Agree. I think it refers to M4 Ultra. Reading between the lines, I’d say it contains the following message: it’s fair to expect an annual cadence for the Max Mac Studio, but not for the Ultra Mac Studio.
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+ Chip Generation - Speculation Megathread [MERGED]
- Thread starter GoetzPhil
- Start date
- Sort by reaction score
You can see what I think in my posts directly above, but with regard to your two questions here, I’d say:
[1] Mac Studio will get the hand-me-down M5 Ultra in March 2026, not 2028. March 2027 at the latest (in this alternate scenario, March 2027 would see the M6 Max Mac Studio and the M5 Ultra Mac Studio launch at the same time, much like what happened yesterday, while March 2026 would only see the M5 Max Mac Studio.)
[2] I think, yes, both Studio and Pro will get the Ultra, but the Pro (and the Apple Intelligence PCC servers) will get it first. The only question is how long the lag will be.
On the thermals, the Extreme/“superchip” Mac Pro will require, I would imagine, a larger heat sink. It looks to me like there is room in the current chassis for a 2x Ultra configuration, but anything beyond that, like 4x Ultra, would be found only in the PCC hardware.
Indeed. Here's my (probably fairly obvious, but I'm going to say it anyway) take: if the new M3 Ultra Studio identifiers are 15,x (likely), then we should still expect 17,1 and 17,2 Mac Pros with M5 Ultras (or whatever they end being called) this year. If the M3 Ultra Studios are 17, then the M5 Ultra may not come until next year and the longer timeline of the Studio being updated in 2027 becomes more likely too.
Still possible they announce an M3 Extreme based Mac Pro...
One obvious place to do this might be in the WWDC Keynote, as some sort of "Apple Intelligence is just awesome. How did we fine our largest base models? Using the new Mac Pro, which I'll describe now..."
They are very good at getting the max press bang for the buck by announcing genuinely new (and somewhat unexpected) items at a different event, not piggybacking on an earlier event. (cf the way no-one expected Pro and Max in the M1 generation, then no-one expected the Ultra, announced via a 3rd event).
Uh, duh...
Just like huge LLMs are the main beneficiary of Blackwell.
Both Apple and nV (and everyone else who isn't named Intel) understand what the shape of computing looks like over the next few years, and is adapting to this! If you're not in LLMs, you'll still benefit, with some delay, from the substantially boosted DRAM capacities and bandwidths, and the general pivot to give these devices much higher throughput.
Presumably at some point, just like GPGPU figured out a variety of new ways to exploit the new style of throughput compute provided by GPUs, people will figure out new ways to exploit the new style of compute (eg matrix ops, reduced precision) provided by these new style GPUs and neural engines. Remember, eg, neural radiance fields (NERFs) as the new hotness in creating images (kinda an alternative/augmentation to ray tracing)? They're still around, and presumably people will continue to explore what can be done with them.
You'e assuming that people are buying Apple 16TB SSD models to store their large email folders.This is probably a very dumb assumption.
I expect it's more like people like movie editors who, yes, are storing the entire movie assets on some large company-wide SAN, but who also move specific assets onto local SSD storage while making modifications. These are people who know exactly what labor costs/hour are worth, and the value of the fastest possible storage tier.
Same thing for your naive assumption that a "data center AI card with full fledged HBM that is affordable" is a real thing...
People have run the numbers! M3 Ultra is a lot cheaper than any realistic nV solution, and substantially cheaper than an AMD solution. It's not quite apples to apples of course, for many reasons:
- nV only allows you access to 192GB "without hassle". You have to start structuring your code and using specific additional transport commands to get beyond that
- nV CAN scale a lot larger via nVLink if you need that. Of course Apple can scale a lot larger via TB5 and MLX, but nVLink is higher bandwidth
- the HBM solutions give you higher DRAM bandwidth. That sounds important BUT remember that everyone training large models has ALREADY had to cope with the fact that training is spread over multiple tiers, first of something like a DGX, then communicating between DGX's. So techniques exist (which work well) for splitting training into tiers of local training (making full use of bandwidth) and "synchronized" training, where different entities sync with each other what they learned during the last training epoch (eg, in Apple's case, via TB5).
I am no expert in this, not even close! But people who are experts seem to feel that it's not a real problem. yes, it's a problem if you want to naively run nV code without modifying it (but in that case you buy nV!) If you're willing to make the effort (apparently not much) to move the code to MLX, then there's not much difference. nV trains a certain size granularity with occasional sync via nVLink; Apple trains on a larger granularity because of the 512GB vs 192; and syncs more slowly and less frequently via TB5, but it's expect that this all kinda works out OK. (Of course no-one has ACTUALLY tried this yet; this is all based on experience from earlier MLX on Ultras, M4 minis and so on.)
Someone else pointed out that, for example, you could fine tune (not train from scratch, but fine tune to become an expert in some specific area, if you have the appropriate training data) a leading edge model OVERNIGHT on 4 of these M3 Ultra 512GB machines. (You need 4 to hold the 2TB of DRAM required.)
So the expectation by the people who do this stuff is that, contrary to your claim, these machines ARE actually competitive -- IF you're doing more than just naively running nV code without any attempt to tweak it to match the target. Also depends a LOT on whether your interests are
- production training
- production inference
- R&D (ie want to experiment with different ideas, spending as little time as possible dealing with overhead like "how do I pack the large memory footprint I want to work with into the DRAM/HBM pool available?" Here Apple's unified memory, and a single large 512GB pool, are a real help.
Perhaps the cpu that goes into the Mac Pro this year will not be called M4 Ultra, but something else? To this point, Ultra has meant two fused chips, perhaps the Mac Pro will feature an M4 that is not "Ultra", but a more monolithic design.
Regardless of what the future holds for the Pro I’d have a difficult time dropping $10k+ on a machine that didn’t support SME given the limitations of how you have to utilize AMX and that those skills aren’t fully transferrable, although some are subsumed into SME I think. I guess an updated CoreML might utilize both?
On that note, has there been more information about scheduling SME on a per cluster level yet? The last (~4 months ago) write-up I saw said that there was not which is, to be fair, a big downside of that for the time being. Given the focus on AI/ML I’m surprised I haven’t seen more analysis of this.
On that note, has there been more information about scheduling SME on a per cluster level yet? The last (~4 months ago) write-up I saw said that there was not which is, to be fair, a big downside of that for the time being. Given the focus on AI/ML I’m surprised I haven’t seen more analysis of this.
Last edited:
it will be the M2-Epic just to keep us all on our toes
This is why I'm hoping high VRAM SOC systems like the Halo become the norm (or at least more widespread) on the PC-side as well. As that no longer becomes a bottleneck, what will developers be able to do that they weren't before for standard consumer systems? Professional applications like 3D rendering and AI-inference/tweaking are immediately obvious, but what else is possible that someone simply hasn't thought of yet?
There was actually an unreleased Mac15,14 identifier in the same file that contained all the M4 and M5 Mac identifiers.
I assumed it was a cancelled M3 Mac since I could also see two M2 Mac identifiers that never released.
I think it’s likely that Mac15,14 is the M3 Ultra Mac Studio and not a cancelled product
M3 Ultra Mac Pro or M3 Ultra Mac Studio? Do you have a link to the full list? I agree - just like we saw additional M3 chip designations that seemingly went unused (until now). I just don't want get people's hopes up and say for sure that X is happening.
Last edited:
I meant to say M3 Ultra Mac Studio.
I don’t have a link but I did write down all the Mac identifiers I saw
M2 Macs
Mac14,2: 13” M2 MacBook Air
Mac14,3: M2 Mac Mini
Mac14,5: 14” M2 Max MacBook Pro
Mac14,6: 16” M2 Max MacBook Pro
Mac14,7: 13” M2 MacBook Pro
Mac14,8: M2 Ultra Mac Pro
Mac14,9: 14” M2 Pro MacBook Pro
Mac14,10: 16” M2 Pro MacBook Pro
Mac14,11: Never released
Mac14,12: M2 Pro Mac Mini
Mac14,13: M2 Max Mac Studio
Mac14,14: M2 Ultra Mac Studio
Mac14,15: 15” M2 MacBook Air
Mac14,16: Never released
M3 Macs
Mac15,3: 14” M3 MacBook Pro
Mac15,4: 24” M3 iMac (2 ports)
Mac15,5: 24” M3 iMac (4 ports)
Mac15,6: 14” M3 Pro MacBook Pro
Mac15,7: 16” M3 Pro MacBook Pro
Mac15,8: 14” M3 Max MacBook Pro (16 CPU, 40 GPU)
Mac15,9: 16” M3 Max MacBook Pro (16 CPU, 40 GPU)
Mac15,10: 14” M3 Max MacBook Pro (14 CPU, 30 GPU)
Mac15,11: 16” M3 Max MacBook Pro (14 CPU, 30 GPU)
Mac15,12: 13” M3 MacBook Air
Mac15,13: 15” M3 MacBook Air
Mac15,14: Unreleased (likely M3 Ultra Mac Studio)
M4 Macs
Mac16,1: 14” M4 MacBook Pro
Mac16.2: 24” M4 iMac (2 ports)
Mac16,3: 24” M4 iMac (4 ports)
Mac16,5: 16” M4 Max MacBook Pro
Mac16,6: 14” M4 Max MacBook Pro
Mac16,7: 16” M4 Pro MacBook Pro
Mac16,8: 14” M4 Pro MacBook Pro
Mac16,9: Unreleased (likely M4 Max Mac Studio)
Mac16,10: M4 Mac Mini
Mac16,11: M4 Pro Mac Mini
Mac16,12: Unreleased (likely 13” M4 MacBook Air)
Mac16,13: Unreleased (likely 15” M4 MacBook Air)
M5 Macs
Mac17,1: Unreleased
Mac17,2: Unreleased
Last edited:
Awesome, thanks! Do you know which file leaked the info? Fascinating about the possibility of two unreleased M2s. I can't think of what those might've been - even if the Extreme had once been on the table there is unlikely to have been two models. Hmmm ...
It was one of the Apple Intelligence ones.
I believe it was the com_apple_MobileAsset_UAF_FM_GenerativeModels.xml file.
I’m not sure if that’s the exact file but it was definitely in the generative models folder.
When you opened it, there was a bunch of iPhone, iPad, Mac etc identifiers listed.
They might have removed it recently with a new macOS update since the last time I accessed it was with macOS 15.1 beta
Last edited:
MPX (minus the TB pass-thru stuff) could provide the extra power needed...?
Mn Extreme for the Mac Pro (Tower/Rackmount) and the all-new Mac Pro Cube...! ;^p
The old nodes will continue to run, Apple will add new nodes to their vast data center on the loweest levels of the Mothership... ;^p
Mac17,1 = M5 Extreme Mac Pro Cube
Mac17,2 = M5 Extreme Mac Pro (Tower/Rackmount)
Whereas I am ConfusedUser^2.
It is 100% certain that's it's a new tapeout. They could not have had a working TB5 controller on the original M3. I think it's also known that there was no UF.
M3 is not on a "more advanced node". It's different. It's slightly denser, but slightly slower and hotter. Nothing about it (N3B) would enable UF, in contrast to N3E. After all we got UF on the M1. That really was on a significantly less advanced process.
Definitely not, because:
That's not how it works. Packaging has its own losses. Beyond that, the result is less efficient due to spending energy moving bits around between chiplets (much more expensive than on a single monolithic chip) and die area on all the inter-chip comms. Then there's the cost of packaging the chiplets.
In short, using chiplets to build Pro- and Max-class chips would be insane. Apple would never make such a boneheaded move.
The file is still there but they have removed mentions of unreleased Macs as far as I can tell. This is the list of supported devices now:
I see two paths towards more processing power in the high-end Apple Silicon offerings:
I don't know, I just want to see Apple produce a real fire-breathing (in processing power, not in power usage or heat output) chip product for use in a high-end desktop personal workstation, wrapped up in an all-new Mac Pro Cube chassis...! ;^p
- Monolithic designs that push the reticle limit, one CPU+ chip using UltraFusion to connect to one GPU+ chip, two heads of the Hidra
- Chiplets design using an underlying interconnect, allowing customization of assorted cores to meet the needs of end users
I don't know, I just want to see Apple produce a real fire-breathing (in processing power, not in power usage or heat output) chip product for use in a high-end desktop personal workstation, wrapped up in an all-new Mac Pro Cube chassis...! ;^p
It could also be that the original M3 Max already had TB5 controller but restricted to TB4 back in 2023. I guess we will never know for sure.
M4 Ultra Geekbench CPU score.
Mac15,14 - Geekbench
Benchmark results for a Mac15,14 with an Apple M3 Ultra processor.
browser.geekbench.com
M4 Ultra Geekbench Compute score.
Mac15,14 - Geekbench
Benchmark results for a Mac15,14 with an Apple M3 Ultra processor.
browser.geekbench.com
This reply there reported a multi-core of 39044, which is more inline with the M3 Max binned score of around 19000?
I mean, they’re wrong without proof. Geekbench 6 doesn’t scale like that.
I agree, at least the entry above had to go through Geekbench's system. Also I looked closer, the roughly 1.4x multiplier is more inline with what how the M2 Ultra scored above the M2 Max.