be able to do in 1-3 years. I think Apple has missed the boat there, but it will take a few more years to bite them in the butt.
Here's the actual quote from Parfitt:
"One thing that caught my eye is that the maximum RAM on it is 192 GB. And for most applications, that's plenty. But I can tell you as a composer here writing orchestral kind of music with large sample libraries, when I have everything open, and Logic is running, Vienna Ensemble Pro's running, Pro Tools is up, everything is just ready to go, my resting template now is over 300 gigs." [ The 2023 Apple Mac Pro with Apple Silicon - Thoughts from a Pro - YouTube ]
So wouldn't it be more accurate to say this is uncommon rather than "absolutely pathological"? The difference being that pathological implies someone is using the machine in a strange or otherwise non-standard way that creates an edge case, while uncommon means someone using it in a standard way, but for a task that is less common (orchestral music rather than popular music).
And that's precisely what that top RAM option is supposed to be for--those uncommon tasks that need that much RAM.
Last edited:
Another uncommon use is from Lunar Animation, that I still remember. Exactly 3 years since they wrote a story about it.
But I do not know if they still use the Mac Pro 7.1
Mac Pro first year in the studio
But I do not know if they still use the Mac Pro 7.1
Mac Pro first year in the studio
Seeing this use case and a few others that hav been posted around makes me hope Apple focusses more on RAM capacity soon.
I think Apple will be able to hit 256 GB if they do nothing more than double the M3 Max this year. Hopefully they go further and find a way to push to higher RAM ceilings because I believe that is a larger block than any compute performance limits that Apple Silicon may have.
Don't forget the developer, virtualization and containerization use case for the Mac Pro. I _always_ had top-of-the-line Mac workstations, until the 2013 Mac Pro came out. I do now own a 2013 MP, but it's for collection value rather than daily use. The 2019 Mac Pro would work for me, but is far, far too expensive for a dead-end machine. Instead, I have a similar-performace Ryzen-based Linux tower.
The very low RAM ceiling and lack of standard GPU support preclude using the current Mac Pro for many use cases, including:
- CAD
- virtualization/containerization
- large-scale rendering (video, etc.)
Obviously, Apple is doing just fine with the current M-series ecosystem, and doesn't have to change anything. But, a lot of *nix geeks, like me, who otherwise love the Mac, are currently using other platforms because of the fact that Apple doesn't any longer serve our markets. It makes me sad, too, as I have been a Mac user since 1984, when my mom first brought a 128k Mac home, and I fell in love with MacPaint and MacDraw.
Being a life-long *nix lover, I was immediately smitten with Mac OS X, and ran it on my clamshell iBook and then on my first brand-new Mac, a TiPb G4/667. I continued to use Macs as my primary desktops and portables until the retina MacBook Pro of 2012, which was my last, since I couldn't expand the RAM. I replaced it with a ThinkPad when it was retired.
I keep hoping that the Apple of today will throw a bone to folks like me, but I do think that it's rather unlikely. Don't get me wrong, I still use an iPad Pro, an Apple Watch, and an iPhone, but the desktop Mac just isn't suitable for my everyday work.
- Alex
The very low RAM ceiling and lack of standard GPU support preclude using the current Mac Pro for many use cases, including:
- CAD
- virtualization/containerization
- large-scale rendering (video, etc.)
Obviously, Apple is doing just fine with the current M-series ecosystem, and doesn't have to change anything. But, a lot of *nix geeks, like me, who otherwise love the Mac, are currently using other platforms because of the fact that Apple doesn't any longer serve our markets. It makes me sad, too, as I have been a Mac user since 1984, when my mom first brought a 128k Mac home, and I fell in love with MacPaint and MacDraw.
Being a life-long *nix lover, I was immediately smitten with Mac OS X, and ran it on my clamshell iBook and then on my first brand-new Mac, a TiPb G4/667. I continued to use Macs as my primary desktops and portables until the retina MacBook Pro of 2012, which was my last, since I couldn't expand the RAM. I replaced it with a ThinkPad when it was retired.
I keep hoping that the Apple of today will throw a bone to folks like me, but I do think that it's rather unlikely. Don't get me wrong, I still use an iPad Pro, an Apple Watch, and an iPhone, but the desktop Mac just isn't suitable for my everyday work.
- Alex
Last edited:
Just out of curiosity, how much RAM do you need? How much do you have in your Linux tower?
Since you describe 192 GB (the MP's current ceiling) as being "very low", I assume it's quite a bit more than that.
The Ultra uses 8 RAM modules. Once Apple moves to LPDDR5x, it should be able to offer up to 512 GB (8 modules x 64 GB/module) with the Ultra chip. And if they ever make a 4x Max "Extreme" (I don't know if they will), that could offer 1 TB RAM.
We can see this with the NVIDIA Grace-Hopper superchip, which uses 8 LPDDR5x modules to obtain 480 GB RAM, which would be 512 GB (64 GB/module) after adding back what's set aside for ECC.
Of course, that doesn't address your GPU needs.
Last edited:
I have 512gb of RAM in most of my servers, the most up-to-date of which is a Dell R440 with 2x Xeon Gold CPUs and 8TB worth of 1TB SATA SSDs. That is the machine that does the work that my last Mac Pro, a 2009 with a pair of 8-core 2.66ghz Xeons and 64GB of RAM. My current desktop is a Ryzen 7 5700 that maxes out at 256gb of RAM.
I do a _lot_ of virtualization (and containerization) for infrastructure testing, as I am a Linux systems architect, by trade.
The current Mac Pro's CPU and (mostly) GPU would serve my needs well enough, except for the difference in cost, when compared to my 2009. What does not serve is the low RAM ceiling, and the lack of expandable storage. These both drive up initial cost, and make the machine age prematurely, as storage and memory drop quickly in cost over time.
I only mentioned the GPU issue because peers of mine would like to do vectorizable calculations (ML, etc.) on their desktops, and the lack of GPU expansion precludes that.
In comparison, I purchased a 2006 Mac Pro 2.66 ghz, as a refurb, from Apple. It came with a single 250gb SATA disk and 8gb of RAM. I expanded it significantly, as prices dipped, and then traded it for the significantly more powerful 2008, which I kept until 2016.
Cost is a significant factor, of course. Here is approximately what I spent for my desktops:
$2000 2006 Mac Pro 2.66 2x 2-core 8gb/250gb (refurb)
$1000 2008 Mac Pro 2.8 2.66 2x 4-core 16gb/250gb (purchased used)
That first Mac pro would be about $4500, now, and the current Mac Pro is about $6000 for the cheapest model available as an Apple refurb, which doesn't have enough RAM or storage to be useful. My minimal useful config would be 256gb RAM and 4tb of SSD, which is about $8200 for the 192gb/4tb Mac Pro (refurb). Based on Apple's pricing, I'd guess that a theoretical 256gb Mac Pro would be about $9-10,000.
The Dell R440 that currently does my old Mac Pro's work was about $1000 used, including the upgrade to 512gb of RAM and the 8x 1tb SATA SSDs. It's not a perfect replacement, of course, since it lacks nvme storage, but I can install nvme modules, should I wish to speed up the disks.
Please don't misunderstand me; I am not angry with Apple, and I do not think that they are fools. I just wish that they still had an interest in my market segment, that of the computer engineer.
- Alex
I do a _lot_ of virtualization (and containerization) for infrastructure testing, as I am a Linux systems architect, by trade.
The current Mac Pro's CPU and (mostly) GPU would serve my needs well enough, except for the difference in cost, when compared to my 2009. What does not serve is the low RAM ceiling, and the lack of expandable storage. These both drive up initial cost, and make the machine age prematurely, as storage and memory drop quickly in cost over time.
I only mentioned the GPU issue because peers of mine would like to do vectorizable calculations (ML, etc.) on their desktops, and the lack of GPU expansion precludes that.
In comparison, I purchased a 2006 Mac Pro 2.66 ghz, as a refurb, from Apple. It came with a single 250gb SATA disk and 8gb of RAM. I expanded it significantly, as prices dipped, and then traded it for the significantly more powerful 2008, which I kept until 2016.
Cost is a significant factor, of course. Here is approximately what I spent for my desktops:
$2000 2006 Mac Pro 2.66 2x 2-core 8gb/250gb (refurb)
$1000 2008 Mac Pro 2.8 2.66 2x 4-core 16gb/250gb (purchased used)
That first Mac pro would be about $4500, now, and the current Mac Pro is about $6000 for the cheapest model available as an Apple refurb, which doesn't have enough RAM or storage to be useful. My minimal useful config would be 256gb RAM and 4tb of SSD, which is about $8200 for the 192gb/4tb Mac Pro (refurb). Based on Apple's pricing, I'd guess that a theoretical 256gb Mac Pro would be about $9-10,000.
The Dell R440 that currently does my old Mac Pro's work was about $1000 used, including the upgrade to 512gb of RAM and the 8x 1tb SATA SSDs. It's not a perfect replacement, of course, since it lacks nvme storage, but I can install nvme modules, should I wish to speed up the disks.
Please don't misunderstand me; I am not angry with Apple, and I do not think that they are fools. I just wish that they still had an interest in my market segment, that of the computer engineer.
- Alex
Last edited:
So if the next generation of MP (if they release one) had 512 GB of LPPDR5x (as I described above), would that meet your needs? The PCIe slots do provide expandable fast storage. Plus there's an internal SATA connection for installing an HDD.
Or do you need the option to go beyond 512 GB RAM as your needs increase?
Yes, that amount of RAM would indeed serve, but I doubt that I could afford it. 
I would indeed always prefer to have modular RAM, as I will likely need multi-terabyte RAM configs in the future. I don't need top-of-the-line processing power, but I do need lots of RAM and lots of storage for my work.
- Alex
I would indeed always prefer to have modular RAM, as I will likely need multi-terabyte RAM configs in the future. I don't need top-of-the-line processing power, but I do need lots of RAM and lots of storage for my work.
- Alex
It sounds like you need a very specific tool for the job, one that definitely is more server than typical workstation territory. It remains to be seen if Apple is interested in servicing that market. Apple Silicon model does have a lot of advantages, but it also has drawbacks.