M4+ Chip Generation - Speculation Megathread [MERGED]

[DrWojtek]

Judging from what Nvidia has done with its Grace Blackwell Superchips, x4 is perhaps more likely to be two M5 Ultras on a substrate (CoWoS-L). Assuming Apple sticks with the Max as the building block, hardly a safe assumption now.

One possible explanation for the M3 Ultra delay might be that the entire production, until recently, has gone into Private Cloud Compute hardware.

I’m just throwing guesses out, but an advantage with my speculation is:
1 chip design for all tiers

Great yields: the worst ones becomes lowest tier, for iPads and iMac, and the rest can go to whichever tier they need more SoCs for. Lowest possible waste.

…which in turn provides improved margins.

If they combine this with the on-package tech, they can exclude media engine and AI from the main SoC and put one of these on every complete package.

 

[komuh]

I’m just throwing guesses out, but an advantage with my speculation is:
1 chip design for all tiers

Great yields: the worst ones becomes lowest tier, for iPads and iMac, and the rest can go to whichever tier they need more SoCs for. Lowest possible waste.

…which in turn provides improved margins.

If they combine this with the on-package tech, they can exclude media engine and AI from the main SoC and put one of these on every complete package.

Are you sure it would have greater margins over low->top design (assuming you sell 100x low tier chips compared to high tier chips?)

 

[DrWojtek]

Going by your supposition the GPU core count in an Ultra drops from 80-core to 60-core...?

I know, I havent it all figured out yet 😄

 

[DrWojtek]

Are you sure it would have greater margins over low->top design (assuming you sell 100x low tier chips compared to high tier chips?)

Nope, simple guesswork

 

[Boil]

...the possibility of the Mac Pro getting an early M5.

M5 Ultra & M5 Extreme for the Mac Pro, should drive sales for those needing the highest performance from macOS hardware...?

Judging from what Nvidia has done with its Grace Blackwell Superchips, x4 is perhaps more likely to be two M5 Ultras on a substrate (CoWoS-L). Assuming Apple sticks with the Max as the building block, hardly a safe assumption now.

Maybe a M5 Ultra paired with a M5 Ultra-sized GPU cluster, for more GPU horsepower...?!?

One possible explanation for the M3 Ultra delay might be that the entire production, until recently, has gone into Private Cloud Compute hardware.

The most likely reason...!

 

[leman]

If M3 Ultra is 1.5 cpu performance of M2 Ultra, that’s about 30000 GB multi core score. M4 Max is 27000. Not that great if true.

A puzzling release. Huge LLMs seem to be the main beneficiaries.

Or about 2900 in Cinebench2024 (which is quite a bit better than M4 Max), if you like that benchmark. GB quickly runs into diminishing returns with high core counts.

 

[MRMSFC]

My guess is either Apple is making something bespoke for the Mac Pro or going to dump it entirely.

 

[deconstruct60]

up to 512 Ram and 16TB ssd

chuckle. The 16TB SSD costs more than a Mac Studio Ultra. ( baseline MS-U $3,999 .... just he 16TB upgrade $4,600. which is more than double the 8TB update $2,200 ).
[ that 16TB version on a Mac Pro highly likely will not sell as well. As most folks will opt for a second (or more) alternative drives for bulk storage. ]


Ditto for the 512GB ($4K) RAM. [ relative to a data center AI card with full fledged HBM that is affordable ]

 

[OptimusGrime]

chuckle. The 16TB SSD costs more than a Mac Studio Ultra. ( baseline MS-U $3,999 .... just he 16TB upgrade $4,600. which is more than double the 8TB update $2,200 ).
[ that 16TB version on a Mac Pro highly likely will not sell as well. As most folks will opt for a second (or more) alternative drives for bulk storage. ]


Ditto for the 512GB ($4K) RAM. [ relative to a data center AI card with full fledged HBM that is affordable ]

You can get a gpu with 512GB of ram for less? TIL.

 

[deconstruct60]

Are you suggesting that the M3 Max for the Ultra is a separate tape out, if ever so slightly different at one side of that chip?

Probably. Similar to how the M1 Pro and M1 Max has high amount of shared overlap. M1 Max 'add on' is vastly bigger, but same general technique where just attach to the internal network and expand.

May not be just one side if the TBv5 isn't lying dormant on the M3 Max die. But yeah UltraFusion and any additions needed for Mac Pro PCI-e backhaul to its PCI-e switch for slot bandwidth distribution. That stuff is entirely dead silicon for the MBP deployment. Just wasted silicon. If sell 1-2M MBP Max's that adds up.

If so, I wonder why they didn’t just use M4 Max, but we of course know that the M3 is based on a slightly more advanced node, maybe they needed that for UltraFusion.

If the M3 Max+ dies didn't start getting made until the plain M3 Max dies stopped .... then there are no augmented M4 Max dies to used. And probably wouldn't be until retired the plain M4 max die ... which isn't going to happen for most of this year.

 

[deconstruct60]

It is entirely possible that the Ultra uses a new iteration of the M3 Max chip which includes UltraFusion. Given that Apple did not intent to use the first batch of M3 Max chips for the Ultra, it only makes sense to save the money and chop off the redundant controller part in production.

It is more so cost shifting than cost savings. The UltraFusion stuff isn't being amortized over the whole MBP 14/15 max unit volume. UltraFusion now has to only get the return on investment on where it is actually used.

Pretty good chance that is why Apple is saying Ultra won't necessarily show up on every M-series generation.

 

[krell100]

Well there has been an ultra in every generation now that M3 Ultra is out. We’re just missing the M4 ultra which may be in a pro or the 2026/7 studio.

 

[deconstruct60]

Well there has been an ultra in every generation now that M3 Ultra is out.

Decent chance the M2 Ultra was a kluge due to the pandemic. The M3 Ultra would be late so they weaved in a M2 Ultra to lower the risk profile. Also probably a bit skittish about killing off the large screen iMac 27" for the Studio. If bad user reaction to that what they would need in Ultra space would shift substantively. So the key issue for M1/M2 Ultra was to subsidize them with volume from the MBP 14/16" Max sales. M1 Ultra was to trial ballon with imac->Studio shift. M2 is Mac Pro Ultra only shift. At this point those two 'experiments' with user adaption are done.


The M3 Ultra appears to be more decoupled from the plain/laptop Max . That is going to bring a change. (and is substantively different from what Apple did in the first generation. ).

Decent chance if M3 has been 'on time' for late 2022 / early 2023 that Apple would have gone odds ( M1 , M3 , M5) from the start.

We’re just missing the M4 ultra which may be in a pro or the 2026/7 studio.

The only product that is 100% grounded in the Ultra is the Mac Pro. For over the last decade the Mac Pro has been no where near yearly updates. 2013->2019 2019->2023 if Apple got to a 2 year iteration cycle that would be a huge 'upgrade' in cycle times. Yearly? It likely isn't happening.

 

[tenthousandthings]

It is more so cost shifting than cost savings. The UltraFusion stuff isn't being amortized over the whole MBP 14/15 max unit volume. UltraFusion now has to only get the return on investment on where it is actually used.

Pretty good chance that is why Apple is saying Ultra won't necessarily show up on every M-series generation.

I like your suggestion (espoused in a different thread) that this is about creating a landing spot for Max/Ultra silicon after its primary deployment(s). In this case, we have the initial deployment in the M3 Max in the MBP, then it gets adapted to 2x Max = Ultra, with its primary, initial deployment into Private Cloud Compute. Then, finally, it gets released into the Ultra Mac Studio. That’s at least a three year production run for M3 Max/Ultra on N3B, from 1H 2023 (at the latest) to 1H 2026 (at the earliest).

Future cases, like M5, would probably be structured differently (depending on what happens with Private Cloud Compute), but the end result would be the same, with the Ultra Mac Studio being the last stop for the Max/Ultra, after more glamorous deployments elsewhere.

 

[deconstruct60]

I like your suggestion (espoused in a different thread) that this is about creating a landing spot for Max/Ultra silicon after its primary deployment(s). In this case, we have the initial deployment in the M3 Max in the MBP, then it gets adapted to 2x Max = Ultra, with its primary, initial deployment into Private Cloud Compute. Then, finally, it gets released into the Ultra Mac Studio. That’s at least a three year production run for M3 Max/Ultra on N3B, from 1H 2023 (at the latest) to 1H 2026 (at the earliest).

If Apple was wlling to deal with a standard 6 or 8 pin molex power connector for a PCI-e card they could probably toss a Mn Max+ die/SoC onto a PCI-e card also. ;-) Mac-on-a-card. [ I think it is bit too close to the 75W standard bus power limit for something without a AuX power supplement. ]

Future cases, like M5, would probably be structured differently (depending on what happens with Private Cloud Compute), but the end result would be the same, with the Ultra Mac Studio being the last stop for the Max/Ultra, after more glamorous deployments elsewhere.

Not sure Private Cloud Compute is going to be glamorous. I won't be surprised if Apple never shows details to anybody. Unless Apple starts charging for Apple Intelligence, it is just a massive cost sink for Apple operations. Cost sinks aren't 'sexy'. The more amorphous and intangible it is, the more likely stockholders won't ask where a huge chunk of their money is going.

 

[deconstruct60]

My guess is either Apple is making something bespoke for the Mac Pro or going to dump it entirely.

The Mac Pro is in somewhat the same position the iPad's are. It has deeper needs for more capable software than hardware.

If virtualization could assign direct mapped I/O to guest OS VM instances than other operating system could chose to support other cards or not. If had virtual ethernet between host OS and PCI-e card could have other OS instances running on cards in the slots ( and do fast 'NAS' storage shares so can swap data quickly). More 'cluster in a box' software. More Xgrid like framework to lash system instances together.

A lot less flakey higher end PCI-e SSD storage. ( the Apple attitude of ignoring other SSD vendors is a problem ).

Expand the market for non visual GPU cards. The list that Sonnet has for PCI-e cards ( https://www.sonnetstore.com/collections/thunderbolt-expansion-systems/products/echo-2dv-desktop see supported PCI-e cards list).
How much of this stuff is still on legacy I/O Kit drivers and how much on newer DriverKit model? Apple has a stagnation problem there that is moving slow motion.

TBv5 will./should open door to handle something better than 25GbE cards.

Apple seems to be majority catering to folks with huge costs stuff into fairly old card designs. ( A/V space ). That is the major issue. There is little 'growth' there.


If there was more value in having the slots then they would sell incrementally more Mac Pros. Sell enough and it won't be hard to keep it around and updating at a 2-3 year rate.

 

[novagamer]

According to Gurman the Mac Pro is supposed to get the M4 Ultra. Someone shared a private message with him from last April where he specifically told them Apple was testing the M4 Max / M3 Ultra for this year’s studio, and he said the M4 Ultra would be the differentiator for the Mac Pro.

Obviously take it with a grain of salt, but I found that interesting. I wonder if Hidra is just a code name for the Ultra, but it does make Apple’s statement about “not every generation getting an Ultra” seem odd if that’s true, unless they’re moving straight to M5 which would be a (welcome) surprise.

As someone else mentioned the M3 not having SME makes this whole thing really weird from a consumer standpoint, my assumption si Apple did use M3 Ultra for their private cloud and now has enough capacity to offer it to consumers but who knows. I never expected M3 Ultra at all given what we allegedly knew about the die shots etc.


Edit: for clarity, someone earlier today shared a conversation with Gurman from 11 months ago where he specifically mentioned the above including the M3 Ultra in the Studio and M4 Ultra in a Mac pro. Again, not saying it’s correct – just want to make explicit what I’m referring to. This is new information as far as I can tell, the way he phrased his Bloomberg reporting never named the CPUs specifically until yesterday once the press briefings happened.

Not sharing it because I necessarily believe it, but it does feel relevant to speculation about what’s next, especially with the context that he was correct (in the text) about the M3 Ultra and M4 Max being in the Studio that far ahead of time, which may give some veracity to his follow-up text that said the M4 Ultra was only being used for the Mac Pro.

At this point I don’t even have a prediction about the Mac Pro, the M3 Ultra was so out of left field for me that I can’t even speculate what they’re going to do with the even more expensive option. Giving it an M3 Ultra seems weird since it didn’t happen today, keeping it as M2 Ultra makes no sense if an update is more than 6 months away, having an entirely new CPU or next generation option seems unlikely but Apple being so specific about not every generation getting an Ultra to multiple media outlets including Gruber make me think the M4 Ultra isn’t happening either. No clue.

My comedy guess is M3 Extreme. You heard it here first.

 

[deconstruct60]

According to Gurman the Mac Pro is supposed to get the M4 Ultra. Someone shared a private message with him from last April where he specifically told them Apple was testing the M4 Max / M3 Ultra for this year’s studio, and he said the M4 Ultra would be the differentiator for the Mac Pro.

".... and he said the M4 Ultra ..." above


"... Gurman suggests that Apple might opt to go with an M3 Ultra in the ‌Mac Studio‌ in order to distinguish it from a ‌Mac Pro‌ with a future M4 Ultra chip. ..."

Mac Studio With M4 Max and M3 Ultra Chips Could Launch This Week

An updated version of the Mac Studio could launch as soon as this week, reports Bloomberg's Mark Gurman. The new machine is expected to be...

www.macrumors.com

You have to be careful about where Gurman says "someone at Apple said ..." versus " I think Apple should" or "I think Apple might .." . Those are two different things. He sometimes slip-slides between those two modes in the same article.

Apple 'might' use M4 Ultra to differentiate Mac Pro is different Apple 'is suppose to'.

It appears Apple has modified the M3 die to get to a M3 Max+ in order to use in an Ultra is expensive. (especially where it somewhat looks like the M4 Max in the Mac Studio isn't using the same variant die. ). The volume of Studio Ultra model is small. To do it again on a M4 Max would be even more expensive. Makes even less sense given the Mac Pro user base at this point is even smaller.

Somewhat detaching the Max" desktop die from the laptop die is going to cost money. Burning a deeper hole isn't going to help the Mac Pro over the long term. And Mac Pro is missing more I/O uplift and further PCI-e slot value-add enable ment than a pissing contest on M-series version numbers bragging rights.

Apple has stated that there will not necessarily be Ultras on every iteration. Gurman has lots of article content about Mn Extreme that hasn't seen light of day either.

As someone else mentioned the M3 not having SME makes this whole thing really weird from a consumer standpoint, my assumption si Apple did use M3 Ultra for their private cloud and now has enough capacity to offer it to consumers but who knows. I never expected M3 Ultra at all given what we allegedly knew about the die shots etc.

Private Cloud doesn't make any money ( Apple isn't charging for it). AS long as it is in the state that it generation no revenue I'm wonder how it is suppose to be functioning as some kind of Ultra cost overhead saviour that affords churning for pay SoCs faster. Not sure how just throw clodue nodes in the trash can every 12 months to save the next gen Mac from cost overuns.

More likely both are going to iterate slower ( just as the Mac Pro has for over a decade. ).

 

[Macintosh IIcx]

Probably. Similar to how the M1 Pro and M1 Max has high amount of shared overlap. M1 Max 'add on' is vastly bigger, but same general technique where just attach to the internal network and expand.

May not be just one side if the TBv5 isn't lying dormant on the M3 Max die. But yeah UltraFusion and any additions needed for Mac Pro PCI-e backhaul to its PCI-e switch for slot bandwidth distribution. That stuff is entirely dead silicon for the MBP deployment. Just wasted silicon. If sell 1-2M MBP Max's that adds up.



If the M3 Max+ dies didn't start getting made until the plain M3 Max dies stopped .... then there are no augmented M4 Max dies to used. And probably wouldn't be until retired the plain M4 max die ... which isn't going to happen for most of this year.

Apple might also just have s surplus of node N3B production that they need to burn on something. They might be bound by a contract to buy a certain amount even here in 2025. N3E is clearly more suited to the mobile SKUs whereas N3B is more advanced so probably a good fit for the M3 Max UF sku.

 

[Pressure]

The power supply was also reworked and got an increase from 370 Watt to 480 Watt.

That should tell us something about the M3 Ultra at least, which still has to be able to supply 6 x 15 Watt for all the TB5 ports and 2 x 7 Watt for the USB-A ports.

 

[tenthousandthings]

[…] Private Cloud doesn't make any money ( Apple isn't charging for it). AS long as it is in the state that it generation no revenue I'm wonder how it is suppose to be functioning as some kind of Ultra cost overhead saviour that affords churning for pay SoCs faster. Not sure how just throw clodue nodes in the trash can every 12 months to save the next gen Mac from cost overuns.

More likely both are going to iterate slower ( just as the Mac Pro has for over a decade. ).

Maybe look at it this way. Apple is about products, right? That’s the mantra. The product here is Apple Intelligence. PCC is arguably its most important component, and it requires Apple silicon. PCC doesn’t generate revenue, but the same could be said for all Apple silicon. It’s the products that generate revenue. From macOS to iOS, from iPhone to Mac Pro.

It’s not like the M2 Ultras will be shut down because the M3 Ultras are online. I’d be surprised if they are pulling out all the M2s and replacing them with M3s.

A two-year cadence for the Ultra would be a vast improvement. For Apple itself and many of its Ultra customers, a four-year replacement cycle would be the target for the marketing comparisons.

M1 Max/Ultra came out exactly three years ago, in October 2021/March 2022. Let’s assume M2 Max/Ultra was originally planned for October 2022/March 2023 (both launches were pushed back three months by Zero Covid). I think the step taken yesterday with M3 Ultra is about establishing a two-year cadence for the Ultra.

M5 Ultra will feature in the “Hidra” Mac Pro introduction at WWDC 2025. The M5 Ultra Mac Studio will launch in March 2026. The M7 Ultra will appear at WWDC 2027, with the M7 Ultra Mac Studio coming in March 2028.

 

[komuh]

Maybe look at it this way. Apple is about products, right? That’s the mantra. The product here is Apple Intelligence. PCC is arguably its most important component, and it requires Apple silicon. PCC doesn’t generate revenue, but the same could be said for all Apple silicon. It’s the products that generate revenue. From macOS to iOS, from iPhone to Mac Pro.

It’s not like the M2 Ultras will be shut down because the M3 Ultras are online. I’d be surprised if they are pulling out all the M2s and replacing them with M3s.

A two-year cadence for the Ultra would be a vast improvement. For Apple itself and many of its Ultra customers, a four-year replacement cycle would be the target for the marketing comparisons.

M1 Max/Ultra came out exactly three years ago, in October 2020/March 2021. Let’s assume M2 Max/Ultra was originally planned for October 2022/March 2023 (both launches were pushed back three months by Zero Covid). So we can already see the two-year cadence in place. I think the half-step taken yesterday with M3 Ultra is about restarting the planned two-year cadence when the M5 Ultra Mac Studio launches in March 2026.

I think M5 Ultra will also feature in the “Hidra” Mac Pro introduction at WWDC 2025. Then, two years later, the M7 Ultra will appear at WWDC 2027, with the M7 Ultra Mac Studio coming in March 2028.

I'm not sure if there is even big enough audience for "Hidra" chip and this isn't just some sort of server chip for their own cloud which won't hit consumer in next few years.

 

[Antony Newman]

The power supply was also reworked and got an increase from 370 Watt to 480 Watt.

That should tell us something about the M3 Ultra at least, which still has to be able to supply 6 x 15 Watt for all the TB5 ports and 2 x 7 Watt for the USB-A ports.

… and perhaps Apple have improved the thermal dissipation capabilities using their M4 Mac Mini design.

<snip> I think M5 Ultra will also feature in the “Hidra” Mac Pro introduction at WWDC 2025. Then, two years later, the M7 Ultra will appear at WWDC 2027, with the M7 Ultra Mac Studio coming in March 2028.

If Apple skip from M5 Ultra -> M7 Ultra .. would that mean the Mac Studio would get a hand-me-down M5 Ultra in 2028 when Laptops get the M8? Deliberately keeping the Mac Studio on 2-3 year older SoCs sounds like a recipe for bitterness.

If the Studio already has the thermal dissipation for the Ultra chip that goes into the Mac Pro (more likely now with TSMCs new packaging solution) - perhaps we will end up with the Studio and Pro getting the same Ultra chips from the M5 onwards?

If the M4 Ultra chip is in fact delayed - (perhaps for a faster Ultrafusion for lower latency / higher bandwidth Max GPUs performance to scale properly) - why couldn’t Apple Also offer this in the Mac Studio at the same time (with a suitable price hike)?

With all the delays, confusion, and 2-3 year desktop cadence around the Studio - perhaps they should simply release the Max in the Studio yearly when the Laptops are rolled out? In Oct 2025 … would you rather buy a Mac Studio with an M3 Ultra for $$$, an M4 Ultra for $$$$ or an M5 Max for $$?

 

[MayaUser]

Apple official said that not all M generations will get the Ultra...until now, its false but maybe was a statement for the M4 ultra?!

 

[tenthousandthings]

Maybe look at it this way. Apple is about products, right? That’s the mantra. The product here is Apple Intelligence. PCC is arguably its most important component, and it requires Apple silicon. PCC doesn’t generate revenue, but the same could be said for all Apple silicon. It’s the products that generate revenue. From macOS to iOS, from iPhone to Mac Pro.

It’s not like the M2 Ultras will be shut down because the M3 Ultras are online. I’d be surprised if they are pulling out all the M2s and replacing them with M3s.

A two-year cadence for the Ultra would be a vast improvement. For Apple itself and many of its Ultra customers, a four-year replacement cycle would be the target for the marketing comparisons.

M1 Max/Ultra came out exactly three four years ago, in October 2020/March 2021. Let’s assume M2 Max/Ultra was originally planned for October 2022/March 2023 (both launches were pushed back three months by Zero Covid). So we can already see the two-year cadence for the Ultra in place. I think the half-step taken yesterday with M3 Ultra is about establishing a parallel two-year cadence for the Ultra Mac Studio, a year behind the Ultra itself. The M5 Ultra Mac Studio will launch in March 2026.

M5 Ultra will also feature in the “Hidra” Mac Pro introduction at WWDC 2025. Then, two years later, the M7 Ultra will appear at WWDC 2027, with the M7 Ultra Mac Studio coming in March 2028.

EDIT: Corrected basic counting error (see strike-through above), and adjusted the subsequent argument in that paragraph accordingly…

Thus, if I’m right about the Ultra, the only remaining question is the Max Mac Studio. Will it be on an annual cadence like the Max, or will it also skip generations like the Ultra? I think the message yesterday is annual, but it will lag behind. It will never launch in October, it will always launch the following March. Every other year, it will get last year’s Ultra.

I don’t think whatever happens with the still-unknown Hidra really affects this at all, because I think the Ultra will be a component in that. Even if it is the only component, Ultra can still have a life of its own, alongside its role in Hidra…