M4+ Chip Generation - Speculation Megathread [MERGED]

[Basic75]

It seems like the cache would follow the GPU, but it also makes the Pro-> Max GPU tiling more sus because the bandwidth of the Pro smaller pro should have half the smaller Max, but isn't. But then again the smaller Pro has less CPU cores, so maybe the memory controller follows it?

Just as they disable a few GPU cores on the lower bin surely they can independently disable a couple of memory channels.

 

[leman]

It seems like the cache would follow the GPU, but it also makes the Pro-> Max GPU tiling more sus because the bandwidth of the Pro smaller pro should have half the smaller Max, but isn't. But then again the smaller Pro has less CPU cores, so maybe the memory controller follows it?

Could be two memory controllers disabled and two RAM modules less, or could be that there are multiple chopped die variants. On the previous models, is the "big" Max physically the same as the lower-end Max, or are they different dies?

 

[Populus]

Hey, do you mind if I ask something that, while “on topic”, is not what you’re discussing now?

It’s about the M5 chip on the MacBook Air. I know this is pure speculation and we better wait for next Wednesday for the first benchmarks but:

Given that the M5 is clocked higher than the M4, and has a more powerful GPU… do you think it will either overheat or throttle? I’m considering a MacBook Air now instead of the Mac mini, and my only concern is the lack of active cooling.

Thank you.

 

[leman]

Hey, do you mind if I ask something that, while “on topic”, is not what you’re discussing now?

It’s about the M5 chip on the MacBook Air. I know this is pure speculation and we better wait for next Wednesday for the first benchmarks but:

Given that the M5 is clocked higher than the M4, and has a more powerful GPU… do you think it will either overheat or throttle? I’m considering a MacBook Air now instead of the Mac mini, and my only concern is the lack of active cooling.

Thank you.

The M-series in Airs thermal throttle by design, that’s how they manage power. What exactly is your worry? That the laptop will be excessively hot to touch or that the performance will be disappointing? I doubt either is likely, but as usual one needs to empirically test your use case. Sounds to me like one of the situations Apples returns policy was designed for.

 

[diamond.g]

The M-series in Airs thermal throttle by design, that’s how they manage power. What exactly is your worry? That the laptop will be excessively hot to touch or that the performance will be disappointing? I doubt either is likely, but as usual one needs to empirically test your use case. Sounds to me like one of the situations Apples returns policy was designed for.

For short bursty workloads it should be fine, sustained will be interesting...

 

[DrWojtek]

For short bursty workloads it should be fine, sustained will be interesting...

The throttling is nothing compared to how it was with Intel. I play BG3 on my M4 MBA. I don’t notice the throttling, though I’m sure it happens.

 

[crazy dave]

Which Qualcomm cores are these? The Oryon cores for Snapdragon 8 Elite is just two levels.

"...Rather than a mix of large, medium, and small CPU cores as it has used in the past, the 8 Elite has two “Prime” cores for hitting that high peak clock speed, while the other six are all “Performance” cores that peak at a lower 3.53 GHz. ..."

Qualcomm brings laptop-class CPU cores to phones with Snapdragon 8 Elite

New CPU cores replace more complicated mix-and-match of CPU architectures.

arstechnica.com

if talking about the non-Nuvia evolution stuff that is still on Arm design IP ... that isn't a good match to what Apple is doing here.

Intel also is a couple generations away from 'unifed core' implementations.

This appears more like AMD's cloud cores where they stripped the max single thread drag racing elements out of the design and shrink the footprint of the 'large' core into something laid out smaller. Save space so can have more cores but don't 'give away' much on the what is implemented. The Apple E-cores are in a very different zone of giving up space and more enhanced lower energy savings.

Apple is never using all three in same die. Super+E or Super+P but no Super+P+E. In any one system implementation there is no 'middle'. [ Having 3 different one makes the OS scheduling more painful, for probably little impact. ]

Middle core doesn't refer to needing 3 levels in the same chip. It refers to a design philosophy or role of that core in the SOC. As you said yourself, Middle cores' primary goal is to achieve power and area efficiency in highly multithreaded scenarios while Efficiency cores are primarily meant to achieve low power operations. I agree with you. These can obviously overlap as we'll see below. Thus, Intel "E-cores" were Middle cores before the introduction of LP-E cores and are still so when LP-E cores aren't present. Intel doesn't need to have a unified core design to be compared to here. Like AMD's "C-cores", Intel also uses its "E-cores" in a density server chip design. In fact, if the rumors are accurate for the new Apple Middle core description, then the Intel E-core and the new Apple Middle core are more similar in design ethos than they are to AMD's "C-core" - though it's a matter of degree since they are all ways to build Middle cores. Many ARM designs including Qualcomm and Nvidia include Middle cores, not E-cores - though I'll admit that the Qualcomm one might be considered more of an E-core given the SOC design. It's still clocked significantly higher than most E-cores and appears to be larger (remains to be seen), but it may still be smaller and is definitely clocked less than most Middle cores as well. With only 6 of them but 12 "Prime" cores it's arguable that it's still functioning closer to the role of an E-core in the SOC. The 10 A725s on the GB10 and upcoming N1X however is more of a density design wrt to the 10 X925 performance cores.

 

[crazy dave]

Hey, do you mind if I ask something that, while “on topic”, is not what you’re discussing now?

It’s about the M5 chip on the MacBook Air. I know this is pure speculation and we better wait for next Wednesday for the first benchmarks but:

Given that the M5 is clocked higher than the M4, and has a more powerful GPU… do you think it will either overheat or throttle? I’m considering a MacBook Air now instead of the Mac mini, and my only concern is the lack of active cooling.

Thank you.

To expand on @leman's and @diamond.g's points here, here's a sustained power efficiency chart:

Previously with the M3 and older chips, Apple didn't let its base chip unleash as it were even in environments where thermals were less of a concern. So it really depends on how you want think about it - the M5 in the Mini and MBP will likely net you significantly more sustained performance than the M5 in the Air, but the M5 in the Air will still be significantly better than the M3 or M4 in the Air. For ST workloads requiring sustained performance, device thermals don't matter as much for Apple cores (not pictured).

For burst MT workloads, the differential of the Air should make less difference. For instance, the difference between the various M4 chips in GB MT is basically zilch compared to the sustained difference and the Mini is even near the bottom indicating most of the difference here is random noise not thermals.

So what's your workload? Burst or sustained? ST or MT?

 

[deconstruct60]

Could be two memory controllers disabled and two RAM modules less, or could be that there are multiple chopped die variants. On the previous models, is the "big" Max physically the same as the lower-end Max, or are they different dies?

Disabling controllers with defects is way cheaper than coming up with three GPU chiplets. (if Pro and Max were two monolithic chips. If go to three chiplets are replacement.. can get cheaper. Go to four chips to replace two chips is heading back up in costs. ) The "max" GPU chiplet is still going to be relatively large. Going to have dies with broken GPU core and broken memory controllers .. just combine those into one alternative binning. Some have broken one or the other (and on very good days with right demand nothing is broken substantially and just turn them off to get higher volume to match demand. )

It is also not just defects. Tightly bundled memory is also a system cost point. For what Apple charges for memory, two less module is a significant enough change in system price. It is margin management also. (same tract with going with just one alternate SKU. Cleaner inventory, etc. )

 

[Populus]

The M-series in Airs thermal throttle by design, that’s how they manage power. What exactly is your worry? That the laptop will be excessively hot to touch or that the performance will be disappointing? I doubt either is likely, but as usual one needs to empirically test your use case. Sounds to me like one of the situations Apples returns policy was designed for.

Well, you’re right, I think the best thing to do will be to test it with real workloads. Most of them aren’t that demanding, but I have a couple of AAA games that may either make it throttle or make the machine uncomfortably hot (Myst & Riven).

Also I will use Handbrake quite often (CPU) and I definitely plan on experimenting with local LLMs (GPU) so… I don’t know.

I won’t mind having the Air closed in clamshell mode, I’d only there was a good stand with a dissipator or a fan…

 

[Populus]

To expand on @leman's and @diamond.g's points here, here's a sustained power efficiency chart:

View attachment 2610002

Previously with the M3 and older chips, Apple didn't let its base chip unleash as it were even in environments where thermals were less of a concern. So it really depends on how you want think about it - the M5 in the Mini and MBP will likely net you significantly more sustained performance than the M5 in the Air, but the M5 in the Air will still be significantly better than the M3 or M4 in the Air. For ST workloads requiring sustained performance, device thermals don't matter as much for Apple cores (not pictured).

For burst MT workloads, the differential of the Air should make less difference. For instance, the difference between the various M4 chips in GB MT is basically zilch compared to the sustained difference and the Mini is even near the bottom indicating most of the difference here is random noise not thermals.

So what's your workload? Burst or sustained? ST or MT?

Thank you for your chart, it’s quite descriptive: the M5 has outstanding performance per Watt if I’m not reading it wrong, right?

 

[name99]

Which Qualcomm cores are these? The Oryon cores for Snapdragon 8 Elite is just two levels.

"...Rather than a mix of large, medium, and small CPU cores as it has used in the past, the 8 Elite has two “Prime” cores for hitting that high peak clock speed, while the other six are all “Performance” cores that peak at a lower 3.53 GHz. ..."

Qualcomm brings laptop-class CPU cores to phones with Snapdragon 8 Elite

New CPU cores replace more complicated mix-and-match of CPU architectures.

arstechnica.com

if talking about the non-Nuvia evolution stuff that is still on Arm design IP ... that isn't a good match to what Apple is doing here.

Intel also is a couple generations away from 'unifed core' implementations.

This appears more like AMD's cloud cores where they stripped the max single thread drag racing elements out of the design and shrink the footprint of the 'large' core into something laid out smaller. Save space so can have more cores but don't 'give away' much on the what is implemented. The Apple E-cores are in a very different zone of giving up space and more enhanced lower energy savings.

Apple is never using all three in same die. Super+E or Super+P but no Super+P+E. In any one system implementation there is no 'middle'. [ Having 3 different one makes the OS scheduling more painful, for probably little impact. ]

We have no idea what Apple will be doing! Will the A20 have 2P, 2M and 2 E cores? Who knows?
Can an M-core clock down low enough (hence low enough energy to be feasible for an aWatch?) Who knows?

Speculation is fun. But strong claims about what will or won't happen are just foolish! As are claims about how hard the OS-scheduling is.
Learn from the strong GPU claims (Apple can NEVER create a decent GPU, all the patents are taken!)
Learn from the strong cellular claims (Apple can never compete with a QC modem. RF is magic and hard!!)

 

[name99]

But Apple said.

They are talking bout both the Pro and Max there. ( Seems more like Fusion 2.0 than "all new" Fusion because combining two dies is what they have done with the Ultra for a couple of iterations. The interconnect tech is likely better and the size of the individual dies is getting closer to actual chiplet oriented design, but generally the same approach done at higher overall unit volume. The dies mated not being exactly the same size is a big leap for "all new" label. )

There are many more concerns to connecting two (or more) chiplets together than just the physical connection.
Obviously you need that and you want it to be fast, low power, low area.

But you also have concerns like how the setup boots? How is power control and clock handled between the two chiplets? How are interrupts distributed? How is coherency handled? What does the NoC look like? etc etc

It's quite feasible that, having learned from iteration one, Apple massively redesigned all these elements to fix all the flaws and limitations they discovered.
We have precedent. There have been at least three very differently architected designs for the memory controller/SLC/coherency system as the A then M series evolved.

 

[name99]

I do find it interesting the base M5 Max has a different bandwidth config than the full part, even though the Pro doesn't have the same "limitation". Do we know if the memory controllers are on the GPU, IO, or CPU die?

They will obviously be on the GPU die. Along with the SLC.
This is just common sense given that it's the GPU's that demand that bandwidth.

 

[name99]

I do find it interesting the base M5 Max has a different bandwidth config than the full part, even though the Pro doesn't have the same "limitation". Do we know if the memory controllers are on the GPU, IO, or CPU die?

Ah, I think I understand what's confusing people.
Consider the Pro class of machine. That has four of let's call them memory controllers. (A has 1, M has 2.)
If we take the base memory bandwidth of the available DRAM as, say, 50 GB/s, that gives our Pro 200 GB/s.
All easily understood.

BUT
when it comes to segmentation, starting with the M3, Apple added a wrinkle. The lower Pro is deficient in THREE ways
- you lose CPU cores
- you lose GPU cores AND
- you lose a memory controller.
(If you understand the issue, this is not quite trivial. You now need to hash addresses across three slices, not four, so you can't just hash by two physical address bits fairly high in the address. Apple actually have a patent for a circuit to do the relevant hashing.)

I THINK at the M3 generation they didn't do with the Max.
But with the M4 they did. Look at the M4 Pro and Max (if the Apple pages are still up.)
In both cases the lesser version
- loses CPU cores
- loses GPU cores AND
- has 3/4 memory bandwidth because one of the four (or two of the eight) memory controllers are disabled.

Does this extend to also losing a quarter of SLC capacity? I believe it does – in terms of HW (and the memory hashing) that's really the only option that makes sense.
To the extent that this is driven by yield binning (rather than mostly driven by market segmentation) that also makes sense. The area of an SLC slice is reasonably larger, larger than the memory controller, so if something's going to fail because of yield bad luck, it's more likely to be the SLC slice than the associated memory controller.

M5 Pro and Max are just following this tradition.

 

[crazy dave]

Thank you for your chart, it’s quite descriptive: the M5 has outstanding performance per Watt if I’m not reading it wrong, right?

Aye, it's getting a bit cluttered ... but yes here are the numbers for the M5 (light pink):

M5 ST: 200pts , 16.1pts/W
M5 MT: 986pts, 29.51pts/W => this is from the MBP

You'll notice the M4 (orange) has 3 points on the MT chart for the 13" Air, 15" Air, and either MBP or mini I can't remember and while the performance goes down in the Airs, the efficiency is incredible compared to the M3 (50% more courtesy of 2 extra E cores plus microarchitecture changes) - peaking at 48pts/W in the 13" Air.

 

[Populus]

Aye, it's getting a bit cluttered ... but yes here are the numbers for the M5 (light pink):

M5 ST: 200pts , 16.1pts/W
M5 MT: 986pts, 29.51pts/W => this is from the MBP

You'll notice the M4 (orange) has 3 points on the MT chart for the 13" Air, 15" Air, and either MBP or mini I can't remember and while the performance goes down in the Airs, the efficiency is incredible (courtesy of 2 extra E cores) - peaking at 48pts/W in the 13" Air.

You say the M4 on the MBA has 2 extra efficiency cores compared to the M4 Mac mini or MBP? Is the M5 on the MacBook Air the same as that on the MacBook Pro?

 

[crazy dave]

You say the M4 on the MBA has 2 extra efficiency cores compared to the M4 Mac mini or MBP? Is the M5 on the MacBook Air the same as that on the MacBook Pro?

Yikes, reread my post. Sorry for being unclear, I've edited it now for clarity. The base M4 has 2 more efficiency cores than the base M3. That is why the dark pink M3 only has an efficiency rating 33 pts/W at the same power level that the M4 nets 48 pts/W (courtesy of not just those 2 extra E-cores but also the microarchitecture changes from the M3 to M4 cores of course, but the two extra E-cores is the big impact change here). I expect the M5 to be as efficient as the M4 in the Air as well (if not more so) when it gets tested, but with better performance to boot.

 

[Populus]

Yikes, reread my post. Sorry for being unclear, I've edited it now for clarity. The base M4 has 2 more efficiency cores than the base M3. That is why the dark pink M3 only has an efficiency rating 33 pts/W at the same power level that the M4 nets 48 pts/W (courtesy of not just those 2 extra E-cores but also the microarchitecture changes from the M3 to M4 cores of course, but the two extra E-cores is the big impact change here). I expect the M5 to be as efficient as the M4 in the Air as well (if not more so) when it gets tested, but with better performance to boot.

Yeah… the thing is, if I play any AAA game (sparsely) or make use of local LLMs, it’s gonna get hot… but I know, that’s a flaw of the MacBook Air.

On the other hand, waiting for the M6 MacBook Air is tempting, as it will be the last LCD MBA and the last one my eyes are going to tolerate… but the international situation is pushing me to get my Mac NOW, before oil prices and NAND & RAM scarcity reflect on the end consumer prices, before China takes Taiwan, etc… I feel like I’m being too risky waiting but ideally the M6 will be much fresher due to being a node shrink.

 

[crazy dave]

Yeah… the thing is, if I play any AAA game (sparsely) or make use of local LLMs, it’s gonna get hot… but I know, that’s a flaw of the MacBook Air.

On the other hand, waiting for the M6 MacBook Air is tempting, as it will be the last LCD MBA and the last one my eyes are going to tolerate… but the international situation is pushing me to get my Mac NOW, before oil prices and NAND & RAM scarcity reflect on the end consumer prices, before China takes Taiwan, etc… I feel like I’m being too risky waiting but ideally the M6 will be much fresher due to being a node shrink.

I'm assuming portability is also something you want? Otherwise mini+monitor of your choice might be ideal for your use case and especially eye issues - that way you can keep the monitor you already know you like if you want to switch computers. The M5 mini probably won't be out until the summer, but, while I can't guarantee anything, unless the world ends before then, in which case we will all have bigger problems, I don't think Apple will massively increase the M5 mini's price beyond what it did with the Air and Pro/Max models (and the Air just returned to its M3 price structure with simply better specs) - e.g. effectively knock out the cheapest storage model and make the 2nd cheapest model the new base (and in the Air's case reduce the price by $100).

 

[leman]

Well, you’re right, I think the best thing to do will be to test it with real workloads. Most of them aren’t that demanding, but I have a couple of AAA games that may either make it throttle or make the machine uncomfortably hot (Myst & Riven).

Also I will use Handbrake quite often (CPU) and I definitely plan on experimenting with local LLMs (GPU) so… I don’t know.

I won’t mind having the Air closed in clamshell mode, I’d only there was a good stand with a dissipator or a fan…

If you are using Handbrake a lot, then a model with active cooking might be a better fit. The Air will do fine of course, but the performance won't be as good.

 

[Populus]

If you are using Handbrake a lot, then a model with active cooking might be a better fit. The Air will do fine of course, but the performance won't be as good.

Actually… the Air may be good for active cooking lol

 

[Populus]

I'm assuming portability is also something you want? Otherwise mini+monitor of your choice might be ideal for your use case and especially eye issues - that way you can keep the monitor you already know you like if you want to switch computers. The M5 mini probably won't be out until the summer, but, while I can't guarantee anything, unless the world ends before then, in which case we will all have bigger problems, I don't think Apple will massively increase the M5 mini's price beyond what it did with the Air and Pro/Max models (and the Air just returned to its M3 price structure with simply better specs) - e.g. effectively knock out the cheapest storage model and make the 2nd cheapest model the new base (and in the Air's case reduce the price by $100).

See, my internal struggle between the mini and the air, is the following: for my academic tasks, my iPad Pro is enough 95% of the times, and that’s what I carry with me. For commute: iPad and iPhone. For working from the countryside , having a mini isn’t a problem either: I could grab it, put it on the suitcase and when I arrive there, I have a monitor and a TV, plus another set of keyboard and mouse.

However, some people are encouraging me to go abroad to complete my formation or even work… I still haven’t decided it yet, but if I do it, it’s one of this two things: either go with a MacBook Air, Which is the most convenient solution, or I grab an 80€ HDMI Chinese screen from Amazon just to be able to boot into my Mac mini and then pair it via sidecar with my iPad Pro. Those are my two options.

I cannot play the card of buying and selling in a year, because I really want the top RAM configuration, and that makes it more complicated to sell it without losing a lot of money…

Anyway, I don’t want to extend the off-topic much more.

 

[thenewperson]

then a model with active cooking might be a better fit.

The 2020 Intel Air?!

 

[DrWojtek]

Look up throttling on the M4 MBAs. Like I said, this is not at all the issue it was during the Intel era. I can get back to you, maybe later today, and let you know whether gaming throttles and if so, by how much.