Apple announces M3 Ultra—and says not every generation will see an “Ultra” chip

It could explain why we’re getting an M3 Ultra this deep into the M4 rollout.

See full article...

 

[mainmac]

Wow, it's a $1500 price bump to add 4 performance cores to the Ultra. I'll consider the base Ultra, but there's no way I'll pay $1500 for another 4 cores.

That 4 performance cores allows you to bump up the RAM from 256gb to 512gb (for an additional $4,000). For some server uses it's an overall $5,500 additional cost to get 512gb RAM.

 

[radarskiy]

Knowing nothing about the details, but making a completely uneducated guess: You have to have two perfect M3 Max chips and a perfect interposer and you've gotta join them up perfectly. I'm thinking this is an issue with yields. If doing it were easy, we wouldn't have seen delays in the M3Ultra rollout and the M4Ultra would be available on day one. (Do the chip designs have additional P and E cores to improve yields? )

It's not the yield per se, but the two die need to have matched performance curves. They're probably also from a lower leakage bin.

Die level cherry picking has probably taken this long to come up with a quantity worth selling.

 

[Thunderracker]

Omitting FCP is fine, but you'll still want a display, keyboard, trackpad and AppleCare+, which brings the total to... $17064 USD. And that's with just a single Studio Display, not even a Pro Display XDR!

edit: With 8 Pro Display XDRs you can get the total up to $74592, but that's probably a little excessive.

It was a joke. The people who need that machine have jobs that make it worth it. So it doesn’t bother me. Although, 490+ gb of vram on a bus that is nearly a KB is amazing.

I don’t think such a person is buying a Pro Display XDR everytime either.

 

[kenTrotts]

High end video, VFX, and modeling.

Genuine question - what kind of high end video? Like finishing, color, or ad work? Most post companies I know that do scripted TV like The Bear use basic stations and 1080p proxy workflows with Lucid.

 

[adespoton]

One observation that stuck with me was: while they doubled the memory size they're using in their chips over the M2 Ultra, the memory bandwidth is still the same as it was in the M1 Ultra.

Blazing fast, but it's interesting that that component hasn't really changed since the M1.

 

[dropadrop]

Honest question: what are the people who need the Ultras doing with them?

Not the Ultra, but this question made me remember my old workplace over 15 years ago.

Half of the employees were developers working on a Java stack that took around 20 minutes to restart if you wanted to test your changes. This pushed you to do a lot of changes at a time, or you would spend all day waiting.

Unrelated, somebody really wanted a SSD in their new workstation and it was accepted as a ”test” (due to the extremely high cost). Within days of the workstations arrival SSD’s were ordered for every developers computer that allowed swapping the drive. It dropped the time to restart the services by over 90%, and while the drive was expensive as a computer component, it was not expensive when gaining hours of work time per week.

The only clear advantage the Ultra has is the ability to have more ram. If you need it then either it will make a huge difference or make something new possible. But for some work just doing it 40% faster can make the computer seem cheap.

 

[kennetd]

@Andew Cunningham & Lee Hutchinson, I'd love if you wrote a few profiles of the people in roles that make full use of the Ultra platform or other high-end hardware. Pair with @EricBerger for high performance computing in weather. Talk to a film & television post production house. And of course, work-loads for AI modeling. I know less about game engines and physics, but I know a lot of readers have interesting work and would be willing to talk.

Honest question: what are the people who need the Ultras doing with them?

 

[InfiniteJester]

The best news about new Apple chips is bargains for older Macs as the great hermit crab new shell shuffle paradigm eventually puts a computer that has more power than I need at a price I’m willing to pay.

Yours,

that typical user you all talk about

 

[zogus]

Apple didn't make a decision, rather, something technical left them no choice but they will never admit to that.

The only decision they made was what spin they were going to use to explain it.

-kp

Since Apple is not going to explain this, do tell us what exactly is the technical problem here. I mean, you sound way too confident not to have some inside information. C'mon, don't be coy.

 

[tipoo]

Unusual statement to come out of Apple, who usually opt just not to say anything. Smells ike the Mac Pro will jump right to M5 Ultra? Creating a somewhat artificial distinction with the M3 Ultra Studio, since they keep cancelling and delaying a theoretical quad tile Extreme chip

 

[jaberg]

since they keep cancelling and delaying a theoretical quad tile Extreme chip

Can a theoretical chip be canceled or delayed?

 

[ijdf]

Though it's undeniably odd they spent the cash to spin up a new M3 Max variant with the interposer compatibility, you have to assume this is because many people were redirected to work on the M4/M5 variants destined for the in-house AI servers, which will presumably also be used for pro desktops in future.

From memory though, wasn't the M3 big core a bit of a miss? The IPC didn't match the increased decode width, and that was more than fixed in the wider-still M4 big core.

 

[Faceless Man]

Can a theoretical chip be canceled or delayed?

Well, they've repeatedly cancelled and delayed the Apple Television set.

 

[repoman27]

It was a joke. The people who need that machine have jobs that make it worth it. So it doesn’t bother me. Although, 490+ gb of vram on a bus that is nearly a KB is amazing.

I don’t think such a person is buying a Pro Display XDR everytime either.

I got the joke, and was also kidding—for the most part.

Although plenty of people will buy the Mac Studio and have no need of Apple's accessories, I felt it worth noting that the Studio, like the mini, comes with nothing else but a power cable in the box. For a machine that is ostensibly Apple's replacement for the 27-inch iMac, it presents a massively different value proposition. Once again, many customers, especially video editors, are completely fine with that. For the home user who just wants a giant Retina display, like my dad, there's currently a glaring hole in Apple's lineup where that 27-inch iMac used to sit.

 

[Fristie Blade]

Can't speak for everyone, but for me I have a home-grown astronomy app. The telescope shoots at 24MP, in mono, with 7 (automated) filter wheels, for R,G,B,L,S2 H-A, O3 images. These are all combined to provide a multi-spectral image, but to provide temporal averaging, I'll typically shoot a few hundred frames over a session, with the mount tracking what I'm shooting.

I'll frequently end up with 50,000 MP of image-data to reconcile, which is a mathematically-intensive process, even when I do it on the GPU. You need to register all the frames with each other before combining them, to get the best image; you need to watch out for Elon [redacted] Musk's satellites all over the place, as well as the occasional plane going overhead, etc. I use an FFT auto-correlation on the GPU to self-register the images but there are other methods.

Once all that is done, there's a massive "combine them into one image" to do... It ends up looking gorgeous though. Overall processing can take hours with everything maxed out.

I used to have an M1 Ultra, bought it as soon as it came out. I upgraded to an M4-max MBP when that came out, and now (even if it's "only" the M3) I'm looking at 2x the CPU/GPU/memory of the M4-max if I got the M3-ultra... I still love the portability of the MBP so I'm a little torn right now, but I am considering it...

I would love to see the output of that. One of those gorgeous images as I can only imagine they are indeed.

 

[repoman27]

One observation that stuck with me was: while they doubled the memory size they're using in their chips over the M2 Ultra, the memory bandwidth is still the same as it was in the M1 Ultra.

Blazing fast, but it's interesting that that component hasn't really changed since the M1.

For the Ultras released thus far, this is true.

Looking at the entire M-series though, the M1 (no adjective) used LPDDR4X-4266. The M1 Pro/Max/Ultra through M3 and M3 Pro/Max/Ultra all come equipped with LPDDR5-6400. The M4 is paired with LPDDR5X-7500, and M4 Pro/Max uses LPDDR5X-8533.

 

[jaberg]

Well, they've repeatedly cancelled and delayed the Apple Television set.

Gene? Is that you?

 

[macosandlinux]

The shared memory is high enough bandwidth you can run large language models locally at decent speed.

The first consumer PC with 512GB of it, I think, as of right now.

 

[macosandlinux]

Not cheap, yes, but $1,500 also includes 20 more GPU cores.

Looking what a middle of the road 5070 costs, not a bad deal.

 

[invisible21]

So is Apple just dripping stuff over these few days? Hoping the Mac Pro is getting a similar refresh since I'm about to need to buy one for a project at work and am really dreading buying an M2 model when it seems inevitable.

 

[Ted August]

It's not literally as simple as gluing them together obviously - they use die-to-die interconnects which requires its own design considerations. But those considerations would have already been made when the Max chip was designed and shipped. Otherwise, yes, M1 and M2 Ultra were very literally two M1/M2 Max dies joined at the interconnect interface.

Given that M3 Max had no interposer, I'm wondering if Apple is moving away from UltraFusion altogether and M3 Ultra is actually just its own unique die. Nowhere does Apple seem to mention M3 Ultra using UltraFusion like the previous generations.

EDIT: Well, that's wrong. The M3 Ultra press release does say it uses UltraFusion with two M3 Max dies. So hell if I know what's going on.

The M3 manufacturing is quite mature at this point. They may have never intended to create M3 Max silicon with the interconnects on them - and interconnects add to the cost. They are likely doing a separate production of the M3 Ultras with the interconnects added on that specific run of CPU's. There's nothing preventing Apple from changing their silicon manufacturing strategies on their M series chips based on yields and experience. I expect this is how Apple is going to operate for every generation - we're never going to nail down exactly every cycle of CPU introductions and what the features will be. Every generation is going to be slightly different based on whether engineering, manufacturing, and pricing expectations are met.

 

[perrosdelaguerra]

Honest question: what are the people who need the Ultras doing with them?

Working on LLMs, based on what I hear from our ML teams. They want to do as much as possible locally.

 

[tigas]

What I don't understand is, since the "Ultra" chips are just two Max chips fused together, shouldn't it be (relatively) trivial for them to make an M4 Ultra given the M4 Max already exists?

yes
They'll reheat the M4 Ultra when they launch the M5 Max.
Some jobs are better for a white van (slower cores but more of them and lots of bandwidth and memory), other could use a sportscar (faster cores, less of them)

 

[The Wonderful Wizard]

This is entirely from memory, but doesn't the M3 lack the AV1 encode functionality that M4 has? This seems most likely the reason for segmentation, here, to try and force studios into full-fat Pros with coprocessors/GPUs instead of Studios.

Would a productions studio be using hardware compression anyway? At least for h264 and h265, their hardware compression was worse than FFMPEG’s, unless you needed soft-realtime compression or were particularly concerned about power.

 

[plectrum]

Not the Ultra, but this question made me remember my old workplace over 15 years ago.

Half of the employees were developers working on a Java stack that took around 20 minutes to restart if you wanted to test your changes. This pushed you to do a lot of changes at a time, or you would spend all day waiting.

Unrelated, somebody really wanted a SSD in their new workstation and it was accepted as a ”test” (due to the extremely high cost). Within days of the workstations arrival SSD’s were ordered for every developers computer that allowed swapping the drive. It dropped the time to restart the services by over 90%, and while the drive was expensive as a computer component, it was not expensive when gaining hours of work time per week.

The only clear advantage the Ultra has is the ability to have more ram. If you need it then either it will make a huge difference or make something new possible. But for some work just doing it 40% faster can make the computer seem cheap.

I do FPGA development which has 1-8 hour build times. Cutting those down can really improve the number of test cycles you can do a day. FPGAs are notoriously opaque, so sometimes when debugging a problem I'd use the time during the first build to try 4 different modifications to try to smoke out the bug, set them all building, and then test them once each one finished building. That kind of debugging is really quite mind bending...

I'd be really interested trying these Macs for FPGA dev, but since the major vendors have been bought out by Intel and AMD then Aarch64 builds of the ain't happening.

 

[esqua]

Honest question: what are the people who need the Ultras doing with them?

GPU 3D rendering. Math is quite simple in that case: a M3 ultra will roughly render twice as fast as a M3 max. An 1 hour rendering on a M3 max will render in 30 minutes on an Ultra. At the end of the day you'll have save quite some times that can justify the price.

On my current set up, an M3 ultra performance would be somewhere between a 5070ti and a 5080 (with a bit of an edge on Vram size with unified memory). This is where not getting a M4 ultra is a bit disappointing. On a Laptop Apple silicon is a no brainer for me. Weight/autonomy/performance ratio is above all. On a desktop power consuption and size is less of an issue and a nvdia based workstation will be substantially cheaper.

 

[Ouroborus]

Honest question: what are the people who need the Ultras doing with them?

Late game Civ VI

 

[gnasher729]

The best news about new Apple chips is bargains for older Macs as the great hermit crab new shell shuffle paradigm eventually puts a computer that has more power than I need at a price I’m willing to pay.

Yours,

that typical user you all talk about

Typically refurbished Apple devices are 15% below the price of brand new as long as brand new is on sale. If a product stops being sold with a better replacement for the same price, they typically subtract 15% because there is a better product and 15% for refurbished.

My niece found a store apparently run by idiots who discounted the new product and sold the old one at full price (or tried) for a week instead of the other way round. So she got the latest product for 15% less.

 

[gnasher729]

The M3 manufacturing is quite mature at this point. They may have never intended to create M3 Max silicon with the interconnects on them - and interconnects add to the cost. They are likely doing a separate production of the M3 Ultras with the interconnects added on that specific run of CPU's. There's nothing preventing Apple from changing their silicon manufacturing strategies on their M series chips based on yields and experience. I expect this is how Apple is going to operate for every generation - we're never going to nail down exactly every cycle of CPU introductions and what the features will be. Every generation is going to be slightly different based on whether engineering, manufacturing, and pricing expectations are met.

And if they build too many with interconnect they can still sell them as M3-Max.

 

[gnasher729]

Since Apple is not going to explain this, do tell us what exactly is the technical problem here. I mean, you sound way too confident not to have some inside information. C'mon, don't be coy.

Apple doesn’t sell that many high end chips. And if you look at how much faster M4 Max vs M1 Max is, many people needed M1 Ultra but half of those will be happy with M4 Max.

So it is quite reasonable that every second of even third generation has an ultra version to reduce development cost. Doesn’t mean there are any problems.

 

[fixate]

Apple have the strangest release schedule for their M chips.

If it’s for marketing reasons, it’s been absolutely terrible. If it’s technical reasons, I guess that’s fair enough.

 

[stilgars]

Without taking contention etc. into account:

M4 Max : 16 M4 cores => 1.25x 16 M3 cores => 20 M3 cores
M3 Ultra : 32 M3 cores / 20 M3 cores => 1.6x the speed of the M4 Max

This doesn't take into account the GPU performance increase - as I said above, a huge amount of the stuff I use these beasts for now is using Metal and GPU compute. 80 GPU cores, all with access to 512GB of RAM is ... breathtaking. I chew up RAM, CPU and GPU like there's no tomorrow with all the image-registration stuff.

I've also been thinking about how I can leverage the "Neural Engine", which is effectively a lot of matrix multipliers. I do a lot of matrix multiplication... The Ultra has 2x the NE core count as well.

In terms of numbers you may be right.
But you are omitting one thing: the 32 cores from the M3 Ultra are actually 2x16 cores that have to communicate through a bus. If i remember correctly, there was a definitive loss of efficiency for the previous Ultras, versus a pure monolithic Mx Max architecture.
Let's wait for the benchmarks.

 

[Tagbert]

Fwiw it’s nice of Apple to admit they are using M3 but it’s super confusing to the consumer that M3 Ultra is better than M4 Max

the average consumer is not the target market for these expensive machines. The M3 Ultra will be better than the M4 Max for those who need a lot of GPUs or fast RAM. Video and audio processing and LLMs are some uses that come to mind. Those tasks are less dependent on the raw speed of the CPUs and more on GPU and RAM.

 

[geeksman]

I have got M4 Pro Mac mini with following, Still under 14day return policy Should I return and get The base Model M4 Max Mac Studio or just keep the M4 Pro Mac Mini ? (Value, work, stability Purpose) ?

I Always Keep Mac for 3-4Year. I just do simple Photoshop work here an there very light video editing may be once in or twice a month, but use many application at once in 3 display, Some app runs 24/7 even some chrome tab run 24/7 (Transport Gps Tracking in chrome tab) also many tab open 24/7 for other purpose. Heavy use of Canva. Also always Keep open CRM / help desk software and other business software for few brands. Transfer Large amount of file from and to Synology NAS. Looking after 7-10 Vps and Dedicated Server, Looking after 16-20 PHP SAAS Platforms (All Required Heavy Browser Usage), terminal (like tabby) Also want to try self hosted ai in the Mac like ollama and deepseek. Some Other Networking Softwares as well. Running some docker instance as well.

Currently I have :

M4 Pro Mac mini

14‑core CPU, 20‑core GPU,

64GB unified memory

1TB SSD storage

10 Gigabit Ethernet

 

[phuul]

I have got M4 Pro Mac mini with following, Still under 14day return policy Should I return and get The base Model M4 Max Mac Studio or just keep the M4 Pro Mac Mini ? (Value, work, stability Purpose) ?

I Always Keep Mac for 3-4Year. I just do simple Photoshop work here an there very light video editing may be once in or twice a month, but use many application at once in 3 display, Some app runs 24/7 even some chrome tab run 24/7 (Transport Gps Tracking in chrome tab) also many tab open 24/7 for other purpose. Heavy use of Canva. Also always Keep open CRM / help desk software and other business software for few brands. Transfer Large amount of file from and to Synology NAS. Looking after 7-10 Vps and Dedicated Server, Looking after 16-20 PHP SAAS Platforms (All Required Heavy Browser Usage), terminal (like tabby) Also want to try self hosted ai in the Mac like ollama and deepseek. Some Other Networking Softwares as well. Running some docker instance as well.

Currently I have :

M4 Pro Mac mini

14‑core CPU, 20‑core GPU,

64GB unified memory

1TB SSD storage

10 Gigabit Ethernet

Well it all depends. If you don't mind lowering your memory to 32GB then the base Mac Studio with 1TB storage would be $300 cheaper. The upside is you would have 12 additional GPU cores (32 total). To get 64GB of memory you would have to spend $400 more than your M4 Pro Mac Mini but that gets you 2 more CPU cores (16 total) and 20 more GPU cores (40 total).

Based on your usage description I would say that you would probably be better off with more memory rather than more GPU cores so stick with your M4 Pro Mac Mini. On the other hand if you do spend the $400 more you get a better port selection (2 USB-A, HDMI, SDXC, an additional TB 5 port) and you double your GPU cores which may help with the AI dabbling. It's not really a slam dunk one way or the other.

 

[aramando]

It depends on what you do, but for developers there's often a solid business case for faster machines. The biggest cost for many business is salaries. If buying a new machine can make your devs x% more productive, then you can earn back the cost of the machine in a relatively short length of time.

Even if you have a long refresh cycle, even making somebody $1 an hour more productive adds up over the working year.

Perfectly stated. It reminds me of the charity I used to be a web & software developer at who wouldn't spend £200 or so to buy me a 2nd monitor. Quite a breathtaking false economy on their part, thinking back. That was early in my career, I'd be more forthright and demanding now!

 

[nigratruo]

Not unusual in this price range for desktop workstations - but yeah, it's cool here as it means it can benefit from that insane 819.2GB/s bandwidth.

It is so funny how many buy into the bandwidth marketing bullshit argument. It is not the bandwidth that does the work, it is the cores and if you don't have enough cores and/or they are not clocked high enough, then all the bandwidth is useless.

It is not hard to calculate how much memory bandwidth your system needs to not create a bottle neck there, every PC is designed like that, with fast enough RAM so the cores don't get slowed down.

You do realize that if you make a drainage pipe thicker, the water that drains through it is not going to grow because of that, once there is no bottle neck anymore with performance, then thicker pipes do absolutely nothing and can never be used fully.

 

[macosandlinux]

It is so funny how many buy into the bandwidth marketing bullshit argument. It is not the bandwidth that does the work, it is the cores and if you don't have enough cores and/or they are not clocked high enough, then all the bandwidth is useless.

It is not hard to calculate how much memory bandwidth your system needs to not create a bottle neck there, every PC is designed like that, with fast enough RAM so the cores don't get slowed down.

You do realize that if you make a drainage pipe thicker, the water that drains through it is not going to grow because of that, once there is no bottle neck anymore with performance, then thicker pipes do absolutely nothing and can never be used fully.

You buy a Mac Studio with 32‑core CPU to make use of these 32 cores. Otherwise you bought the wrong Mac.

 

[Got Nate?]

It is so funny how many buy into the bandwidth marketing bullshit argument. It is not the bandwidth that does the work, it is the cores and if you don't have enough cores and/or they are not clocked high enough, then all the bandwidth is useless.

It is not hard to calculate how much memory bandwidth your system needs to not create a bottle neck there, every PC is designed like that, with fast enough RAM so the cores don't get slowed down.

You do realize that if you make a drainage pipe thicker, the water that drains through it is not going to grow because of that, once there is no bottle neck anymore with performance, then thicker pipes do absolutely nothing and can never be used fully.

Memory bandwidth has been a major bottleneck in computing since the dawn of time. Thats why modern chips have 3+ layers of increasingly faster caches. This video illustrates just how much a modern CPU sits around waiting for memory to return results:


View: https://www.youtube.com/watch?v=PpaQrzoDW2I

 

[adespoton]

Well, they've repeatedly cancelled and delayed the Apple Television set.

The Macintosh TV came out in 1993? Were you expecting a refresh?