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.
Would have bought an M4 Ultra Studio - decided not to get the M3 Ultra or maxed out Max - anyone else?
- Thread starter M4M
- Start date
- Sort by reaction score
My problem with the M3 Ultra is that the Ultra chips never scaled well to begin with, and the problem is compounded when the Max model is a generation newer than the Ultra model. In many tests the M4 Max is very close in performance to the M3 Ultra.
When comparing M3 and M4 generations - the M4 Pro can come pretty close to the M3 Max when running LLMs with MLX. I need to see more tests and comparisons, but I believe that the M4 Max and M3 Ultra are going to come pretty close in tokens/sec, especially when running MLX.
Without the gains of what a M4 Ultra would provide, the value of the extra RAM available is diminished to some extent. I am looking to run batch jobs with data out of a database, and feeding it through APIs into the LLM server. I just don't see how really large models, that can utilize 512GB RAM, will run with any acceptable speed on the M3 Ultra. If it was an M4 Ultra at least the performance should have been a bit better.
I'm hoping to be proven wrong, but the price difference between a specced up M4 Max and the M3 Ultra does not seem like it will provide the performance advantage to justify the cost. Running a model that can be loaded into the amount of VRAM that 512GB can provide does not seem like it could possibly offer the performance that could justify the increased cost. For my purposes multiple M4 Max machines running models sized to give appropriate performance would make much more sense.
The one benefit of the M3 Ultra, is that in reality it received a price drop - maybe to offset how close in performance the M4 Max is to the M3 Ultra. The M2 Ultra used to come with 64GB RAM, where the M3 ultra now comes with 96GB standard - and in the Apple world that is worth around $400. So in essence the consolation of getting an M3 Ultra instead of an M4 Ultra is the fact that Apple knocked about $400 off of the price.
Useful reviews are just starting to land, and I am hoping to see more before deciding to make an investment. If Nvidia cards were not in such short supply (and ridiculously over priced when they can be found), the decision might have been easier to just go the x86 route for this particular use case - but that is not today's reality.
If benchmarks coming out show that paying substantially more money for an M3 Ultra, provides substantially better performance in tokens/sec over the M4 Max, then I would be happy to pick up an M3 Ultra. I just don't believe that will be the case, though.
When comparing M3 and M4 generations - the M4 Pro can come pretty close to the M3 Max when running LLMs with MLX. I need to see more tests and comparisons, but I believe that the M4 Max and M3 Ultra are going to come pretty close in tokens/sec, especially when running MLX.
Without the gains of what a M4 Ultra would provide, the value of the extra RAM available is diminished to some extent. I am looking to run batch jobs with data out of a database, and feeding it through APIs into the LLM server. I just don't see how really large models, that can utilize 512GB RAM, will run with any acceptable speed on the M3 Ultra. If it was an M4 Ultra at least the performance should have been a bit better.
I'm hoping to be proven wrong, but the price difference between a specced up M4 Max and the M3 Ultra does not seem like it will provide the performance advantage to justify the cost. Running a model that can be loaded into the amount of VRAM that 512GB can provide does not seem like it could possibly offer the performance that could justify the increased cost. For my purposes multiple M4 Max machines running models sized to give appropriate performance would make much more sense.
The one benefit of the M3 Ultra, is that in reality it received a price drop - maybe to offset how close in performance the M4 Max is to the M3 Ultra. The M2 Ultra used to come with 64GB RAM, where the M3 ultra now comes with 96GB standard - and in the Apple world that is worth around $400. So in essence the consolation of getting an M3 Ultra instead of an M4 Ultra is the fact that Apple knocked about $400 off of the price.
Useful reviews are just starting to land, and I am hoping to see more before deciding to make an investment. If Nvidia cards were not in such short supply (and ridiculously over priced when they can be found), the decision might have been easier to just go the x86 route for this particular use case - but that is not today's reality.
If benchmarks coming out show that paying substantially more money for an M3 Ultra, provides substantially better performance in tokens/sec over the M4 Max, then I would be happy to pick up an M3 Ultra. I just don't believe that will be the case, though.
I'll make the point that Apple won't give the 2025 Mac Studio longer support based on which CPU it has. The M3 Ultra model will be supported just as long as the M4 Max variant.
Case in point: the A10X iPad Pros stopped at iPadOS 17, while the A10 iPad 7th gen received iPadOS 18 solely because it was on sale longer.
Case in point: the A10X iPad Pros stopped at iPadOS 17, while the A10 iPad 7th gen received iPadOS 18 solely because it was on sale longer.
This is a pet peeve of mine. Words have meaning…
The point of the video is that the pricing structure...and what you get...for buying the M3 Ultra version do not make much sense - especially if you upgrade to the top tier chip.
And he's right about that.
The cost that it takes to upgrade to the M3 Ultra version from the M4 Max version do not give you the rate of return that you should be getting...especially when the M4 Max version is way better at single-core work and it comes surprisingly close in some of the multi-core benchmarks as well.
But, not enough for a human interaction. If an M3 Ultra gets 6 tokens/s and your hypothetical M4 Ultra gets 8 tokens/s, that 8t/s is still going to feel slow compared to using an LLM over the internet. If you want realtime interactivity that a human won't find boring, then you'll need to buy one of those newly-announced DGX-workstations. They will cost a lot more than an M3 Ultra but you'll be able to chat to your heart's content.
I wanted an M4 Ultra, but wasn't really ever going to buy one. I could afford it if I wanted, but wasting money unnecessarily isn't worth it for just a little FOMO. I don't need an Ultra.
My tricked out M4 Max was delivered today (16/40-core chip, 64GB RAM, 2TB SSD). It's fantastic! Besides, web browsing and email aren't going to happen any faster on an Ultra anyway.
My tricked out M4 Max was delivered today (16/40-core chip, 64GB RAM, 2TB SSD). It's fantastic! Besides, web browsing and email aren't going to happen any faster on an Ultra anyway.
Web browsing and email are basically gonna be the same on a base M4 Mini as a top of the line M4 Max.
Enjoy that Studio! It sounds like a beast. Would love to hear how it works out for you.
I don't understand the resale value thing. It's a tool for a job. Does it matter how much it loses in value if the tool helps make you more productive, or the most productive? There is no other computer Apple offers that will compete, even an M5 Max won't on the upper end of performance, which is what the Mac Studio w/ Ultra chip is all about. Buy what you need and forget the rest of the noise. If you really don't need an Ultra processor, why even consider buying it?
Last edited:
Given Ultra chips are just two Max chips is there any reason they can't make an M4 Ultra?
Well, that was the case. But having watched Luki Mianni's video it seems the M3 Ultra only excels in a few specific tasks. For everything else the performance difference between the M4 Max is negligible.
That said a person needs to evaluate a system based on their specific needs and buy accordingly. Use the tool which makes you the most productive even if it uses "older" technology.
I was being a little facetious the web and email comment, but thanks!
Are they really? We do know that all M1Max had the option to connect to another chip meaning any non defect pair on a wafer could have turned in an Ultra. Same for the M2Max but not for early M3 and current M4Max chips.
At which point one has to wonder if the option to split a pair still exists for M3Ultra or if that designation has to be when the wafers are made.
In many cases the M4 Max Studio is better/faster than the M3 Ultra. If the Ultra was M4 based there would have been none of those negatives as everything on the Ultra would have been as good as the Max or better.
As reviewers are saying: "They have basically set it up for failure by basing it on last year's chip"
We've been told that's what they are. I have no reason to disbelieve that information.
I agree 100%. That's essentially what my post you responded to said. However I did add the statement that one has to evaluate a product based on their needs and, if the "older" technology product is the better choice then buy it despite it being "older" technology.
All this said where does this leave the Mac Pro?
Apple Trade In on a Base Mac Studio M2 Max 32GB/512GB SSD is $910. The cost was $1799 education (or sale) pricing two years ago.
Retained 51% of value after two years
$889 loss in two years, $444.50 loss per year
Base M2 Ultra 64GB/1TB SSD Apple trade in is $1220. The cost was $3599 education (or sale) pricing two years ago.
Retained 34% of value after two years
$2379 loss in two years, $1189.50 loss per year
These are indeed large drops in value for such powerful machines. If the computers were never used but kept as gold bricks, one could consider these decreases as "losses." Even if only serving for bragging rights and never powered up, they would have had some value.
Retained 51% of value after two years
$889 loss in two years, $444.50 loss per year
Base M2 Ultra 64GB/1TB SSD Apple trade in is $1220. The cost was $3599 education (or sale) pricing two years ago.
Retained 34% of value after two years
$2379 loss in two years, $1189.50 loss per year
These are indeed large drops in value for such powerful machines. If the computers were never used but kept as gold bricks, one could consider these decreases as "losses." Even if only serving for bragging rights and never powered up, they would have had some value.
But Apple's own trade-in system has never offered the most competitive prices. You'll always do better selling it on your own. That said, Apple's system is clean and quick.
It's a tricky question.
From M1/M2 Ultra, we know that the Ultra performance isn't exactly 2 Max performance. The ultrafusion thing that connects two Max doesn't scale well. So will Apple continue this in every future Max is a question. Two weeks ago Apple already said not every M family chip will have an Ultra version. This is a strong indication that there won't be an M4 Ultra.
Furthermore, we already know that M3 Ultra is NOT two M3 Max, based on two facts:
1. M3 Max does not support TB5 while M3 Ultra does.
2. M3 Ultra can have 512 GB of memory but none of the M3 Max family has 256 GB of RAM.
So no matter what, connecting two M3 Max will not give you an M3 Ultra, and splitting an M3 Ultra will not give you two M3 Max. Given this fact, it's questionable whether M3 Ultra is splittable. Plus, even if it can be split into two M3 Max+, there are no Macs that can take advantage of this. All the other Macs move to M4. So, an M3 Ultra may contain two identical units and each one looks very similar (but not identical) to an M3 Max, but perhaps the two units are never meant to split, unlike the case for M1/M2 Ultra.
1. This is a common misconception and often repeated. It is true, but does not mean what you think it means. Fact is the TB5 ports are supported by custom-designed controllers, they are NOT part of the M3 chip itself but are a separate, integrated component.
2. None of the released M3 Max's have 256 gb ram. We know Apple has binned chips. We don't know they don't have 256 gb Ram chips, or if they do, what the yield might be such that they potentially reserved them for the M3 Ultra. One could just as easily argue the existence of the 512 gb M3 Ultra supports that Apple DOES have 256 gb M3 Max chips.
Only Apple knows.
some reviewers are doing the click bait thing and saying that...
Apple has pivoted to AI because Apple knows it has already conquered the video processing world, more performance just isn't required by most, but AI? The field is wide open..... Apple also knows most people don't need a Mx Ultra, which isn't the same as some just want it as evidenced by this thread. But the better reviewers that actually are paying attention acknowledge for some use cases (AI) the Ultra is a great success. And for the use cases where the Max is enough, well congrats, you get to save money. Whats the complaint?
Oh, you want improved single core use? Why, just how fast do you need to answer email with? Or web browse? Fact is for 99% of the use cases, the M3 single core speed is not perceptibly slower than the m4.
Agreed, the M3 Ultra isnt for everyone. But the ultra series never was.
Yes, we kind of do. LPDDR5 are made mostly by Samsung, Micron, or SK-Hynix. With the M3 Max using four LPDDR5, if you want 256GB of RAM then each LPDDR has to be 64GB. None of the manufacturers ever announced they could do that, and only recently did Samsung claim they could make 32GB LPDDR5x, though they don't have it on their sales page. The M4 Max will use four of those to get 128GB.
As for TB5, while it is true there is a separate IC for the interface, to handle the power delivery requirement, the handshakes, etc., the M3 SoC still has to have IO controllers that are designed to handle the bandwidth, and the display controllers on the SoC have to deal with multiple 8K displays.
As for interconnect between the two sides of the Ultra SoC: a noticeable fraction of the surface area of a die for a large SoC is for wiring all the pieces together, and notably the memory controllers. One does not merely paste two such die together. The engineers will have to had laid out the wiring ahead of time knowing that an interconnect is going to be implemented.