M4+ Chip Generation - Speculation Megathread [MERGED]

[caribbeanblue]

No core increase for the Max. 500GB/s bandwidth. 128GB max.

Slightly disappointed.

Spec upgrades aren’t necessarily symmetrical between different chip SKUs. One reason for that is different parts of the chip consume different amounts of area on different SKUs, i.e A18 Pro has 24MB of SL cache while the M4 only has 8MB, but products that have an M4 can have up 4 Thunderbolt 4 ports whereas an A18 Pro only gives you 1 Gen 2 USB 3.2 port. Sometimes they’re trying to optimize for the thermal capacity of the product the chip will go into as well, and sometimes (from what I understand) they segment their SKUs differently like we saw last year. The M4 Pro is appealing for more people than the M3 Pro was, kind of like how the regular iPhone 16 is more appealing this year.

On the plus side, last year they started binning the memory bus of the binned M3 Max model, and this year even the binned M4 Max has a higher memory bandwidth than the highest-end M3 Max last year.

 

[caribbeanblue]

Was not the rumor from Gurman about the mid-tier chip, codename Brava, going to be the M4 Pro; but it would have an interconnect, so two Brava chips would make up the laptop variant M4 Max...?

Therefore:

• M4 Max (laptop variant)
• 28-core CPU (20P/8E)
• 40-core GPU
• 32-core Neural Engine
• 546GB/s UMA bandwidth

That P-core count is way too high in terms of die area and power consumption for the Max SKU. With that spec the non-binned version of the Max chip would cost as much as the Ultra chip of the prior generation and be unsuitable for a 14" MacBook Pro's thermal and power envelope.

 

[Chancha]

That P-core count is way too high in terms of die area and power consumption for the Max SKU. With that spec the non-binned version of the Max chip would cost as much as the Ultra chip of the prior generation and be unsuitable for a 14" MacBook Pro's thermal and power envelope.

This thread now has different uses of the names juggling around. In the above quoted context he meant the M4 Max would be the new M4 Ultra, but still can fit in a laptop thermal envelop. Or that's what I understood the conversation?

 

[caribbeanblue]

This thread now has different uses of the names juggling around. In the above quoted context he meant the M4 Max would be the new M4 Ultra, but still can fit in a laptop thermal envelop. Or that's what I understood the conversation?

I'm not fully aware of the context of the entire thread so if I'm missing anything, I apologize. However I think an Ultra-tier chip in terms of power consumption could only fit in the 16" MacBook Pro, and it would definitely be partially thermally and power limited (like the Max-tier chips partially are in the 14").

If I may say my guess on Apple's chip scaling strategy will be, I think that depends on how high they're still planning to scale up. If they still want to make a chip that is a tier above the Ultra chip, which is the highest tier today, one that will exclusively go into the highest-end Mac Pro, I think they might make the Ultra chip a monolithic chip instead of consisting of two Max chips in the future, and give that chip an interconnect to make the highest tier chip by connecting two Ultra chips, if they don't run into reticle size issues (which might be the whole reason why they made the Ultra chip out of 2 chips to begin with). But they could still figure out how to make it with 4 Max-tier chips, like the previous rumored but supposedly canceled Mac Pro chips made out of 4 M1 Max or 4 M2 Max chips).

 

[Confused-User]

@mslilyelise for the win! So pleased they did this! In retrospect, it makes sense especially if no M4 is coming. In my mind, this move practically confirms Mac17,1 and Mac17,2 are the M5 MacBook Air.

It could just be that the M4 Air is coming early next year and they don't want to sell any devices that can't run Apple Intelligence in the meantime. No way to know until it happens.

Speculation ended: m4max is basically m3 max on another node but clocked higher and some minor tweaks?

Um, no. The differences between the two are enormous. The CPU and GPU cores, NPU, display controllers, TB controllers, memory controllers, etc... almost everything has seen meaningful improvements.

I am a little disappointed we're not seeing more P cores though.

Now that we know M4 Max is 12P + 4E, what does the P-core cluster look like? Does M4 Max have two 6P clusters different from the 5P ones in M4 Pro, or even the 14-core variant M4 Pro has two P cores fused off?

From the “die shot” Apple gives in the press release, can we know whether M4 Max supports UltraFusion?

I still find it hard to believe M4 Ultra, if exists, will be a new design rather than two Max fusion. For such a large chip selling at very low volume, the cost must be too high, no?

There is no way to know what the cluster configuration is until we have one in hand (to run cache and latency tests), or a real die shot. Apple did not provide any die shots. In fact without die shots we still may not know if the Pro is a chop of the Max or not (if we're surprised by 3x4, then we'll know it's not, but with 2x6 we won't).

 

[mslilyelise]

@mslilyelise for the win! So pleased they did this! In retrospect, it makes sense especially if no M4 is coming. In my mind, this move practically confirms Mac17,1 and Mac17,2 are the M5 MacBook Air.

Hey they doubled the memory in the Air! They announced it at the end of the Pro video! That’s nice. While watching the video on the Pro today I was feeling kinda bummed when they said this was the last day of announcements, cause I figured we’d have to wait, but the whole lineup basically got the 16GB RAM bump. And I mean you could argue they should have done that a couple years ago, but at least it’s there now. Pushes me much closer to buying a new Air since I won’t have to pay for the 16GB upgrade. 😆 A new Mini for my desktop work and a new Air are in my future now.

 

[Confused-User]

If I may say my guess on Apple's chip scaling strategy will be, I think that depends on how high they're still planning to scale up. If they still want to make a chip that is a tier above the Ultra chip, which is the highest tier today, one that will exclusively go into the highest-end Mac Pro, I think they might make the Ultra chip a monolithic chip instead of consisting of two Max chips in the future, and give that chip an interconnect to make the highest tier chip by connecting two Ultra chips, if they don't run into reticle size issues (which might be the whole reason why they made the Ultra chip out of 2 chips to begin with).

You are describing exactly what Gurman claims is going to happen middle of next year.

 

[tenthousandthings]

Hey they doubled the memory in the Air! They announced it at the end of the Pro video! That’s nice. While watching the video on the Pro today I was feeling kinda bummed when they said this was the last day of announcements, cause I figured we’d have to wait, but the whole lineup basically got the 16GB RAM bump. And I mean you could argue they should have done that a couple years ago, but at least it’s there now. Pushes me much closer to buying a new Air since I won’t have to pay for the 16GB upgrade. 😆 A new Mini for my desktop work and a new Air are in my future now.

Yes, that’s what I was referring to. You called it — I don’t recall anyone else suggesting it in endless discussion around here about it.

 

[novagamer]

Now that we know M4 Max is 12P + 4E, what does the P-core cluster look like? Does M4 Max have two 6P clusters different from the 5P ones in M4 Pro, or even the 14-core variant M4 Pro has two P cores fused off?

From the “die shot” Apple gives in the press release, can we know whether M4 Max supports UltraFusion?

I still find it hard to believe M4 Ultra, if exists, will be a new design rather than two Max fusion. For such a large chip selling at very low volume, the cost must be too high, no?

One thing to keep in mind is that Apple needs these for their own use now with the push to private cloud compute. That may justify some pretty nice R&D expenses that trickle down into the low-volume high end.

 

[name99]

M1 Max to M3 Max are all single-die chips. Why would Apple move from monolithic to fusion at the Max tier?

Not to answer this (sensible, rhetorical) question but to steer this thread back to something vaguely technical, and based on chiplets.

We have been brainwashed by AMD and Intel into seeing chiplets as essentially a technology of extending logic. Why limit yourself in that way? If you think outside the box, what ELSE could you do with chiplets?
Enter https://patents.google.com/patent/US12068324B2
This is an astonishing patent that is so far out the box I've no idea if it will ever become reality.
The idea is that below your main SoC you attach a secondary chip which is kinda like a dramatically simplified FPGA. All it has on it is capacitors and switches on one sort or another. A (LARGE) number of power pins, from all the different IP blocks on the SOC connect to different capacitors on the lower chip.

You then configure this lower "passive FPGA" so that the different capacitors are isolated as appropriate, each right-sized for the expected switching demands of the device. You can even dynamically resize the per-IP-block capacitance if you see that the workload has switched from CPU-intensive to GPU-intensive. It's absolutely insane, and absolutely genius! And the best part - you can build this passive chiplet on a trailing fab, nothing about it demands N3.

Will it ever ship? Who knows? Apple has been polishing the idea since 2013, which on the one hand means it ain't shipped yet (as far as I know) so maybe there is some fundamental problem (like the ill-fated power pad).
On the other hand, maybe it's JUST a cost thing; at some point the costs vs the payoff will work out and we start to see it?
The latest version (I did a quick compare of the 2013 vs 2021 version) seems to suggest that this secondary chiplet under the SoC, because it doesn't need to be that large to hold the desired capacitance, especially w/ deep trench capacitors, could also host "memory" (which, realistically I interpret to mean MRAM or ReRAM, something that can also be fabbed in a trailing process).
And maybe this additional idea will kick the concept into viability?

 

[wmy5]

One thing to keep in mind is that Apple needs these for their own use now with the push to private cloud compute. That may justify some pretty nice R&D expenses that trickle down into the low-volume high end.

Good point. But still, if we assume a slight increase in transistor count from M2 Ultra, M4 Ultra would be a massive >140B transistors chip. Huge chips are inherently expensive to make (yield, etc.).

 

[name99]

Wasn't it Steve Jobs who said something along the lines of, "if we won't get our own lunch someone else will"?

To me it makes very little sense to not roll out say an M3 Max Mac Studio despite the M3 Ultra not being ready or even not existed. But Apple decided to go that route, letting both Studio and Mac Pro rot at M2 gen.

The highest end Apple products have ALWAYS been updated on a slower cadence than the mainstream products.
It's hardly an outrage, or even surprising, that the Ultra tier and the Studios skipped M3 (and will likely skip M5 and so on).

 

[Boil]

As for the Studio/Pro, my prediction is the the Studio will focus on the Ultra chip and the Pro will focus on the "Extreme" These will be chips that will be relevant within a 4-year timespan.

I still want to see an all-new Mac Pro Cube with an Extreme chip...!

one wonders what that hidra chip will be then and if there will be any max version of the studio at all.

My initial thoughts were of Apple using the Hidra chips for their high-end desktops offerings; the Mac Studio & Mac Pro...

But now I could see Apple using the new M4 Max for both the MBP laptops and the "low-end" Mac Studio, thereby allowing the laptop market to "fund" the M4 Max in the (lower overall sales volume) Mac Studio market...?

Sad that there is no actual hidra rumors but maybe you more competent people might have some clue based on the pro and max configs?

My totally non-engineer layman thoughts; single Hidra chip for M4 Ultra, dual Hidra chips for M4 Extreme...

M4 Ultra

• 32-core CPU (24P/8E)
• 96-core GPU
• 32-core Neural Engine
• 256GB LPDDR5X RAM
• 1TB/s UMA bandwidth

M4 Extreme

• 64-core CPU (48P/16E)
• 192-core GPU
• 64-core Neural Engine
• 512GB LPDDR5X RAM
• 2TB/s UMA bandwidth

There could be more RAM, there could be RAM with an inline-ECC set-up, there could be faster RAM if using the 9600MT/s chips...

I still find it hard to believe M4 Ultra, if exists, will be a new design rather than two Max fusion. For such a large chip selling at very low volume, the cost must be too high, no?

One thing to keep in mind is that Apple needs these for their own use now with the push to private cloud compute. That may justify some pretty nice R&D expenses that trickle down into the low-volume high end.

Good point. But still, if we assume a slight increase in transistor count from M2 Ultra, M4 Ultra would be a massive >140B transistors chip. Huge chips are inherently expensive to make (yield, etc.).

Apple using the same Hidra chips in their private cloud compute server farm would allow overall lower costs for the personal workstation end-user market...?

 

[Antony Newman]

Is that also taking into account powering devices over Thunderbolt?

The existing M2 Ultra Mac Studio has 6 x TB4 - and had a maximum continuous power of 370W.

Even if Apple upgrade all the ports to TB5 in their M4 - power consumption of those existing plug-in components should be similar.

This would mean that if Apple are able to dissipate and additional (guestimate) 300W from the same enclosure - they could use the lions share of this for additional ‘M4 compute’.

I speculate that Apple would have the ability to dissipate enough heat to cater for 2 x M2 Ultra‘s in the same enclosire.

If M4 Ultra based on 3nm uses less power than the 5nm M2 Ultra - it would not surprise me if the Mac Studio was also able offer the M4 Extreme (or whatever the top compute offering is in the M4 Mac Pro).

Apple could still differentiate the Mac Pro with up to 8TB of RAM; Something that would not comfortably fit in the Mac Studio enclosure.

 

[Pressure]

The highest end Apple products have ALWAYS been updated on a slower cadence than the mainstream products.
It's hardly an outrage, or even surprising, that the Ultra tier and the Studios skipped M3 (and will likely skip M5 and so on).

Let's test that hypothesis out.

Power Macintosh G3
November 10, 1997
January 5, 1999

Power Mac G4
August 31, 1999
July 19, 2000
January 9, 2001
January 28, 2002
August 13, 2002

Power Mac G5
June 23, 2003
June 9, 2004
April 27, 2005
October 19, 2005

Mac Pro
August 7, 2006
April 4, 2007
January 8, 2008
March 3, 2009
December 4, 2009
July 27, 2010
June 11, 2012
December 19, 2013
June 3, 2019
June 5, 2023

Judging by the data I would argue the slower cadence or negligence of the high-end is a new thing in the post Jobs era and it hasn't, in fact, always been like that 😅

 

[Chancha]

In the latest Intel Mar Pro iterations, one could attribute it to Xeon being slower to move a generation as well. And Apple famously admitted they designed itself into the corner. But with the Mac Studio it looks like they control almost the whole stack. The M3 Max exists on the MBP, with similar if not the same TDP envelope, sitting inside a chassis with hardly anything improved over the M2 Max MBP. And yet we don't see it in the Studio.

Marketing wise it is easy to see they don't want to make the Studio M2 Ultra sitting along side an M3 Max looking too embarrassing, and by proxy the Mac Pro too. There is otherwise no technical reason that the chip can't also be sold inside a Studio.

 

[Basic75]

I think they might make the Ultra chip a monolithic chip instead of consisting of two Max chips in the future

That would push far beyond the current reticle limit, not going to happen anytime soon.

 

[MayaUser]

So next year Macbook air and mac studio/pro will be probably on M5 family since Apple presented the way they did the M4 Max

 

[Dulcimer]

So next year Macbook air and mac studio/pro will be probably on M5 family since Apple presented the way they did the M4 Max

How’d they present the M4 Max that gives you that impression?

I know there were some leaked M5 Mac identifiers in a recent Sequoia build (https://forums.macrumors.com/threads/found-m4-and-m5-mac-identifiers.2432863/) which could be M5 MBAs.

 

[MayaUser]

deleted

 

[Xiao_Xi]

So next year Macbook air and mac studio/pro will be probably on M5 family since Apple presented the way they did the M4 Max

Would it make sense for Apple to release new Mx SoCs every year, but not update all products with them?

 

[Basic75]

Would it make sense for Apple to release new Mx SoCs every year, but not update all products with them?

The iMac skipped the M2, the Mac mini, Pro and Studio skipped the M3, must have made some sense to Apple.

 

[AgentMcGeek]

How likely is it that the next MBA actually skips M4 and gets the M5 straight away?
Since this is the most popular Mac, I doubt that they would have enough volume for it. Going the iPad Pro route first makes sense in that regard.

 

[yuda398]

My totally non-engineer layman thoughts; single Hidra chip for M4 Ultra, dual Hidra chips for M4 Extreme...

M4 Ultra

• 32-core CPU (24P/8E)
• 96-core GPU
• 32-core Neural Engine
• 256GB LPDDR5X RAM
• 1TB/s UMA bandwidth

M4 Extreme

• 64-core CPU (48P/16E)
• 192-core GPU
• 64-core Neural Engine
• 512GB LPDDR5X RAM
• 2TB/s UMA bandwidth

There could be more RAM, there could be RAM with an inline-ECC set-up, there could be faster RAM if using the 9600MT/s chips...





Apple using the same Hidra chips in their private cloud compute server farm would allow overall lower costs for the personal workstation end-user market...?

I think you are right about “single Hidra chip for M4 Ultra, dual Hidra chips for M4 Extreme” but the core counts are too big to fit in one chip. And also if the Ultra and Extreme are designated to power desktops only I think the E-core are redundant.

I think it will be along the lines of:
18-20 core CPU (all being P-cores)
60-70 core GPU

 

[caribbeanblue]

Good point. But still, if we assume a slight increase in transistor count from M2 Ultra, M4 Ultra would be a massive >140B transistors chip. Huge chips are inherently expensive to make (yield, etc.).

M3 Max is already 92B transistors. If M4 Ultra continues to have double the hardware of the Max chip, it will easily have over 180B transistors.

You are describing exactly what Gurman claims is going to happen middle of next year.

Which part? I haven't seen Gurman talk about exactly how they will go about making the next Ultra chip or a chip one tier above that. The only specification he provided that I've seen was that one of (or the only?) chips that will go into the Mac Pro will have a 512GB memory option, which is 4x what the M4 Max supports, which makes me think they'll either offer somehow a 4x higher capacity memory option on the Ultra chip than the Max chip (and perhaps also a 256GB option on the Max for desktop hardware), or we'll see a chip tier on the Mac Pro that has either exactly or nearly 4x more hardware than the Max chip in every way.