I don’t know if the comparison is fair or not, but 1) All his tests are GPU-intensive and 2) Even then, the Ultra doesn’t always win by a substantial margin. However, one part that puzzled me is the LLM part. The video he shows claims that deepseek-r1 670B is 400GB. But everybody says it’s 700GB. I went to the ollama site and it says 400GB, and I don’t understand anything. It’s like the files at ollama are half their expected sizes. I’m a newcomer to AI, but I’m certainly not understanding what’s going on. Any search will tell you that you need 700GB for running deepseek-r1 locally.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)
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.
M4 Max (16-Core CPU, 128GB RAM) vs. M3 Ultra (28-Core CPU, 96GB RAM)
- Thread starter seastar
- Start date
- Sort by reaction score
Yeah I'm leaning in that M4 Max 128GB direction as well. Eventually some more useful videos shall surface to make that final decision. In a couple of weeks I hope to be a bit wiser/informed enough🤞
Can't comment on the LLM as it's not my field of interest. Which definitely puts me in the M4 Max camp for now🤫 M3 Ultra is no doubt a great machine for "great" price, but M4 Max price/performance value is undeniable when compared to its stronger and "older" sibling.
The option you’re leaning towards is not aligned with the budgeting you’re presenting.
To select 128GB for the M4 Max, you must also opt into the higher core count. This puts you at $3,499 vs $3,999 for the M3 Ultra. That is nowhere near “twice as much” and arguably a poor value prop to gain an additional 32GB of RAM.
Based on your use case, have you considered building a dedicated AI PC with an NVIDIA GPU and using an M4 Mac Mini or a base M4 Max Mac Studio for everything else?
If AI workloads are your main priority, a high-end PC with an NVIDIA GPU could offer significantly better AI performance and greater modularity compared to the M3 Ultra. Meanwhile, for general computing, development, and creative work, an M4 Mac Mini or a base M4 Max Mac Studio would still be a great choice at a lower cost.
Additionally, a custom-built AI PC would provide greater expandability and flexibility. You could upgrade the GPU, add more RAM, or swap components as needed which isn't possible with a Mac Studio or a Mac Mini.
I see your point, and I appreciate the clarification. When comparing the M3 Ultra (96GB RAM, 28-core CPU) at $4000 to the M4 Max (128GB RAM, 16-core CPU) at $3700 (with equal SSD capacity), the price difference is actually only $300.
However, for my use case, I prioritize newer architecture, better efficiency, and higher RAM capacity over additional CPU cores. The M3 Ultra has more cores, but they are previous-generation performance and efficiency cores, and since it combines two M3 Max chips using UltraFusion, it may lead to higher power consumption, more heat due to the dual-CPU design, and potential fan noise in real-world usage.
Paying an extra $300 for only a ~25% multi-core CPU performance gain while sacrificing 32GB of RAM doesn’t seem like a good trade-off for me. That’s why I see the M4 Max with 128GB RAM as the better overall value.
Simple explanation. Full DeepSeek R1 requires more than 400GB, but Q4 (quantization so weights are reduced to only 4 bits (on average) from fp8 - 8 bits), roughly halves the amount of memory required for the weights. A Q4 model is NOT the full model, and not as good as the full model, but may be okay depending on purpose. A Q3 model will use around 3 bits (on average), a Q2 2 bits to store weights. If you want to run full DeepSeek R1, it'll take two 512GB Studios.
Last edited:
I'm afraid for me the Windows/PC ship sailed a long time ago. I can absolutely see the appeal for ultimate power and DIY freedom with future upgradeability. However, Studio offers compactness and power efficiency, given its overall performance it burns less in the long run - despite its initially high price - compared to a capable Windows rig. Less heat, less noise, small(ish). Reliable enough. Real question (for me at least) is whether to upgrade now or give it another 1-2 years and wait for M5 Max/Ultra. It's not that I must own the latest Mac Studio, it's more about I'm thinking about to go for it. Nothing wrong with M1 Ultra, just that M4 Max performs so good and temptation is imminent🫣
If Windows/PC isn’t an option for you, then a Mac Studio is definitely the best way to go. The compact design, power efficiency, and reliability are all strong advantages, and I totally get why you'd prefer that over a modular PC.
The real question, as you said, is whether it’s worth upgrading now or waiting for the M5 Max/Ultra. Honestly, the M4 Max is already a huge leap over the M1 Ultra in terms of single-core performance and efficiency. If your workload benefits from that, it might be worth upgrading now rather than waiting another 1-2 years.
With WWDC coming up in June, we might see a new Mac Pro with an M4 Ultra or a base model that is more powerful than the M3 Ultra. It could be a compelling alternative if you prioritize expandability and future-proofing—but it would likely have active cooling with more fans, which might not align with your preference for a quieter system.
Additionally, if you ever consider AI workloads in the future, a dedicated Linux AI computer with an NVIDIA GPU, running alongside macOS, could provide a seamless hybrid experience. You could handle general computing, software development, and creative tasks on macOS, while offloading AI processing to a Linux system via remote inference or containerized workflows. This way, you get the best of both worlds.
If silence and efficiency are key factors, then the M4 Max Mac Studio still seems like the best fit for you right now.
Thanks a lot for the clarification. It puzzles me how people pretend they run the full deepseek locally, when the reality is that they don’t (I’m talking about the youtube video that claims to run the complete deepseek on a 512 Studio).
"Fair" is not something of concern here.
A professional photographer/videographer should be concerned about throughput.
If an M3 Ultra allows them to do more work in a day then the $2k-$4k price bump is not very much, and once amortized over the years is basically a couple of dollars a day.
You are right, I misunderstood the video when I first watched it. I should have watched it completely before commenting. That was my mistake, and I have apologized for it.
Instead of paying that price bump, it might be better to wait for the Mac Pro, which will likely be announced at WWDC in June. It could offer more advantages within the same price range
That viewed like a guy trying to convince himself his purchase was justified. There was 0 comparative info.
agreed. But I think my point is, for $300 more, you get more GPU cores and 2 extra TB5 ports. But you lose 32GB of RAM. But in his video he didn't go above 77GB with everything running. So if GPU and multicore use is more important for you, I think the M3 is the way to go.
To be technically accurate, for $300 it includes 12 more CPU cores, 20 more GPU cores, an additional 409GB/s memory bandwidth, two extra Thunderbolt 5 ports, one additional video decode engine, two additional video encode engines, and two more ProRes encode and decode engines.
In exchange for a lose of 32GB ram and unless you really sure you need that additiaonl 32GB ram most of the time and cant accept swapping.
You can’t simply say "12 more CPU cores" without converting their performance equivalence. Let's make a simple linear estimation:
If the 32-core M3 Ultra has 100p performance (where "p" represents a unit of performance), then the 16-core M4 Max has approximately 75p performance.
Based on this assumption:
- 1 M3 core would contribute 3.125p performance (100p / 32).
- 1 M4 core would contribute 4.6875p performance (75p / 16).
A 32-core M3 Ultra CPU has a performance difference equivalent to 5 more M4 Max CPU cores compared to a 16-core M4 Max CPU.
So, the difference between them is not simply "12 more CPU cores"—it's effectively 5 additional M4 Max CPU cores in terms of performance.
I based my calculation on the 32-core M3 Ultra CPU, but if I had compared it to the 28-core version, the results would have been similar—meaning the difference still wouldn’t be 12 extra cores.
If you’d like, I can redo the calculation after verifying whether the benchmarks were conducted with the 28-core or 32-core M3 Ultra
Of course, this is a highly simplified and purely linear calculation. Real-world performance depends on workload efficiency, architecture optimizations, and other factors. However, this estimation might help illustrate the relative performance scaling.
The memory bandwidth difference is not 409GB/s, but 309GB/s.
- M3 Ultra has 819GB/s memory bandwidth.
- M4 Max has 510GB/s memory bandwidth.
Last edited:
Based on the calculations in my previous reply, by paying an additional $300 and sacrificing 32GB of RAM, you would get a system equivalent to a 21-core M4 instead of a 16-core M4, if you choose the base model of the M3 Ultra Mac Studio.
Additionally, you would gain 20 more GPU cores, but at the cost of using an older-generation system, which may lack support for certain small features in the coming years.
To illustrate the impact of generational differences, I recently installed macOS Sequoia on my 2013 Mac Pro (Black Cylinder). While it still has a reasonably powerful CPU and GPU, its performance is sluggish, and the overall experience feels unstable.
Of course, this is an extreme example, but I wanted to highlight how older architectures can gradually lose efficiency and compatibility over time.
Additionally, you would gain 20 more GPU cores, but at the cost of using an older-generation system, which may lack support for certain small features in the coming years.
To illustrate the impact of generational differences, I recently installed macOS Sequoia on my 2013 Mac Pro (Black Cylinder). While it still has a reasonably powerful CPU and GPU, its performance is sluggish, and the overall experience feels unstable.
Of course, this is an extreme example, but I wanted to highlight how older architectures can gradually lose efficiency and compatibility over time.
Last edited:
The upgrade to Apple's 8TB SSD is some $2,000.
Consider that OWC offers their 4TB Envoy Ultras TB5 units for some $550 each, so 8TB can be obtained for around ~$1,100 configuring the Ultras in RAID-0 or as a JBOD. This fact tends to highlight the M3 Ultra with it having more TB5 ports.
Also, consider that RAID-0 two Envoy Ultras will provide the combined 60GB SLC write cache size of each Ultra to be 120 GB. The read and write data rates for this Ultra RAID-0 will likely out perform the Apple's internal SSD.
External SSDs can be added to or replaced if their life-span degrades badly, whereas the Apple's internal SSD suffering the same degraded life-span fate or constantly exhausting its free space will cause one to purchase a new Mac.
I've tested the OWC's Envoy Ultra on my M4 Pro mini's TB5 ports using the Terminal command dd, and see a good 6,000 MB/s for reads and some ~5,000 MB/s for writes done without over subscribing the SLC that I discovered to be 60 GB.
For me the TB5 feature offered on these new Studios is a game changer for me as my most used CFD application writes out enormous checkpoint files, runs for days and sometimes weeks, is CPU and memory intensive, and will grab as many cores as it can or is allowed to grab. This aspect drives me to selecting the M3 Ultra
Consider that OWC offers their 4TB Envoy Ultras TB5 units for some $550 each, so 8TB can be obtained for around ~$1,100 configuring the Ultras in RAID-0 or as a JBOD. This fact tends to highlight the M3 Ultra with it having more TB5 ports.
Also, consider that RAID-0 two Envoy Ultras will provide the combined 60GB SLC write cache size of each Ultra to be 120 GB. The read and write data rates for this Ultra RAID-0 will likely out perform the Apple's internal SSD.
External SSDs can be added to or replaced if their life-span degrades badly, whereas the Apple's internal SSD suffering the same degraded life-span fate or constantly exhausting its free space will cause one to purchase a new Mac.
I've tested the OWC's Envoy Ultra on my M4 Pro mini's TB5 ports using the Terminal command dd, and see a good 6,000 MB/s for reads and some ~5,000 MB/s for writes done without over subscribing the SLC that I discovered to be 60 GB.
For me the TB5 feature offered on these new Studios is a game changer for me as my most used CFD application writes out enormous checkpoint files, runs for days and sometimes weeks, is CPU and memory intensive, and will grab as many cores as it can or is allowed to grab. This aspect drives me to selecting the M3 Ultra
Last edited:
MicroCenter has the base M3 Ultra for only $3400. Get then while they are in stock!
I’m in the exact same situation as you I’m undecided between M4M or M3U . I would like to speed up the import in Lr of my Canon CR3 files (R5 MII), which is currently very slow. Exporting is quite fast, except for PDF generation. I’m not sure if it’s worth it, and I don’t know if I would get any substantial benefits. I’m curious about what you will choose😌
Last edited by a moderator:
For $3,399.99 at MicroCenter (N.B. AVAILABLE FOR IN-STORE PICKUP ONLY.)
SKU: 831073
Apple Mac Studio MU973LL/A (Early 2025) Desktop Computer; Apple M3 Ultra 28-Core CPU; 96GB Unified Memory; 1TB Solid State Drive; 60-Core GPU
This MicroCenter price of $3,399.99 compares with Apple's $3,999.00 for exact same model specs. That's a $600 discount!
If it had the 2TB internal SSD I'd jump for it ASAP. This $3,399.99 is a good deal IMO.
For more see https://www.microcenter.com/search/...4294967292+4294819353+4294806409&myStore=true
Last edited:
Same here. I have now ordered and canceled 3 times - silly I know. For very similar workflows - video and photo (incl. conversions but not much rendering), but also as my daily system (and yes, I do play native games when I have time) - I am between M4 max 16/40 128GB or Ultra 3 with 32/80 (probably 96GB, but I could do 256GB).
It's not so much a question of $ and ROI but, like most others are asking, if it's worth spending considerable more on the U3 for my use cases. I am coming from an m1 Ultra (the high-spec version). Dunno. I feel the m4 for 2-3 years would serve me better 'now' (I don't keep these more than 3 years) but this forum help me keep changing my mind