Apple Executive Discusses New Mac Pro's Lack of Graphics Card Support

[fluidmodeler]

Simply put, without NVIDIA GPU support, the Apple Silicon Mac Pro is dead in the water for any serious ML/AI endeavors.

I invite anyone fixated on the unified memory architecture to compare M2 Ultra’s GPU compute to the power offered by high-end professional GPUs (e.g., NVIDIA H100) — there’s simply no comparison.

In the absence of NVIDIA GPU support, who is this machine for?

 

[enb141]

Simply put, without NVIDIA GPU support, the Apple Silicon Mac Pro is dead in the water for any serious ML/AI endeavors.

I invite anyone fixated on the unified memory architecture to compare M2 Ultra’s GPU compute to the power offered by high-end professional GPUs (e.g., NVIDIA H100) — there’s simply no comparison.

In the absence of NVIDIA GPU support, who is this machine for?

- People that hate windows or need Logic or Final Cut and need PCI express cards For Audio, Video Capture, Storage or Network.

 

[singhs.apps]

I think everyone complaining about lack of ram needs to post how much they are using now.

That's one using more than 192. No others. Thought so. Thanks.

What is the point of these comments? What do these have to do with Apple’s inability (not intent) to make 192+GB Mx systems at present?

A few months ago the max configuration was 128 GB. Are you suggesting just a few months ago most people didn’t need more than 128 GB? If yes, why did Apple release a 192 GB configuration just a month back?
What prompted to apple to offer an increase ? Did some new magic workflow emerge that forced Apple ?
Did no 192 GB+ system (Mac/pc) exist before June 2023?

Why isn’t Apple selling a 192 GB+ Mx system right now, if it will be win-win - according to you - today?

 

[xbjllb]

Simply put, without NVIDIA GPU support, the Apple Silicon Mac Pro is dead in the water for any serious ML/AI endeavors.

I invite anyone fixated on the unified memory architecture to compare M2 Ultra’s GPU compute to the power offered by high-end professional GPUs (e.g., NVIDIA H100) — there’s simply no comparison.

In the absence of NVIDIA GPU support, who is this machine for?

Me, and I HAVE an NVIDIA Quadro K5000 in my Mid 2010 Mac Pro and use it all the time for rendering.

 

[spaz8]

Simply put, without NVIDIA GPU support, the Apple Silicon Mac Pro is dead in the water for any serious ML/AI endeavors.

I invite anyone fixated on the unified memory architecture to compare M2 Ultra’s GPU compute to the power offered by high-end professional GPUs (e.g., NVIDIA H100) — there’s simply no comparison.

Sadly agreed, in my recent research about if I should personally get a maxed out Mac Studio Max or Ultra.. I found some blogs from Apple enthusiasts doing ML work with the latest pytorch and tensorflow ports to metal vs. Nvidia

The top spec M2 Ultra w 76 core GPU is 2.2x slower (about a 1080 TI at ML) than a 4090 at ML work .. samples per second (663 vs. 1470).. still lots of work to do, but at least Apple has finally 8 yrs? later put a horse in the race. Yes the Mac Studio could throw more memory at the issue.. but I didn't see these testers using bigger batch sizes solving things.. plus training large models that would not fit into the 4090's 24 gb of ram will take Eon's to train.

Through several leaps of M2 Metal scores next to the closest AMD card's metal score.. and then comparing that AMD card to its closest NV card on userbenchmark. The equivalency I came up with for general compute not ML (would be less flattering)..

Mac Studio M2U 76 core = approx RTX 3080 -and about 1080 Ti at ML/AI tasks
Mac Studio M2U 60 core = approx RTX 3070
Mac Studio M2M 38 core = approx RTX 4060 Ti
Mac Studio M2M 30 core = approx RTX 4060

TLDR.. for my needs the diminishing returns of GPU performance at the top end is not worth the $2000 cost in Apple land currently for me. So taking that money and building a dedicated 2nd PC system for that ML work.

 

[martinX]

Simply put, without NVIDIA GPU support, the Apple Silicon Mac Pro is dead in the water for any serious ML/AI endeavors.

I invite anyone fixated on the unified memory architecture to compare M2 Ultra’s GPU compute to the power offered by high-end professional GPUs (e.g., NVIDIA H100) — there’s simply no comparison.

In the absence of NVIDIA GPU support, who is this machine for?

To be fair, I think the H100 is a little more costly.

 

[Patroclus]

A lot of people don’t understand Apple M chip architecture. They think that memory is used the same way as before, in Intel machine. It is not. This is why Apple limits the memory in these new M machines. Don’t you think they do benchmarks in their labs with all the most used professional software?
This machine is not for the regular user. These are for professional studios that deals with large amount of data.
I’ll take any upgrade that takes 1 day to generate whatever to 3 days. (3x faster).

 

[Patroclus]

Sadly agreed, in my recent research about if I should personally get a maxed out Mac Studio Max or Ultra.. I found some blogs from Apple enthusiasts doing ML work with the latest pytorch and tensorflow ports to metal vs. Nvidia

The top spec M2 Ultra w 76 core GPU is 2.2x slower (about a 1080 TI at ML) than a 4090 at ML work .. samples per second (663 vs. 1470).. still lots of work to do, but at least Apple has finally 8 yrs? later put a horse in the race. Yes the Mac Studio could throw more memory at the issue.. but I didn't see these testers using bigger batch sizes solving things.. plus training large models that would not fit into the 4090's 24 gb of ram will take Eon's to train.

Through several leaps of M2 Metal scores next to the closest AMD card's metal score.. and then comparing that AMD card to its closest NV card on userbenchmark. The equivalency I came up with for general compute not ML (would be less flattering)..

Mac Studio M2U 76 core = approx RTX 3080 -and about 1080 Ti at ML/AI tasks
Mac Studio M2U 60 core = approx RTX 3070
Mac Studio M2M 38 core = approx RTX 4060 Ti
Mac Studio M2M 30 core = approx RTX 4060

TLDR.. for my needs the diminishing returns of GPU performance at the top end is not worth the $2000 cost in Apple land currently for me. So taking that money and building a dedicated 2nd PC system for that ML work.

The Mac Pro is not meant for AI. AI applications run on servers which you can spin the Cloud (AWS, Azura, Digital OCean) where they have dedicated GPU intensive configured machine.

 

[Patroclus]

Top 5 Companies, Worldwide PC Workstation Shipments, Market Share, and Year-Over-Year Growth, 2022 (shipments in thousands of units)

Company	2022 Shipments	2022 Market Share	2021 Shipments	2021 Market Share	2022/2021 Growth
1. Dell Technologies	3,171.2	41.4%	2,979.6	39.8%	+6.4%
2. HP Inc.	2,580.4	33.7%	2,549.3	34.0%	+1.2%
3. Lenovo	1,860.0	24.3%	1,920.9	25.6%	-3.2%
4. ASUS	24.5	0.3%	19.7	0.3%	+24.3%
5. NEC	20.1	0.3%	26.1	0.3%	-22.7%
Total	7,656.2	100.0%	7,495.6	100.0%	+2.1%

Source: https://www.idc.com/getdoc.jsp?containerId=prUS50454823

Mac Pro is not meant to eat these PC workstation market share. It’s saturated anyway.
iphone is Apple butter, along with iWatch, Ipad and services.

 

[Patroclus]

Seems like a weird decision, at least it can still grate thicc mozzarella shreds.

I’ll take a Mac Pro over an ugly PC workstation with lights inside, water coolant, 10 fans ,and a disco ball.
These new Mac Pro M chip probably smoke PC workstation anyway (for the same config price).

 

[ilikewhey]

These new Mac Pro M chip probably smoke PC workstation anyway (for the same config price).

lol, it certainly does not, do you even have benchmark figures to back that up?

 

[ilikewhey]

A lot of people don’t understand Apple M chip architecture. They think that memory is used the same way as before, in Intel machine. It is not. This is why Apple limits the memory in these new M machines. Don’t you think they do benchmarks in their labs with all the most used professional software?
This machine is not for the regular user. These are for professional studios that deals with large amount of data.
I’ll take any upgrade that takes 1 day to generate whatever to 3 days. (3x faster).

yeah we get it its a shared memory architecture. but guess what, 10 tabs on chrome in ram usage is still 10 tabs on chrome rin ram usage, regardless if its m2 or ryzen 5800.

 

[ilikewhey]

I think everyone complaining about lack of ram needs to post how much they are using now.

 

[enb141]

I’ll take a Mac Pro over an ugly PC workstation with lights inside, water coolant, 10 fans ,and a disco ball.
These new Mac Pro M chip probably smoke PC workstation anyway (for the same config price).

Just because the Mac Pro has Apple brand name doesn't means will smoke a PC, actually the Mac Pro is being smoked by an Intel Core i9 13 Gen with an nvidia 4090 for a way cheaper price.

As the other guy mentioned, Chrome or any browser doesn't cares if RAM is shared or not.

 

[-BigMac-]

Mac Pro is not meant to eat these PC workstation market share. It’s saturated anyway.
iphone is Apple butter, along with iWatch, Ipad and services.

What is an iWatch

 

[cookedart]

A lot of people don’t understand Apple M chip architecture. They think that memory is used the same way as before, in Intel machine. It is not. This is why Apple limits the memory in these new M machines. Don’t you think they do benchmarks in their labs with all the most used professional software?
This machine is not for the regular user. These are for professional studios that deals with large amount of data.
I’ll take any upgrade that takes 1 day to generate whatever to 3 days. (3x faster).

Many here understand the M series architecture just fine. Apple is forced to limit the maximum memory because it cannot be upgraded after the fact, and would be difficult to sell in a very large memory configuration (i.e. 1TB or larger). Because of this, they have effectively bowed out of the market that would find this kind of memory usage useful.

That being said, they would for sure be ahead in graphics-heavy workloads specifically where the large amounts of addressable GPU memory would be an advantage - in theory their GPU can access up to 192 GB of VRAM vs the largest Quadro cards would cap at 48GB. Very large simulations come to mind as being a potential advantage here, but I think this specific use case would be very, very niche.

However I think losing the flexibility of adding more overall system memory past 192GB makes this a less viable workstation overall. Essentially Arm-based Mac OS computers are capped at 192GB , so developers have no reason to target and test against anything larger than this, or to release software or features that can address larger amounts of memory. Add to the fact that Metal and the M series GPUs are hardly industry standard in creative industries and Apple needs to do a lot of convincing to bring people over, and I don't really see that happening with lineup they have.

We also have seen the benchmarks of these machines, and unfortunately, especially in the GPU front, they do not impress compared to the other workstations available on the market, especially in the price range. At least in many instances of the Mac Pro before, when Apple released a new version, it would be competitive or a bargain compared to the competition - but in this case the value proposition skews heavily towards a Mac Studio which is even less flexible in a production environment. Users who need the PCI expansion are being punished/taxed for needing something critical to their job, which in my opinion is not a great way to get people on board with their platform.

 

[SFjohn]

Just because the Mac Pro has Apple brand name doesn't means will smoke a PC, actually the Mac Pro is being smoked by an Intel Core i9 13 Gen with an nvidia 4090 for a way cheaper price.

As the other guy mentioned, Chrome or any browser doesn't cares if RAM is shared or not.

It depends on the software you’re using, Final Cut doesn’t run on a PC…

 

[H2SO4]

This time it’s a software limitation, not hardware

Regardless of what the limitation was people voted with their wallets.

 

[H2SO4]

Good luck…
Apple is going after bigger fish …

Like what exactly?

 

[Velli]

View attachment 2260043

I have asked a few times now what actually happens if you try to perform one of those tasks that people claim ise more RAM than Apple Silicon supports, on an actual Apple Silicon computer. I have documents that will not load on my 16 GB Windows work computer, with an error message about something something memory. Other documents work, but are sluggish and RAM is maxed out. While the same documents not only load, but works fluently on my 8 GB M1 Macbook Air.

I am simply curious if the same situation applies to larger workloads, but so far noone has provided a real life example. Only theories, or like yours, stats for memory usage, rather than tales about the actual experience of one vs the other.

 

[ilikewhey]

I have asked a few times now what actually happens if you try to perform one of those tasks that people claim ise more RAM than Apple Silicon supports, on an actual Apple Silicon computer. I have documents that will not load on my 16 GB Windows work computer, with an error message about something something memory. Other documents work, but are sluggish and RAM is maxed out. While the same documents not only load, but works fluently on my 8 GB M1 Macbook Air.

I am simply curious if the same situation applies to larger workloads, but so far noone has provided a real life example. Only theories, or like yours, stats for memory usage, rather than tales about the actual experience of one vs the other.

need more details, what document on what program, what you are experiencing its probably memory swapping, apple uses decently fast ssd so you might not notice it as much vs your work issued computer.

i had a 16gb mba before my 64gb mbp, it pretty much behaved the same when memory start swapping, however, because my mbp ssd is so much faster than the ssd in the m1 mba, the swap operation is alot more fluid.

bare in mind, memory swap is not a good thing, and apple can't just make more ram pop out of thin air, your 8gb ram on the mba is still 8gb of ram.

also, one thing i noticed is window uses alot less ram for the same tasks, when my mbp was in the shop i used my gaming pc, and what normally requires 50gb of ram on my mbp, only took up 30gb on my pc.

 

[Alan Wynn]

Many here understand the M series architecture just fine. Apple is forced to limit the maximum memory because it cannot be upgraded after the fact, and would be difficult to sell in a very large memory configuration (i.e. 1TB or larger).

Do you have any evidence that is why they did not offer larger memory configurations? Why do you think this machine would be any harder to sell than the previous Intel based system with 1.5TB as an option? Apple did not produce a machine that they expected would only be filled with third party RAM.

Because of this, they have effectively bowed out of the market that would find this kind of memory usage useful.

How many machines are sold into this market? What software is used by this market? How much of it runs on macOS? Before they “effectively bowed out” of this market, in 2020 when the 2019 Mac Pro was brand new, how many Mac Pros were sold into this market.

That being said, they would for sure be ahead in graphics-heavy workloads specifically where the large amounts of addressable GPU memory would be an advantage - in theory their GPU can access up to 192 GB of VRAM vs the largest Quadro cards would cap at 48GB. Very large simulations come to mind as being a potential advantage here, but I think this specific use case would be very, very niche.

All systems non-virtualized systems that use over 64GB are very niche. The question is simply is this a profitable niche.

However I think losing the flexibility of adding more overall system memory past 192GB makes this a less viable workstation overall. Essentially Arm-based Mac OS computers are capped at 192GB , so developers have no reason to target and test against anything larger than this, or to release software or features that can address larger amounts of memory. Add to the fact that Metal and the M series GPUs are hardly industry standard in creative industries and Apple needs to do a lot of convincing to bring people over, and I don't really see that happening with lineup they have.

I agree that it take effort to get people to port to macOS. What I have never understood, is that if the best system Apple can build based on Intel or AMD chips cannot really be meaningfully faster than a commodity machine from Dell, HP or Lenovo, why would anyone waste time porting to macOS?

We also have seen the benchmarks of these machines, and unfortunately, especially in the GPU front, they do not impress compared to the other workstations available on the market, especially in the price range. At least in many instances of the Mac Pro before, when Apple released a new version, it would be competitive or a bargain compared to the competition - but in this case the value proposition skews heavily towards a Mac Studio which is even less flexible in a production environment. Users who need the PCI expansion are being punished/taxed for needing something critical to their job, which in my opinion is not a great way to get people on board with their platform.

These machines are substantially more power efficient (something that matter in a state with electricity costs as high as California’s), but are not competitive for every workload. On the other hand, we just benchmarked one of the leading MS Windows based Photogrammetry applications on a Threadripper system with a top of the line nVidia card against Apple’s Object Capture on an M1 Ultra Studio and found that the Mac took one fifth the time to produce better results. As soon as we have our M2 Ultra Mac Pro, I will rerun the benchmarks and post the results.

No one is being ”punished” for needing PCIe expansion slots, but they have become a very small niche of users and as such should expect to pay more. Just as users who need 192GB of RAM do. For most users (even at the high end), a Mac Studio is perfectly sufficient. For those (like us) who need internal expansion slots, paying more is not really a new phenomenon.

 

[Velli]

need more details, what document on what program, what you are experiencing its probably memory swapping, apple uses decently fast ssd so you might not notice it as much vs your work issued computer.

i had a 16gb mba before my 64gb mbp, it pretty much behaved the same when memory start swapping, however, because my mbp ssd is so much faster than the ssd in the m1 mba, the swap operation is alot more fluid.

bare in mind, memory swap is not a good thing, and apple can't just make more ram pop out of thin air, your 8gb ram on the mba is still 8gb of ram.

also, one thing i noticed is window uses alot less ram for the same tasks, when my mbp was in the shop i used my gaming pc, and what normally requires 50gb of ram on my mbp, only took up 30gb on my pc.

You ask me to share which files, yet don’t share specifically about yours… but Ok. In a few scenarios, PDF’s couldn’t open while I had a few PowerPoints and simple Excel files open. Opened twice as much on the Mac, zero issue. I don’t remember file sizes, but none of the files above a few hundred MB.

Usually this is where people tell me my company’s managed Windows install is faulty… but that doesn’t really help me, and doesn’t answer the question.

I’m well aware that there are tough workloads that my MBA is slow for. I’m pointing towards lots of people claiming that they work with files that “require” more RAM than on Intel. You still didn’t answer that - the files used less RAM on Windows, but how was it to work with? A 64 GB computer doesn’t run any slower with 50 GB used than with 30 GB used, so for all we know it can show the opposite of what one may think, that it actually uses the RAM more efficiently (but I acknowledge I am a layman and this is just a theory)

 

[Velli]

Like what exactly?

I disagree with the sentiment. What we see is, Apple plays the long game. We’re unlikely to have seen yet what Apple’s bigger plan for the Apple Silicon is.

 

[cocoua]

need more details, what document on what program, what you are experiencing its probably memory swapping, apple uses decently fast ssd so you might not notice it as much vs your work issued computer.

i had a 16gb mba before my 64gb mbp, it pretty much behaved the same when memory start swapping, however, because my mbp ssd is so much faster than the ssd in the m1 mba, the swap operation is alot more fluid.

bare in mind, memory swap is not a good thing, and apple can't just make more ram pop out of thin air, your 8gb ram on the mba is still 8gb of ram.

also, one thing i noticed is window uses alot less ram for the same tasks, when my mbp was in the shop i used my gaming pc, and what normally requires 50gb of ram on my mbp, only took up 30gb on my pc.

I'm with you about the memory, but there is something I dont understand completely about this, I have a rMBP2012 with 16GB RAM and a MacPro 2013 64GB RAM, I use both for similar task, but the rMBP when working at offices or second house, the swap memory in the laptop is around 8-16GB (would be 32GB in total) while in the MP 0-8GB, similar tasks, same FCP projects, it seems MacOS uses all RAM available eventually and it uses swap no matter the amount once it reaches some point, But the rMBP should show >50GBGB (16+50) swap sometimes as the MP uses more than 64GB RAM, isnt it? or how does it works? or page file is never larger than the physical RAM?


on the other hand, I have a MacStudio Max 32GB RAM, and the workload is much heavier than the MP with multiple 4k streams and AI (topaz video) rendering but the swap never got beyond 14 GB RAM

All 3 machines have at least 200GB of spare SSD for chache files or VR.

I'm a bit lost about how all this works