M4 Max (16-Core CPU, 128GB RAM) vs. M3 Ultra (28-Core CPU, 96GB RAM)

[throAU]

Well, for one, I might be able to allocate just 3–4 CPU cores to my VMs instead of 6–8, and that could be sufficient. Additionally, some software may include optimizations specifically for ARMv9 CPUs, which would naturally lead to better performance.
If I already had an M3 Mac, upgrading to M4 wouldn’t make much sense. However, since I don’t own a Mac right now, choosing the latest architecture is the better option for me.

Dude.

No not really. It’s maybe 10-15 percent performance per core difference. In no universe are you going to be using half the cores in your VMs.

Flip side is you get 2x the cores and more max memory.

Not saying don’t buy m4 max if it meets your use case (for most people it’s probably the best choice) but m3 ultra vs m4 max is a no brainer if you need lots of GPU or lots of ram.

If you need lots of throughput for long running tasks go ultra.

If you don’t and your workload is more about interactive responsiveness and don’t need more than 128gb of ram or uptime GPU performance go max.

It’s only one generation and Apple software doesn’t just not run on older arch. The libraries will fall back to the best the cpu can do. Armv9 isn’t that big a step.

 

[MacPoulet]

As MaxTech mentioned, ARMv9 cores are more efficient, which means higher single-core and multi-core performance in benchmarks. How does this impact my workflow? Well, for one, I might be able to allocate just 3–4 CPU cores to my VMs instead of 6–8, and that could be sufficient. Additionally, some software may include optimizations specifically for ARMv9 CPUs, which would naturally lead to better performance.
If I already had an M3 Mac, upgrading to M4 wouldn’t make much sense. However, since I don’t own a Mac right now, choosing the latest architecture is the better option for me.

I'm not sure if ARMv9 will pay those kind of dividends, but again, not my area.

Maybe someone who knows better, but I thought Apple's previous implementation of ARMv8 also included a bunch of ARMv9 specs already. Can someone (much) smarter than me confirm and/or clarify?

 

[picpicmac]

I'm not sure if ARMv9 will pay those kind of dividends

See the ARM definitions:

A-Profile Architectures

The Arm Application-profile (A-profile) architecture targets high-performance markets.

www.arm.com

AFAIK Apple has not stated publicly how they exploit the version of extensions they have used, or plan to use.

In the short term I think it probably makes no difference for the end user. Perhaps some future version of MacOS will be ARMv9 specific, but that is clearly not the case now.

 

[bxs]

Subject t: Apple's internal SSD vs. external TB5 SSD

Both the new Studios offer TB5 on separate buses. The Ultra offer two more TB5 than for the M4 Max.

New TB5 SSD devices are becoming available, and OWC currently offers them for pre-order with cycle-2 deliveries already completed. They are now continuing with cycle-3 preorders that are expected to deliver within the next few weeks, plus they just dropped the cost of the 4TB Envoy Ultra by $50.

These OWC TB5 SSDs are exceptionally fast, capable of delivering approximately 6,000 MB/s reads and ~5,000 MB/s writes within their embedded ~60 GB write cache. I have the 4TB OWC TB5 Envoy Ultra and have already tested it to verify these specifications.

These OWC TB5 data rates are comparable to Apple’s internal SSDs, and it is important to note that even Apple’s SSDs have an embedded write cache of some size. For instance, my current 512 GB SSD has a ~40 GB write cache.

My workload necessitates extremely fast external SSD devices, as it generates approximately 100 GB checkpoint files two or three times per hour.

However, the current OWC Envoy Ultra’s write cache of ~60 GB is insufficient for my needs. Therefore, my plan is to combine two Envoy Ultras in RAID-0 configuration, which will double the write cache size and likely provide close to 10,000 MB/s reads or approximately 7,000 MB/s writes. These data rates are likely to match the performance of Apple’s internal SSDs.

The requirement for TB5 alone compels me to prioritize the M3 Ultra Studio.

With above in mind my order for the stock M3 Ultra Studio has been placed with a delivery of around Mar 19, 2025. What sweetened the deal for me is that I have a $390 Apple gift card to offset the Ultra's price.

 

[throAU]

I'm not sure if ARMv9 will pay those kind of dividends, but again, not my area.

Maybe someone who knows better, but I thought Apple's previous implementation of ARMv8 also included a bunch of ARMv9 specs already. Can someone (much) smarter than me confirm and/or clarify?

Back to back testing has shown that m3 vs m4 is a wash efficiency wise. M4 burns more power when it runs faster clock. M4 generation runs hotter and louder under sustained load.

And the 10-15 percent vs m3 isn’t an estimate. Its benchmark results.

Anyone thinking they’ll be able to use half the cores in a vm is just not living in reality.

 

[seastar]

Then you should really know what you need, and if you don't it probably doesn't make much difference what you get.

Does being a computer engineer mean having a crystal ball to know everything with absolute certainty? I guess estimations and real-world testing are useless then.

 

[novagamer]

SME is not supported on the M3 Ultra, full stop, which is a big deal for people who work on AI or ML, and I specifically mean developers not people who simply run models. You don't want to have to rely on CoreML, and AMX makes you do that. AMX is obviated entirely by SME. Learning to leverage an old, dead architecture is a waste of time. SME is a standard and goes beyond the Apple ecosystem, although Apple was very likely involved with its creation since it entirely replaces their old one.

INT-8 performance is also tremendously higher in M4 vs M3. There are benchmarks which bear this out: https://blog.roboflow.com/putting-the-new-m4-macs-to-the-test/ Having 2 "last gen" neural engines will overcome this deficit if they can be used in parallel, but I'd sure like to have the new ones if I'm paying the kind of money required for a high-end Ultra config.

Finally, Xcode introduced a feature recently which requires an M4 or better: https://developer.apple.com/documentation/xcode/analyzing-cpu-usage-with-processor-trace?changes=la

If you need a machine now, need a ton of memory, and are OK with a Mac Pro or possibly even an M5/M6 laptop obliterating your $10k workstation within the next 6-18 months, the Ultra makes sense. For most people who don't need to run local models right this very second and aren't using Logic projects with sample libraries in the 100s of Gigabytes it's a pretty bad buy imo.

...

Edit: I'm also curious about the "VM workloads" people have. Outside of docker containers for development work I've personally found Apple Silicon a step back in this area since you can't run Linux (Asahi is a pet project more than an actual distribution and won't even work on modern silicon) or x86 Windows without a bunch of workarounds or inconsistent behavior that is generally a pain in the ass.

For people that are still using a lot of VMs often, what are you doing with them specifically that requires CPU intensive effort? Genuinely wondering this, maybe I'm missing a major use case or haven't used some tool that is out there and I missed.

 

[seastar]

Edit: I'm also curious about the "VM workloads" people have. Outside of docker containers for development work I've personally found Apple Silicon a step back in this area since you can't run Linux (Asahi is a pet project more than an actual distribution and won't even work on modern silicon) or x86 Windows without a bunch of workarounds or inconsistent behavior that is generally a pain in the ass.

You’re saying Apple Silicon is a step back for VM workloads, but that really depends on the use case. If you need native x86 Windows, sure, it’s not ideal—but neither is ARM-based Windows on x86 hardware. On the other hand, ARM-based Linux and Windows VMs run extremely well, and solutions like Parallels and CrossOver allow x86 applications to run with solid performance. Just because it’s different from x86 virtualization doesn’t mean it’s worse.

I’ve personally been using Parallels Desktop for years, running both Windows 11 for ARM and Kubuntu Linux. I do Qt-based software development on both platforms, and I’ve been able to test and refine my cross-platform applications with ease. I can also compile for x86 and x64 Windows targets directly within the ARM version of Windows 11 when needed, without any issues. Even Steam’s x64 version runs smoothly on Parallels Desktop—at least for the games I play, it performs more than well enough.

Also, regarding Asahi Linux—while you called it a 'pet project,' it's actually an actively developed and well-supported distribution specifically optimized for Apple Silicon. It has made significant progress, and many users successfully run it on modern M-series chips.

For people that are still using a lot of VMs often, what are you doing with them specifically that requires CPU intensive effort? Genuinely wondering this, maybe I'm missing a major use case or haven't used some tool that is out there and I missed.

As I mentioned in my previous response, I’ve been using Parallels Desktop for years, running both Windows 11 for ARM and Kubuntu Linux. I do Qt-based software development on both platforms, and being able to compile, debug, and optimize software in multiple environments without switching machines is a huge time-saver.

Additionally, I frequently compile code for x86 and x64 Windows targets directly within the ARM version of Windows 11, and it works flawlessly. Even Steam’s x64 version runs smoothly on Parallels Desktop—for the games I play, it performs more than well enough. This level of flexibility in a virtualized environment is incredibly useful for development and compatibility testing.

If you're interested in setting up something similar, I’d be happy to help. But honestly, you could also just ask AI—it's more than capable of guiding you through the process.

 

[novagamer]

Also, regarding Asahi Linux—while you called it a 'pet project,' it's actually an actively developed and well-supported distribution specifically optimized for Apple Silicon. It has made significant progress, and many users successfully run it on modern M-series chips.

M3 and M4 are still unsupported is what I meant by modern silicon. They really need to be adopted by a larger group who can fund a full development team. It is impressive what they've accomplished through reverse engineering and my personal pov is that Apple should throw some resources at this but I doubt they will unfortunately.

As I mentioned in my previous response, I’ve been using Parallels Desktop for years, running both Windows 11 for ARM and Kubuntu Linux. I do Qt-based software development on both platforms, and being able to compile, debug, and optimize software in multiple environments without switching machines is a huge time-saver.

Additionally, I frequently compile code for x86 and x64 Windows targets directly within the ARM version of Windows 11, and it works flawlessly. Even Steam’s x64 version runs smoothly on Parallels Desktop—for the games I play, it performs more than well enough. This level of flexibility in a virtualized environment is incredibly useful for development and compatibility testing.

If you're interested in setting up something similar, I’d be happy to help. But honestly, you could also just ask AI—it's more than capable of guiding you through the process.

Thanks, nice that Qt development works well for you and I haven't used ARM Windows so didn't know about targeting x86 builds within that environment. Good to know that's a practical reality in real-world use, I will have to look into ARM versions some more.

 

[seastar]

M3 and M4 are still unsupported is what I meant by modern silicon. They really need to be adopted by a larger group who can fund a full development team. It is impressive what they've accomplished through reverse engineering and my personal pov is that Apple should throw some resources at this but I doubt they will unfortunately.


Thanks, nice that Qt development works well for you and I haven't used ARM Windows so didn't know about targeting x86 builds within that environment. Good to know that's a practical reality in real-world use, I will have to look into ARM versions some more.

Yes, Qt development on ARM works very well, and cross-compiling for x64 or x86 is quite practical. You can also install MSYS2 on Windows 11 for ARM64 and set up Qt packages within it for cross-platform development. Additionally, Visual Studio on Windows 11 for ARM64 can compile for x86 and x64 targets as well. If you ever test it out, I’d be curious to hear your thoughts.

 

[seastar]

M3 and M4 are still unsupported is what I meant by modern silicon. They really need to be adopted by a larger group who can fund a full development team. It is impressive what they've accomplished through reverse engineering and my personal pov is that Apple should throw some resources at this but I doubt they will unfortunately.


Thanks, nice that Qt development works well for you and I haven't used ARM Windows so didn't know about targeting x86 builds within that environment. Good to know that's a practical reality in real-world use, I will have to look into ARM versions some more.

By the way, if you're also interested in Linux, I highly recommend Kubuntu, which is KDE-based. Since there isn't an ARM64 version of Kubuntu, you can install it by first setting up the ARM64 version of Ubuntu Server on Parallels Desktop, and then installing the Kubuntu package using the following command:

sudo apt install kubuntu-desktop

After installation, I have a few quick workarounds for some minor issues. Once those are applied, Kubuntu runs excellently on Parallels Desktop on Apple Silicon Macs.

If you're interested in these workarounds, feel free to ask, and I’ll be happy to share them here.

 

[throAU]

SME is not supported on the M3 Ultra, full stop, which is a big deal for people who work on AI or ML, and I specifically mean developers not people who simply run models. You don't want to have to rely on CoreML, and AMX makes you do that. AMX is obviated entirely by SME. Learning to leverage an old, dead architecture is a waste of time. SME is a standard and goes beyond the Apple ecosystem, although Apple was very likely involved with its creation since it entirely replaces their old one.

Agreed with your post and above pretty much 100%.

However the game changer for ai with the m3 ultra though is memory. 512gb of unified memory means that it’s one of the only consumer machines under 50k or more that can actually run the larger models.

Would m4 ultra be quicker due to new instructions? Sure. But it doesn’t exist.

 

[seastar]

Agreed with your post and above pretty much 100%.

However the game changer for ai with the m3 ultra though is memory. 512gb of unified memory means that it’s one of the only consumer machines under 50k or more that can actually run the larger models.

Would m4 ultra be quicker due to new instructions? Sure. But it doesn’t exist.

But my goal isn’t to run AI models, and I don’t have the budget for a 512GB RAM system. I started this discussion simply to find a reasonable system for my use cases within a $4000 price range.
The discussion was about the 16-core M4 Max with 128GB RAM vs the 28-core M3 Ultra with 96GB RAM. Based on real-world tests, the M3 Ultra outperforms the M4 Max by less than 25% in terms of multi-core CPU performance, and this is enough for me to choose M4 Max over M3 Ultra for my use cases.
Of course anybody can say that after the applications are optimized for the M3 Ultra, this gap could increase, but this case will probably be clarified in a few days, and I don't think that this gap will increase dramatically. Otherwise, some computer scientists would warn about this issue, but I haven't seen anything like that. If I see such a thing, I can cancel my M4 Max order or return it before my return period expires to switch to an M3 Ultra Mac Studio.

 

[throAU]

Wasn’t responding to you. Don’t really care for further discussion with yourself.

 

[asiga]

Given that some people were debating between the Max and the Ultra, I gave a second thought to it, even considering the idea of cancelling the M4 Max order and choosing a Ultra instead, like @seastar said. But however, I ended up with the same conclusion: At least for my use case, there's currently no other possible choice but the M4 Max 128GB RAM.

The rationale is as follows:

• The only reason for me considering the Ultra would be being able to order more memory. But this would be related to big LLMs only (because otherwise, for 3D graphics, VMs, and development, 128GB RAM is fine). And then, if you consider deepseek-r1 for example, you need 1TB of RAM for running the complete model locally (I mean the complete 670B model, not the distilled/quantized smaller versions that everybody is running in their NVIDIAs). If you think what the future developments can bring, 256GB is not the amount of RAM that you want for being able to run LLMs. Maybe 512GB would help, but, still, it won't let you run the complete models of the most advanced LLMs. So, this means you are paying over 10000€ for something which is already outdated in terms of the cutting edge LLMs that are available today.
• I guess that we will see increases in supported RAM in next generation chips, but the question is when will 1TB RAM become supported and reasonably affordable in Macs, and, when will M processors have performance enough for running a 670B model at a fast speed (one thing is supporting 1TB RAM, another thing is being able to compute it really fast).
• Another question is: when will stable diffusion run on Macs as fast as on NVIDIA cards? Because I don't find it reasonable to pay more than 5000€ if I'm not getting at least the stable diffusion performance of a 5090.
• Last but not least, from the current benchmarks, the M3 Ultra is not remarkably superior to the M4 Max. Yes, the GPU is 40% faster, but for me it doesn't justify the price increment. Also, I sometimes run single-threaded long CPU tasks, so the M4 would be faster for such tasks (and cheaper).

Putting all of this in perspective, I see no better option than the M4 Max 128GB model for my uses.

 

[ondioline]

Another question is: when will stable diffusion run on Macs as fast as on NVIDIA cards?

The gulf between MPS and CUDA for diffusion is so huge I'd say "never"

 

[xraydoc]

Given that some people were debating between the Max and the Ultra, I gave a second thought to it, even considering the idea of cancelling the M4 Max order and choosing a Ultra instead, like @seastar said. But however, I ended up with the same conclusion: At least for my use case, there's currently no other possible choice but the M4 Max 128GB RAM.

The rationale is as follows:

• The only reason for me considering the Ultra would be being able to order more memory. But this would be related to big LLMs only (because otherwise, for 3D graphics, VMs, and development, 128GB RAM is fine). And then, if you consider deepseek-r1 for example, you need 1TB of RAM for running the complete model locally (I mean the complete 670B model, not the distilled/quantized smaller versions that everybody is running in their NVIDIAs). If you think what the future developments can bring, 256GB is not the amount of RAM that you want for being able to run LLMs. Maybe 512GB would help, but, still, it won't let you run the complete models of the most advanced LLMs. So, this means you are paying over 10000€ for something which is already outdated in terms of the cutting edge LLMs that are available today.
• I guess that we will see increases in supported RAM in next generation chips, but the question is when will 1TB RAM become supported and reasonably affordable in Macs, and, when will M processors have performance enough for running a 670B model at a fast speed (one thing is supporting 1TB RAM, another thing is being able to compute it really fast).
• Another question is: when will stable diffusion run on Macs as fast as on NVIDIA cards? Because I don't find it reasonable to pay more than 5000€ if I'm not getting at least the stable diffusion performance of a 5090.
• Last but not least, from the current benchmarks, the M3 Ultra is not remarkably superior to the M4 Max. Yes, the GPU is 40% faster, but for me it doesn't justify the price increment. Also, I sometimes run single-threaded long CPU tasks, so the M4 would be faster for such tasks (and cheaper).

Putting all of this in perspective, I see no better option than the M4 Max 128GB model for my uses.

I'm going the same route — M4 Max 16/40 with 64GB of RAM. The M3 Ultra would be a lot of fun, I'm sure, but my workflow doesn't even really require the M4 Max, so the M3 Ultra would be wasted. The faster single-thread performance of the M4 Max will be better for me.

Will be placing the order in the next day or two as soon as I get the formal "OK" email from work.

Edit: received the go-ahead, so it's been ordered!

 

[Ekim Venes]

The following video provides a good idea of performance and efficiency of M3 Ultra vs. M4 Max (though not a Studio, but the laptop):

From photo/video content perspective I'd say many creators will have a reasonable (and financial!) preference for 64GB/128GB RAM of 16/40 M4 Max Studio. I personally am still on the fence whether I should "upgrade" from my 64GB M1 Ultra to 128GB M4 Max. From the benchmarks popping up all over slowly but surely, M4 Max is more than a capable contender for M1 Ultra (and even M2 Ultra)

 

[seastar]

The following video provides a good idea of performance and efficiency of M3 Ultra vs. M4 Max (though not a Studio, but the laptop):

From photo/video content perspective I'd say many creators will have a reasonable (and financial!) preference for 64GB/128GB RAM of 16/40 M4 Max Studio. I personally am still on the fence whether I should "upgrade" from my 64GB M1 Ultra to 128GB M4 Max. From the benchmarks popping up all over slowly but surely, M4 Max is more than a capable contender for M1 Ultra (and even M2 Ultra)

This video compares a 32-core M3 Ultra with 256GB RAM against a 16-core M4 Max with 128GB RAM. Given the significant difference in core count, memory, and price, do you think this makes for a fair comparison?

 

[seastar]

The following video provides a good idea of performance and efficiency of M3 Ultra vs. M4 Max (though not a Studio, but the laptop):

From photo/video content perspective I'd say many creators will have a reasonable (and financial!) preference for 64GB/128GB RAM of 16/40 M4 Max Studio. I personally am still on the fence whether I should "upgrade" from my 64GB M1 Ultra to 128GB M4 Max. From the benchmarks popping up all over slowly but surely, M4 Max is more than a capable contender for M1 Ultra (and even M2 Ultra)

This is an unfair comparison between these two systems. The M3 Ultra setup in the video costs nearly twice as much as the M4 Max, making the performance difference less meaningful in a price-to-performance context.

By the way, if you're considering purchasing the M3 Ultra below, I strongly recommend waiting for the Mac Pro. It might offer even more advantages.

 

[Ekim Venes]

This video compares a 32-core M3 Ultra with 256GB RAM against a 16-core M4 Max with 128GB RAM. Given the significant difference in core count, memory, and price, do you think this makes for a fair comparison?

Exactly

 

[seastar]

Exactly

Another unfair comparison video:

 

[Ekim Venes]

This is an unfair comparison between these two systems. The M3 Ultra setup in the video costs nearly twice as much as the M4 Max, making the performance difference less meaningful in a price-to-performance context.

By the way, if you're considering purchasing the M3 Ultra below, I strongly recommend waiting for the Mac Pro. It might offer even more advantages.

View attachment 2491485View attachment 2491483

I'm considering M4 Max 128GB 16/40. Similar price range to what I paid for the M1 Ultra 64GB 20/48 back then, heck even "a little bit" cheaper. Not generally talking for other people and their needs, but M4 Max would be more than likely good enough for those into photo/video process not wanting to spend significantly more than 5000EUR (at least here in Europe that is).

Tyler Stalman's video gives a fair hint of what to expect from latest Studio while not to overpay. Do you think his take on the M4 Max/M Ultra makes for a fair comparison?

 

[Ekim Venes]

Another unfair comparison video:

Ah yeah, many more to come, surely

 

[seastar]

I'm considering M4 Max 128GB 16/40. Similar price range to what I paid for the M1 Ultra 64GB 20/48 back then, heck even "a little bit" cheaper. Not generally talking for other people and their needs, but M4 Max would be more than likely good enough for those into photo/video process not wanting to spend significantly more than 5000EUR (at least here in Europe that is).

Tyler Stalman's video gives a fair hint of what to expect from latest Studio while not to overpay. Do you think his take on the M4 Max/M Ultra makes for a fair comparison?

Under these conditions, the video does appear fair. I apologize for my earlier stance. However, I still feel that the significant price difference deserves more emphasis. Even if the performance is similar, paying nearly twice as much for the M3 Ultra simply doesn't make sense for most users.

For my specific use case, the M4 Max 128GB RAM remains the best choice. It provides an excellent balance of performance, efficiency, and value, making it the most reasonable option within my budget.