Am I the only one being skeptical about a new Mac Pro with M2 Extreme?

[jujoje]

And yet in an video, Apple was showing off Maya’s viewport performance on AS
If Sidefx comes to AS (preview builds available already), rest assured Autodesk won’t be far behind.

I begin to suspect that Maya will never get optimised for Arm: the codebase is a sprawling mess by all accounts, going back to the Alias days, Autodesk don't care much for Maya and less for the Mac. There's been very little innovation in Maya for a long time, just enough to justify the subscription. If it runs through Rosetta it’s probably good enough for them.

I think the Foundry have similar problems with Nuke and Mari - both software they bought rather than developed.

Tbh those are the two main holdouts in terms of programmes for 3D on apple silicon. Perhaps they’ll prove me wrong though and we’ll see them demo’d on a shiny new mac pro.

 

[singhs.apps]

I begin to suspect that Maya will never get optimised for Arm: the codebase is a sprawling mess by all accounts, going back to the Alias days, Autodesk don't care much for Maya and less for the Mac. There's been very little innovation in Maya for a long time, just enough to justify the subscription. If it runs through Rosetta it’s probably good enough for them.

I think the Foundry have similar problems with Nuke and Mari - both software they bought rather than developed.

Tbh those are the two main holdouts in terms of programmes for 3D on apple silicon. Perhaps they’ll prove me wrong though and we’ll see them demo’d on a shiny new mac pro.

If anything, 3D is moving far beyond its traditional haunts. Those that survived the race to the bottom have all the incentive now to push their solutions.

Maya as a brand is right up their with the most recognisable cg apps. Not sure Autodesk is going down without a fight.

Reg foundry… I don’t know what to say.

 

[l0stl0rd]

In video, it's already outperform 3090. Even in 3D, M1 Ultra is 3090's level.

I don’t think in every case… have have done some rendering of the same scene and the 3080ti was 10-12 times faster then the 32 core M1 max. I bet it still would have been 5-6 times faster then the 64 core Ultra.

 

[jujoje]

If anything, 3D is moving far beyond its traditional haunts. Those that survived the race to the bottom have all the incentive now to push their solutions.

Maya as a brand is right up their with the most recognisable cg apps. Not sure Autodesk is going down without a fight.

You might be on to something therethere; it does feel like we're on the cusp of significant changes in the industry, from ML working it's was into everything from animation to rendering to fx, to physics based animation and all kinds of cross pollination between different areas. Feels like there's a lot of cool things popping up lately and new areas up for grabs.

I can't see Maya going anywhere; it's still pretty much the standard for animation (and to a lesser extent layout and modelling) with entire pipelines built around it. Out of curiosity had a look at the Maya 2023 update and it was surprisingly decent, definitely one of the better updates for a while. Perhaps you're right and Autodesk are starting to developing it a bit more seriously. Still a bit dubious about a AS port, but fingers crossed.

 

[deconstruct60]

I begin to suspect that Maya will never get optimised for Arm: the codebase is a sprawling mess by all accounts, going back to the Alias days, Autodesk don't care much for Maya and less for the Mac. There's been very little innovation in Maya for a long time, just enough to justify the subscription. If it runs through Rosetta it’s probably good enough for them.

Primarily, all Rosetta2 does statically compile the Intel binaries into a 'hidden' Arm binary the first time you run the applications. ( there is some corner case stuff that does some dynamic compiles when necessary, but that also is mostly just a compile). There is some overhead/steps to set up an XCode project to compile an Arm build but it isn't a huge project. If the Rosetta compiler is working in vast majority of cases the source code to arm binary compiler is going to work to. ( If there is some 'too clever for its own good' assembler code then perhaps have problems, but non perverse c/c++ code should work. ).

The system requirements for Maya say it only requires SSE4.2 . If there is AVX code in Maya that would be problematical. The other hiccup is all the 'plug ins' have to be switched over also ( there is no mixed-mode with Rosetta2 ).

The bigger Apple Silicon optimization problem that Maya likely has is not the Arm "half" of the SoC, but the GPU 'half'. Mismatch on the GPU optimizations if only trying to talk to the Apple GPU through the now 'frozen in time' OpenGL (and OpenGL shader language code ) . ( which Apple could also just nuke in 4-6 years ). That is probably where Maya is likely leaving performance on the table. Once they have gone through the Rosetta2 translator it is running native code. (so if arm core is clocked faster and has a faster cache properties the code will run faster.) However, frozen-in-time OpenGL code knows nothing about new Metal 3 features that dramatically better leverage tile memory caching, new ray tracing data structures , or etc, etc. All the new GPU driver features rolled out over the last 2 years they don't have direct access to because sitting behind a deprecated API.


The non Metal exposure impacts them on the macOS on Intel side also because there are still some AMD Metal optimizations trickling. Maya really isn't going to get those either even pre-Rosetta2 status.

More likely Maya developers need to find some localized modules that are a computational choke point and possibly just convert a portion so that the Rosetta2 compiler can get a 'better' handle on translating a more modern Intel/Metal chunk of code into something that will run faster. Probably will run faster on an Intel Mac also. That won't be a 'cure all' since pretty several of the computational bottlenecks are all in OpenGL shader code littered with assumptions it is talking to a 5-12 year old AMD/Nvidia GPU .


the other major issues is all the plug-ins and wrap around infrastructure apps. The bigger the ecosystem the slower the migration is likely going to be.

I think the Foundry have similar problems with Nuke and Mari - both software they bought rather than developed.

If don't have the folks who wrote/know the code what apps tend to drift into is new people either spending most of their time re-writing the old stuff the way they would have done it or tip-toe around making making changes largely at the edges where can control interaction with the 'mystery' main code base.

Have Apple come in and say " go rewrite major portions of all your code" and that is typically a disconnect.

Tbh those are the two main holdouts in terms of programmes for 3D on apple silicon. Perhaps they’ll prove me wrong though and we’ll see them demo’d on a shiny new mac pro.

If Apple is looking to show off performance more likely they'd pick an app with a 'native' Metal stack with all the optimization recommendation they made at WWDC 2022 rather than something with Arm code . The latter really isn't hard if it can transit through Rosetta 2.

 

[deconstruct60]

My point is that UltraFusion is potentially unnecessary, and potentially unwanted for a lot of large tasks like rendering or ML.

Batch job 'final' very high resolution rendering , sure. Interactive rendering not so much. Part of the problem with these thread discussions is that sometimes there is very rapid flip-flopping between those two and they have substantively different requirements.

Similar to ML training versus ML inference.


Hard real time requirements bring their own problems when put longer latencies between nodes. There have been many multiple CPU package systems over the last 3 decades. There has very few. non-quirky, highest performance single screen drivers spread over multiple packages over the same time.

 

[DearthnVader]

The Apple SoC Mac Pro won't use a M2 chip, it will be a completely new SoC that supports Slotted ECC RAM and discrete graphics.

Otherwise there will be no reason to build it and Apple can just continue to update the Mac Studio line.

 

[EugW]

The Apple SoC Mac Pro won't use a M2 chip, it will be a completely new SoC that supports Slotted ECC RAM and discrete graphics.

Otherwise there will be no reason to build it and Apple can just continue to update the Mac Studio line.

That's roughly what I was thinking, basically that a new Mac Pro would be based off the same cores, but with a different SoC architecture to accommodate Mac Pro specific needs.

However, we could very well be totally wrong. One thing that some people may have lost sight of is the fact that a large part of the Intel Mac Pro customer base have moved on to Mac Studio. The problem in the past was that the Mac mini and iMac were not enough, and the iMac Pro was a dead end. So, they got the Mac Pro. However, Mac Studio is much more powerful than the (lower end) Intel Mac Pro for many of these users' use cases, and Mac Studio is all they'll ever need. That means the potential AS Mac Pro customer base is that much smaller. However, I'm not entirely sure how this will affect AS Mac Pro's design.

 

[DearthnVader]

That's roughly what I was thinking, basically that a new Mac Pro would be based off the same cores, but with a different SoC architecture to accommodate Mac Pro specific needs.

However, we could very well be totally wrong. One thing that some people may have lost sight of is the fact that a large part of the Intel Mac Pro customer base have moved on to Mac Studio. The problem in the past was that the Mac mini and iMac were not enough, and the iMac Pro was a dead end. So, they got the Mac Pro. However, Mac Studio is much more powerful than the (lower end) Intel Mac Pro for many of these users' use cases, and Mac Studio is all they'll ever need. That means the potential AS Mac Pro customer base is that much smaller. However, I'm not entirely sure how this will affect AS Mac Pro's design.

Just my 2 cents, Apple is likely working on a new SoC for the Pro line, and like I say, it must support ECC RAM and discrete graphics.

They can free up some die "space" without a ton of integrated GPU cores and they are not limited to the thermals of the Studio or the Mini.

The Studio may have been a stop gap, but if it sells very well and Apple can't make a SoC that scales well into a Pro type CPU, then it maybe a very long time before we ever see the SoC Mac Pro.

Apple workstations are a very limited market and those of us that would buy one have our own list of what it must include. Apple knows this, but can they make enough money off such a system?

The SoC MP may end up being vaporware.........

 

[deconstruct60]

My point is that UltraFusion is potentially unnecessary, and potentially unwanted for a lot of large tasks like rendering or ML.

Batch job 'final' very high resolution rendering , sure. Interactive rendering not so much. Part of the problem with these thread discussions is that sometimes there is very rapid flip-flopping between those two and they have substantively different requirements.

Similar to ML training versus ML inference.


Hard real time requirements bring their own problems when put longer latencies between nodes. There have been many multiple CPU package systems over the last 3 decades. There has very few. non-quirky, highest performance single screen drivers spread over multiple packages over the same time.

The Apple SoC Mac Pro won't use a M2 chip, it will be a completely new SoC that supports Slotted ECC RAM and discrete graphics.

Otherwise there will be no reason to build it and Apple can just continue to update the Mac Studio line.

Actually is the the opposite. There is little to no economic reason to decouple the R&D spend of the Mac Pro from the rest of the Mac (and iPadOS ) line up. If too expensive to make then that will be a product inhibitor.

A major constraint the Mac Pro has is that it is realtive very low volume. There are hyper modular advocates , but they collectively don't relatively buy much in the aggregate. The Mac Pro is very low single digits of the Mac unit sales. The danger is that it sinks so low that it rounds to zero.

If it is an SoC that is so expensive 'nobody' wants to buy it then you pragmatically don't have a product either.


Studio has a Apple self imposed constraint that it is a literal desktop system. It has to fit inside of a 7 inch x 7 inche square. Remove that constraint and add some modest PCI-e v4 lane provisioning to the SoC and Apple can make a substantively different product for a different product audience. The don't necessarily have to make a Dell/HP super bulky workstation 'killer'. Just something different enough to separate from the Studio target market.

 

[CasualFanboy]

Now more than ever, we no longer need a desktop computer to have multiple graphics cards or terabytes of memory because of virtualization and cloud based compute resources. If you truly need absurd computation you can just run it remotely, you don't need 4 graphics cards in your desktop.

"You will own nothing and you will love it."

 

[alex00100]

"You will own nothing and you will love it."

Something tells me you don’t work in an industry that requires it. Would you prefer to buy a $5k pro workstation that would still take days to render your scene and be unavailable during that time. And when you don’t need the power it’s just collecting dust. Or would you prefer to render your whole scene in one hour remotely while you work on a gaming laptop?
Remote computing power is absolutely the future.

 

[deconstruct60]

"You will own nothing and you will love it."

Being in the cloud doesn't necessarily mean that your organization doesn't own it. Or that you are running multi-tentant on a system.

Right tool for the right problem. 'Never buy' or 'never lease' are both sweeping generalizations don't particular address specific tactical situations.

 

[Xiao_Xi]

Would you prefer to buy a $5k pro workstation that would still take days to render your scene and be unavailable during that time. And when you don’t need the power it’s just collecting dust. Or would you prefer to render your whole scene in one hour remotely while you work on a gaming laptop?
Remote computing power is absolutely the future.

Workstations such as the Mac Pro are good for intensive interactive tasks, and cloud computing for long non-interactive tasks.

 

[jmho]

The only problem with cloud computing is that it's great for established professionals, but it sucks for newbies. I can't imagine being a student trying to learn 3D at home if the only option is to pay $50 every time you want to render the janky poorly textured bobble-head you're working on.

Obviously the Mac Pro is out of financial reach for almost all students, but this stuff trickles down to the cheaper SKUs, and I really hope that Apple is able to make the Mac Pro a local 3D powerhouse so the tech / APIs / software can trickle down to the consumer models.

 

[Fomalhaut]

Something tells me you don’t work in an industry that requires it. Would you prefer to buy a $5k pro workstation that would still take days to render your scene and be unavailable during that time. And when you don’t need the power it’s just collecting dust. Or would you prefer to render your whole scene in one hour remotely while you work on a gaming laptop?
Remote computing power is absolutely the future.

It depends on how much data you need to upload and download to & from the cloud service, and how long this will take. Most internet connections are 20-100x slower than NVMe local storage bandwidth . If the rendering models are relatively small and don’t require a lot of communication between client and server processes, then remote computing works well.

 

[innerproduct]

Please do remember that the iMac Pro of 2017 has almost exactly the same rendering speed in GPU renderers as a full m1 max of a 2022 Mac Studio. With the screen (that was included in the iMac Pro package) we basically pay exactly the same as 2017 for the same performance as 2017 but in 2022. That is beyond pathetic. The iMac pro could connect up to 4 eGPUS as well for extra boost. the Vega 64 that was inside the iMac Pro had similar perf as the 1080 ti if utilised correctly. But since those days nvidia has doubled perf 3 times. 1080ti-->2080ti->3090ti and soon again with the 4090. Of course pros use cloud or their own farms, but we need decent local perf as well. 50% of the PC competition perf would be ok (that is rendering an image in 1 min on a Mac that takes 30 s on a PC) But having something that is close to a magnitude worse, there is just no way to justify a Mac. So please Apple, fix this horrible GPU-situation and make sure that you at least can match your own old 2019 Mac Pro on actual (not advertised) performance in 3D. (we all know already the m1 max is fast enough for making YouTube videos).

 

[jmho]

The main problem is software. It's great that Apple is finally sending devs to port Blender to Metal, but they need to be hiring, training, and sending Metal specialists to literally every large 3D software developer, for free, and saying "Hey, we'll help you port your stuff (properly)"

The M1 GPUs are incredibly powerful, but a) what's good for an nVidia GPU is usually terrible for an M1 GPU, b) almost nobody knows Metal, c) Apples documentation and sample code for Metal is garbage, d) Apple won't help you and you can't talk to any of their devs.

This is a huge problem because even if Apple release a Mac Pro that can actually go toe to toe with a 4090 and has ray tracing hardware, nobody is going to make use of it.

Compare that with nVidia who release an absolute ton of resources including free books, courses on Youtube etc. Microsoft has great docs and samples for DirectX, and Vulkan has a thriving community.

Apple knew that when they asked all their developers to learn Swift instead of a standard language, it was up to them to popularise it and make it easy to learn so people would actually use it, and in many ways they've done that. With Metal they're just all "Oh hey, here's this new API. Have 6 videos a year at WWDC, almost zero documentation, and some terrible sample code written in objective C so it's confusing for both C++ and Swift programmers. Anyway, good luck!"

 

[deconstruct60]

Please do remember that the iMac Pro of 2017 has almost exactly the same rendering speed in GPU renderers as a full m1 max of a 2022 Mac Studio. With the screen (that was included in the iMac Pro package) we basically pay exactly the same as 2017 for the same performance as 2017 but in 2022.

" .. The base iMac Pro model is priced at $4,999 and comes equipped with a 3.2GHz 10-core Intel Xeon W processor with Turbo Boost up to 4.2GHz, 32GB 2666MHz ECC RAM, 1TB SSD storage, and a Radeon Pro Vega 56 with 8GB HBM2 memory. ..."
https://www.macrumors.com/roundup/imac-pro/

[ Note : this is an updated blurb or the config change later as the iMP aged. Originally started off at 8 cores at that price. ]


M1 Max Studio 32GB RAM 1TB SSD 10 cores 16 GPU (top line Max ) --- $2,300

$2,699 cheaper. (about price of an MP 2013 ;-) ... or perhaps shaving off a Studio display price off the iMP to get back to a M1 Max Studio? )


M1 Ultra 64GB 4TB SSD 20 cores 48 GPU ( binned Ultra ) --- $4,999

The Vega 64 on the iMP would another $500 and 14 cores $800. Same range as going full spec Ultra for the studio. ( where down on core count and single thread performance )

The iMac pro could connect up to 4 eGPUS as well for extra boost.

for substantially more money. Can plug a Max Studio via Thunderbolt and get a stacked cluster render farm too for what 4 eGPU boxes would cost.

 

[deconstruct60]

The main problem is software. It's great that Apple is finally sending devs to port Blender to Metal, but they need to be hiring, training, and sending Metal specialists to literally every large 3D software developer, for free, and saying "Hey, we'll help you port your stuff (properly)"

They don't really need to do that. Part of the issues with the M1 Ultra was that the education and tools release was backwards. All the more highly relevant "Use these steps to write better optimized code" and "here are the better tools for more accurate profiling and tracing " came at WWDC 2022 ... around 3 months after they released the system. At some point Apple's developers education efforts have to become effect. Apple can't run around holding everybody's hand who wants to write a Mac app.

Part of the issue here is the dual edge soft of a "low level only", 'thin' APIs like Metal where responsibility is shewed to the app developers to do a heft share of the optimization work. If have widely dispersed efforts then pragmatically going to get widely different optimizations results.

The M1 GPUs are incredibly powerful, but a) what's good for an nVidia GPU is usually terrible for an M1 GPU, b) almost nobody knows Metal, c) Apples documentation and sample code for Metal is garbage, d) Apple won't help you and you can't talk to any of their devs.

Nobody knows Metal? 10's thousands of iOS developers have been walking around in the dark for the last 4-6 years? If sat through most of the WWDC sessions for the last 3-4 years and never saw Metal that was a deliberate choice to avoid it.

As the mac SoC have divated from the iPhone SoCs in scope and complexity I don't think Apple's depth of Metal talent is as deep as the more mature SoC scopes ( Ann and AnnX SoCs ). I wouldn't be surprised if they are mostly just 6-9 months out in front of folks on the outside on Metal maturity on the "never released before" SoCs.



Apple's technical support documents and general support. Yeah there is lots more effort making them look "Pretty" and "trendy" more so than someone looking hard at the content and educational utility.

This is a huge problem because even if Apple release a Mac Pro that can actually go toe to toe with a 4090 and has ray tracing hardware, nobody is going to make use of it.

I don't see Apple saying they are trying to go 'toe-to-toe' with the 4090. Apple has stated they want to be best iGPU, but at no point I can recall did Apple run around claiming they were going to be king of the whole GPU market in the near or intermediate term. Lots of non Apple folks on these forums claim that, but where did Apple claim this?

That is one of the contributing problems that go Intel into trouble. They were going to do super duper enthusiast level graphics . They were skipping over AMD and going to take Nvidia down .


They don't need a 4090 killer GPU to have a viable , profitable product for the M-series second generation. They don't have to get there in one leap any more than the A-series needed to jump to a huge lead in 1-2 initial generations either in that market segment.

As pointed out there are far more deep seated educational issues to sort out. If don't have the right software the hardware isn't going to make a difference all on its own. Nvidia hardware and no signed macOS drivers to run it, is not a pragmatically effective tool.

the Mac Pro appearing last in the transition roll out is likely not an accident. Apple hasn't done anything like that largely by themselves before either. ( for Intel era they have ready Intel reference designs to guide them .)

 

[jmho]

Nobody knows Metal? 10's thousands of iOS developers have been walking around in the dark for the last 4-6 years? If sat through most of the WWDC sessions for the last 3-4 years and never saw Metal that was a deliberate choice to avoid it.

Where is the community then? Where are the tutorials? Where is the code on GitHub?

I've spent the last year working in Metal, and I've probably worked with every Metal feature there is. I've also ported most of the Metal Sample code to Swift and put it up on GitHub.

I'll be incredibly happy if there is a thriving vibrant community of Metal devs that I've overlooked, but it feels pretty sparse from what I've seen.

 

[Kimmo]

Don't ask me how I got this transcript from a recent, high-level meeting on the new Mac Pro, but I thought you guys might find it interesting.

Don't share this with anyone, especially not Gurman.

• Phil Schiller: "Look, I made promises. John made promises. Craig made promises. Pretty soon there won’t be anyplace in Silicon Valley where we can hang our hats.
• John Ternus: "You don't know these crazy Mac Pro owners. They'll whack us, for sure."
• Johny Srouji: "Try and be patient."
• Ternus: "I’m not asking for your help, Johny, just take off the handcuffs. I need to deliver upgradability and GPU power and I need to do it now!"
• Srouji: "Be patient."
• Ternus: "We gotta protect ourselves. At least give me the chance to add some PCIe slots and some powerful GPU's. After the 6,1 fiasco that's what we told the Mac Pro whackadoodles we'd give them. We did it with the 7,1 and can't just take it away in the 8,1!"
• Craig Federighi: "He's right, you know. Ah, I need to step out and comb my hair."
• Srouji: "No! I don’t want Nvidia to think they have the upper hand. Apple Silicon is a miracle, a miracle I tell you!"
• Schiller : "Johny, you're wrong."
• Ternus: "Godfather, you once said the day would come when me and Phil could form our own workstation company where we'd innovate our asses off and give the pro users what they want, terabytes of RAM and big ass GPU's. Until today I would never think of such a thing, but now I must ask your permission."
• Tim Cook: "Well, Johny's head of hardware technologies now and if he gives his permission then you have my blessing."
• Srouji: "After we finish the move to Apple Silicon you can break off from the Mac Family and go off on your own. After we finish the move."
• Ternus: "How long will that take?"
• Srouji: "Six months."
• Schiller: "Forgive me, Godfather, but with Johny in charge, me and John will come under Nvidia's thumb sooner or later."
• Ternus: "And I hate that Goddamn Nvidia. In six months time there won’t be nothin’ left to build on. The Pro users will all be using PC's."
• Cook: "Do you have faith in my judgement?"
• Ternus: "Yes."
• Cook: "Do I have your loyalty?"
• Ternus: "Yes, always Godfather."
• Cook: "Then be a friend to Johny. Do as he says."
• Srouji: "There are developments being made that are going to answer all of your questions and solve all of your problems. That’s all I can tell you right now."

 

[Boil]

Six months, huh...?

So NAB 2023, Apple has a booth pimping FCP/Logic, but the demo computers appear to be 2019 Mac Pro rackmounts, with the rear of the chassis hidden in a deskside rackmount enclosure... Rackmount RAIDs & A/V I/O gear are also in these deskside enclosures, a full production suite set-up basically...? About This Mac tells you to piss off when you try to see what is under the hood...

WWDC 2023 - Heavy focus on 3D & real-time ray-tracing for mixed reality & vfx/games/entertainment, solid overview on the ASi Mac Pro as asset creation tool, teaser overview on Apple Glasses/Goggles/Headset for mixed reality...

Siggraph 2023 - Apple/Blender booth highlighting Full Metal Blender & the ASi Mac Pro (Mac Studios on-site as well), back room suite previews of hardware ray-tracing & real-time ray-tracing in the viewport, third-party software developers only, extensive NDAs signed and blood oaths sworn to for the back room peek-a-boo action...

 

[dawnrazor]

That's roughly what I was thinking, basically that a new Mac Pro would be based off the same cores, but with a different SoC architecture to accommodate Mac Pro specific needs.

However, we could very well be totally wrong. One thing that some people may have lost sight of is the fact that a large part of the Intel Mac Pro customer base have moved on to Mac Studio. The problem in the past was that the Mac mini and iMac were not enough, and the iMac Pro was a dead end. So, they got the Mac Pro. However, Mac Studio is much more powerful than the (lower end) Intel Mac Pro for many of these users' use cases, and Mac Studio is all they'll ever need. That means the potential AS Mac Pro customer base is that much smaller. However, I'm not entirely sure how this will affect AS Mac Pro's design.

This is where I currently sit…. a top spec Mac Studio would serve my needs right now and probably for a year or two, however with no expandability and no way to upgrade I’d be looking at getting something new in about 3 years max… But the new Mac Pro might offer upgrade paths so it’s a bit more future proofed… that said a decent spec MP is going to be a lot more expensive than the top spec MS… so you’ll be paying for that expandability…

I just want to know what the new Mac Pro is going to offer, I’m holding off getting a Mac Studio even though I desperately need a faster computer and am just waiting for them to announce the Mac Pro… if that doesn’t happen until March like some are predicting, then that’ll me buying a MS in December… the good news is that the wait times have dropped significantly as they catch up on orders etc…