If my view of past Apple Silicon still aligns with reality, an Air, an iMac or an iPad Pro is on the way. 🙂
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.
Apple Reveals M6 as First-Ever 2nm Chip
- Thread starter MacRumors
- Start date
- Sort by reaction score
Given their Mac history, there are plenty of folks still running 10 yr or older systems. 🙂
BorisDG
macrumors 6502a
That was pretty much the same for M5, since it's using the new (same) Apple 10 GPU (A19/M5). New GPU arch with M7 (Apple 11). It's pretty much in specific scenarios.
Basic75
macrumors 68030
That much is confirmed?
The M5 Pro and Max have a 6 + 12 + 0 configuration, the M5 is 4 + 0 + 6 and the A19 and A19 Pro are 2 + 0 + 4.
I have a hard time believing this. The E-cores are for low-power efficiency (Intel calls them LP-E) while the new mid-tier Apple P-cores (Intel calls them E-cores) are for area efficiency: high throughput for high number of threads. Surely using the M5 P-cores as E-cores in the M6 would make them too wide for the low-power role.
Yes, but my argument was mostly about the ratio between P-cores and E-cores in the M6 while your M4 comparison is about S-cores and E-cores (if we retro-actively apply Apple's new naming to the M4 to keep the terminology consistent). What do you need 6 low-power E-cores for? Dropping from 4 to 2 S-cores should have given space for more P-cores, with 2 or 4 E-cores. Intel's 2 + 8 + 2 makes more sense to me. That's why I'm wondering what Apple is thinking here. Are they just stuffing in E-cores for higher numbers without real benefits?
Thank you Apple for the rubbish "Super" core rename that makes talking about this all the more confusing.
Last edited:
Basic75
macrumors 68030
And it would be nice to have single-core numbers for all types of cores, not just the fastest ones.
Populus
macrumors G3
Exactly. I was expecting that, for such a price, they would start at 512GB as other Macs. But adding the upgrade price it ends up easily in the 2000€ territory.
Populus
macrumors G3
Something tells me this is a minor update compared to the M7 coming next year, which will come with a memory bandwidth of 240GB/s. For the M6 I was expecting at least a bandwidth of 200GB/s.
Of course the two extra GPU cores (expected) and the two extra CPU cores (unexpected) make this little machine even more interesting. But that’s it?
Oh, and the dual Neural Engine. Correct me if I’m wrong but most inference tools prefer to use the GPU rather than the NPU. We’ll see if that changes with the new 32 core NE.
I’d love to have the new chip but at that price I’m happy with my M5 MacBook Air. I know I’ll be tempted to sell it in order to get the M6 MacBook Air next year but… honestly, if the M6 mini with 1TB/32GB costs 2059€, I cannot imagine how much the 1TB/32GB M6 MacBook Air will be… 2400€?
Of course the two extra GPU cores (expected) and the two extra CPU cores (unexpected) make this little machine even more interesting. But that’s it?
Oh, and the dual Neural Engine. Correct me if I’m wrong but most inference tools prefer to use the GPU rather than the NPU. We’ll see if that changes with the new 32 core NE.
I’d love to have the new chip but at that price I’m happy with my M5 MacBook Air. I know I’ll be tempted to sell it in order to get the M6 MacBook Air next year but… honestly, if the M6 mini with 1TB/32GB costs 2059€, I cannot imagine how much the 1TB/32GB M6 MacBook Air will be… 2400€?
Last edited:
Populus
macrumors G3
Why not? Apple OS updates are already on the 7 year range. It wouldn’t be strange to stretch that to 8 years of macOS support, given how powerful today’s Macs are. Add in the regular two extra years of security updates and voila, there you have your 10 year old support for the M6 Mac mini.
As I understand it the NPU is used for inputing and "understanding" the initial prompt - the "prefill" stage of processing - whereas the GPU is then used for the subsequent token generation for the output. The doubled NPU performance should reduce time to first (output) token and from then on my understand is that it's the GPU taking the load.
I'm guessing here but I assume the NPU impact will be most noticeable where long context windows are involved - input prompts with image data, very long documents to review etc. There the performance gain might be quite noticeable and welcome in reducing the time before the system responds whereas it might have much less of an impact (perhaps almost imperceptible) for simpler prompts such as "How much more powerful is an M6 vs an M5?".
chucker23n1
macrumors G3
The alternative explanation would seem to be that the M5 Pro / Max / Ultra's performance cores (the middle tier) are cores that do not exist anywhere else; that they were custom-designed just for those three SoCs. The M5 and A19 Pro don't use them; they only use the super and efficiency cores.
That's why one hypothesis was that they're actually the next generation's efficiency cores.
Well, historically, the base M series has had 4p4e (M1), 4p4e, 4p4e, 4p6e, 4p6e (M5) configs. With the new branding, 4s4e (M1), 4s4e, 4s4e, 4s6e, 4s6e (M5).
So they're going from 4s6e to 2s4p6e.
So that's already more non-efficiency cores than they've ever had on the base.
Well, macOS is full of background processes. The M1 Pro had an 8p2e setup (now 8s2e), and that arguably led to poorer battery life.
I'm now on the M5 Pro, with an 6s12p setup, and looking at Stats, six of those p-cores seem to constantly be doing stuff, albeit at below 30% each. Six others seem to be idling. If I then boot a Windows VM, everything's busy: all 12 performance cores (maxing out around 50% each), and all super cores (maxing out at 100%).
Yeah, it's not great. Especially when they launched the original M5 with a different naming, because they didn't want to reveal the middle tier.
Populus
macrumors G3
Yes, I’m also confused. So the “super” cores are the new performance cores, or the middle ones?
And what are the ones on the base M5, super or performance cores?
chucker23n1
macrumors G3
The base M5 has super and efficiency cores, but Apple didn't want to reveal that at the time, so they called them performance and efficiency cores.
Starting with M5 Pro, Max, Ultra, M6, "performance" is now the middle tier.
Populus
macrumors G3
So basically the base M5 has six efficiency and four “super-performance” CPU cores, while the M6 has six efficiency, four middle-performance, and two super-performance CPU cores. Is this correct? I’ll bookmark his message to have it at hand…
What I do not understand, is that Apple is aiming the two super cores on the M6 towards single threaded tasks… I guess the super cores on the M6 are different than the ones on the M5…
The s-cores are aimed at the same things apple used to aim their performance cores at, given the super ones are just new iterations of the former p-cores. Apple apparently decided they don’t need to pay the area premium for many high-perf cores, especially at the sub-pro/max end.
BTW, I much prefer to call the present apple bigLITTLE setup p-core/m-core/e-core, as it preserves the continuity from the previous setup.
nathan_reilly
macrumors 6502
If we are gonna start making up numbers than anything is possible. Let me know how your mac mini is running in 2036.
Populus
macrumors G3
Okay then. Not especulations. Based on what we’ve seen so far, new macs will probably have 9 years of support like the longest supported machines instead of the 10 the other member said: 7 years of software releases and two years of additional security updates.
But that is all subject to change, as we’re entering a new era for the company with a new CEO, the focus is shifting quickly from traditional computing to AI technology... This means that support could be extended once more, or be shortened because the lack of compatibility with future foundational features… yeah, we don’t actually know. But that’s what we do here, isn’t it? Speculation.
If what you’re meaning is that regardless of the years of software support a mac receive, there’s a performance toll that will penalize the machine over the years, I’ll give you that. However, I cannot see the future, but if I have to use the past to infer what could happen in the future, let me tell you that my 12 years old mini, being an old Intel dual core processor and having only 8GB of RAM, still runs macOS 12 quite well.
chucker23n1
macrumors G3
Correct.
We don't really know yet what core designs of them use. Usually, each generation introduces entirely new cores — so, for example, each M4 performance core is faster than each M3 performance core, and each M4 efficiency core is faster than each M3 efficiency core. It's not yet clear how this works with the M5 and M6, where they inserted a third tier. Is that middle tier an entirely new design? Is it a simpler version of the "super" design? Etc.
Clever. Like when Uber Eats introduced fast delivery which turned out to just be normal delivery times.
chucker23n1
macrumors G3
Yeah, I wish Geekbench 7 would've introduced that. We're increasingly in a world of heterogeneous cores, after all.
Agreed, I could be wrong but a base M4 Mac mini can run most local models reasonably for one person.
It’s good have more headroom still but unless you’re working with ridiculous models you should be fine for a few more years.
It’s good have more headroom still but unless you’re working with ridiculous models you should be fine for a few more years.
Populus
macrumors G3
Hey! I've been thinking about this.
At first, I thought like you, like "they are adding even more performance cores to the base chip". However, I think the M6 will be quite balanced to be used on laptops, because if you think about it, they are going from a chip with half the cores being efficient in previous chips (M1, M2, M3) to again half the cores being efficiency chips on the M6, going from the initial 4p4e, to 6p6e. After a couple of node shrinks, I think they were finally capable of increasing in 2 cores both the efficient ones and the performance ones.
But not only that. From this new 50% performant cores on the M6, four of them (the p cores) are a new type of core -introduced on the M5 Pro IIRC- that is a mix between a performance core with a reduced power consumption. So, this 6 non-efficiency cores are, overall, more efficient than the 4 p-cores on the previous chips. All of this helped by the tremendously reduced power consumption of the N2 process, makes the 6 e-cores even more efficient, and the remaining 6 non e-cores probably more efficient than the previous 4 p-cores.
Okay, if you didn't get what I was trying to say, it is totally fine. I myself I'm re-reading it and I'm not sure I quite understand what I wrote... but I think I'm confident enough to say the M6 will probably be more efficient, fresh and cool, than the M5.
*That's without taking into account the 2 extra GPU cores and the extra 16 NE cores... Because that reduction in overall CPU power draw could be used to increase the power draw from a bigger GPU and a double sized neural engine. but that's another story...
chucker23n1
macrumors G3
To be clear, I was mainly responding to "What do you need 6 low-power E-cores for?", with "they've had that many since the M4", two generations ago.
Indeed, they're now inserting that middle tier, which
- in some scenarios where they'd previously need an s-core, they can now save on power
- in some scenarios where they'd previously use an e-core, they can now finish the task faster (which, again, saves on power)
I think the main issue apple are solving with the mid-cores is not so much power efficiency as it is area. Dies are only this big (yes, even with chiplets) and Apple need real estate for those AI circuitries. They’ve already addressed the gpu area problem, but it turned out AI was best addressed at multiple levels at once — at the cpu/gpu/npu. That area has to come from somewhere.
