M4 Mac Announcements

M1 Ultra took 16 months after the release of the M1. The M2 Ultra took a year. So the M3 Ultra taking 16 months suddenly doesn't seem so crazy ... it's just that the M4 came on so fast and it appearing that the regular set of M-series chips will be updated every year, makes these Pro chips look behind now - especially if the M5 (Ultra) gets released soon.



If so, which seems very plausible, that'll be really interesting ...

My prediction (i don’t make a lot of predictions): We are at the end of stitching identical die together to make Ultra’s. When next we see “ultra” it will be hybrid die.
 
I agree this doesn't make sense. So I conclude Apple did this only because they weren't able to produce an M4 Ultra in time. They must be struggling to merge the two Max chips. To make up for this, and keep the model relevant for AI development, the max RAM on the Ultra is, for the first time, 4x the max RAM on the Max (512 GB vs. 128 GB) (rather than 2x) which should be appealing to those who are working on very large models. Though it still does have (two of) the older (20 TOPS) M3 neural engines, rather than the newer (38 TOPS) engine on the M4, and Apple promoted the latter as being a valuable advancement for AI.

This Ultra also differs from past Ultras in that its max storage is upped to 16 TB, twice the Max's, where before they had the same max storage.

Definitely a win for Gurman, but I have to wonder how Apple feels about him stealing their thunder. Normally you'd think Apple wouldn't like this, but I wonder if Apple authorized these leaks, thinking Gurman's audience would bring more attention to this announcement (?).
 
Last edited:
My prediction (i don’t make a lot of predictions): We are at the end of stitching identical die together to make Ultra’s. When next we see “ultra” it will be hybrid die.
Agreed.
I agree this doesn't make sense. So I conclude Apple did this only because they weren't able to produce an M4 Ultra in time. They must be struggling to merge the two Max chips. To make up for this, and keep the model relevant for AI development, the max RAM on the Ultra is, for the first time, 4x the max RAM on the Max (512 GB vs. 128 GB) (rather than 2x) which should be appealing to those who are working on very large models. Though it still does have (two of) the older (20 TOPS) M3 neural engines, rather than the newer (38 TOPS) engine on the M4, and Apple promoted the latter as being a valuable advancement for AI.

This Ultra also differs from past Ultras in that its max storage is upped to 16 TB, twice the Max's, where before they had the same max storage.

Definitely a win for Gurman, but I have to wonder how Apple feels about him stealing their thunder. Normally you'd think Apple wouldn't like this, but I wonder if Apple authorized these leaks, thinking Gurman's audience would bring more attention to this announcement (?).
I prefaced my own hesitation with saying that when Gurman says something is happening in the next couple of weeks, he's usually right (in this case it was one day). I'll admit I was thinking (okay hoping) his own unsure language meant he was guessing, but then I saw the @exoticspice1 post last night/early this morning and as they pointed out there were missing M3 chips, which felt like the clincher. New M3s it is. Still strange even if, as I posted earlier, the timeline is not actually off the previous M1/2 Ultras ...
 
Yeah, I'd like to understand how that works. The annotated die shots I've seen show the TB controllers are on-die, so does this mean that, when they designed the M3, they must have ensured the controllers could accommodate both TB4 and TB5? Calling @Cmaier to the courtesy phone....
I favor the tweaked die hypothesis - also to explain why the ultra fusion was seemingly missing on the M3 Max but now suddenly there. But of course that could be wrong.
 
I favor the tweaked die hypothesis - also to explain why the ultra fusion was seemingly missing on the M3 Max but now suddenly there. But of course that could be wrong.

yeah, i assume this is a new M3 Max die. It’s comparatively easy to just modify some of the small controllers without touching everything else.

But this is all still weird.
 
And it uses Ultrafusion and tb5. Weird. The article says not every generation will get an ultra chip…even though they have so far.

I believe the generation reference is to the chips, not the appliance...?

My prediction (i don’t make a lot of predictions): We are at the end of stitching identical die together to make Ultra’s. When next we see “ultra” it will be hybrid die.

So maybe a hybrid of a standard Mn Max sort of chip and a GPU-specific sort of chip, stitch those together with UltraFusion for a Hidra chip, and then use an underlying interface to stitch two of these hybrid Hidras into an Extreme chip; just something to give more GPU cores to the overall package...?
 
I'm a little disappointed with the release today to be honest.

I really like the M4 Max, but I'd hoped that they would be doing what everybody else was expecting and hoping for with the M4 Ultra.
Probably what I'm most disappointed with is the pricing structure... it could be a reflection of new tarrifs? Not sure (saying that not to get political but more to say that it might explain the product pricing anticipating a weaker canadian dollar).
Needless to say, this is not a great release IMHO for what I was hoping (and waiting) for with a mythical M4 Ultra.
 
Probably what I'm most disappointed with is the pricing structure... it could be a reflection of new tarrifs? Not sure (saying that not to get political but more to say that it might explain the product pricing anticipating a weaker canadian dollar).
The M4 MacBook Air got a surprise $100 price cut, allowing Apple to put their latest CPU at the $999 entry-level price, so it doesn’t look tariff-related to me (for now anyway).
 
I'm a little disappointed with the release today to be honest.

I really like the M4 Max, but I'd hoped that they would be doing what everybody else was expecting and hoping for with the M4 Ultra.
Probably what I'm most disappointed with is the pricing structure... it could be a reflection of new tarrifs? Not sure (saying that not to get political but more to say that it might explain the product pricing anticipating a weaker canadian dollar).
Needless to say, this is not a great release IMHO for what I was hoping (and waiting) for with a mythical M4 Ultra.

I guess if you have highly parallel workloads, today you can get your work done twice as fast as before (with a given number of boxes). So it’s probably good news for a lot of people in the Ultra market, even if it’s not the news they really were hoping for.

I’m not in that market anymore, so I don’t know nuthin. My M1 Max MBP still is plenty fast enough for me.
 
So the next question: Will these offer higher max clocks than the same-gen MPB's? Guessing not, but we shall see. Maybe the M4 offers some additional headroom in that department...

If they were able to boost the max GPU clock on the M3 Ultra, that would make it a more interesting product.
 
yeah, i assume this is a new M3 Max die. It’s comparatively easy to just modify some of the small controllers without touching everything else.

But this is all still weird.
Apparently they didn't modify the memory controllers, which are much larger. The M3 Ultra has twice the M3 Max's memory bandwidth (suggesting it uses LPDDR5-6400, same as the M3), rather than twice the M4 Max's (which uses LPDDR5X-8533).

Also unchanged from the last (M2) Ultra are the specs of the UltraFusion bridge. Both
and
say:
"Apple’s custom-built UltraFusion packaging technology uses an embedded silicon interposer that connects two ... dies across more than 10,000 signals, providing over 2.5TB/s of low-latency interprocessor bandwidth"


Apple is touting the M3 Ultra's advancements for AI:

"M3 Ultra is built for AI, including ML accelerators in the CPU, Apple’s most powerful GPU, the Neural Engine, and over 800GB/s of memory bandwidth. AI professionals can use Mac Studio with M3 Ultra to run large language models (LLMs) with over 600 billion parameters directly on device".
 
Last edited:
"M3 Ultra is built for AI, including ML accelerators in the CPU, Apple’s most powerful GPU, the Neural Engine, and over 800GB/s of memory bandwidth. AI professionals can use Mac Studio with M3 Ultra to run large language models (LLMs) with over 600 billion parameters directly on device".
Seems like this may be a focus for them. The devil’s in the detail, but that’s a big jump.
1741214261787.png
 
It didn’t make sense to me that Apple would try to pass the M4 air at 1099 for 256/16 with the M2 at $999 through 2025 and half of 2026 with how far windows laptops have come (MacBook airs also still use a 60Hz LED display that, while high quality is doubtless depreciated and cheap by now) and where the market will be in Q4/Q1 as well, and developer wise as Dave noted it was also bad strategy.

This isn’t ideal — marginal upgrade pricing for storage/RAM is still too high but like it’s at least very much in respectable territory for sure especially given the chip and build quality, MagSafe, Microphone, trackpad. With sales we will will see 512/16 for $1000.
 
Seems like this may be a focus for them. The devil’s in the detail, but that’s a big jump.
View attachment 34087
I checked the footnote, and unfortunately it doesn't give the test conditions. It would be nice to confirm this 16.9x difference on a model "with hundreds of billions of parameters" is due purely to a difference in compute performance, and not an "artificial" difference created by choosing a model that can't fit into the M1 Ultra's maximum 128 GB RAM (and thus requires some swapping to disk), but can fit into the higher RAM options on the M3 Ultra.

I'm probably overanalyzing Apple's language, but Apple Kremlinology is fun, so....

To qualify as "hundreds of billions", you'd need 200B minimum. According to Apple, 512 GB is big enough for their representative 600B parameter model. And I assume this means just big enough, i.e., if 700B would fit, they would have said that instead. If it's linear with memory size (and the following table indicates it is), a 200B parameter model (of the same type) would not fit into the M1 Ultra's maximum 128 GB RAM.

1741234308678.png


Finally, this indicates it is possible (though not desireable) to run LLMs using swap:
 
Last edited:
Back
Top