Apple's 'M2' Next-Gen Mac Chip Enters Mass Production, Expected to Debut in Redesigned MacBooks Later This Year

[Bug-Creator]

3nm in production to 2022, that what the rumored schedule was few months ago.. which is the next big gen of Apple Silicon.

Back in the days when Intel's Tick Tock really worked they got as much improvement out of arch improvements as out of fab shrinks.

It is really simple, if it is based on the same tech as the A15 call it M2 regardless of fab used.

If it ist just more of everything A14/M1 called it M1x.

Rumors/timing suggest it will be the 1st....

 

[cmaier]

Back in the days when Intel's Tick Tock really worked they got as much improvement out of arch improvements as out of fab shrinks.

It is really simple, if it is based on the same tech as the A15 call it M2 regardless of fab used.

If it ist just more of everything A14/M1 called it M1x.

Rumors/timing suggest it will be the 1st....

Yep, though it’s a little hard to untangle fab from design. I did *many* shrinks in my life. Typically what would happen is we’d do a timing simulation with the new spice models and see, say, a 10% improvement. Management wanted a 20% improvement. And it turns out the design rules are such that you can’t really linearly shrink everything (you shrink wires and transistors by different amounts, spacings by different amounts than polygon dimensions, etc.). So you redesign anyway. The redesign is based on the same microarchitecture, but you also throw in “easy“ changes proposed by the architects (let‘s add this instruction! Let’s change this bus interface! Let’s double the clock rate to the north bridge!). And you make other changes based on your design ideas (the critical path could be eliminated if the load/store unit pre-computes these two control signals and sends them to the scheduler, but then we need to modify the microarchitecture slightly to deal with invalidating them when there is a cache miss!). In the end you get your 20% and you sort of lose track of how much was given to you by the new node vs. design changes.

 

[docbop]

The info in this article is the same things that were talked about when the M1 chips were first announced. What hasn't been mentioned then and now is what will the RAM increase be on the M2 and beyond. How much RAM is Apple willing to put on a SoC. Going off the SoC chip to access RAM is going to be a major bottleneck on performance.

 

[cmaier]

The info in this article is the same things that were talked about when the M1 chips were first announced. What hasn't been mentioned then and now is what will the RAM increase be on the M2 and beyond. How much RAM is Apple willing to put on a SoC. Going off the SoC chip to access RAM is going to be a major bottleneck on performance.

You are wrong. There is no DRAM on the SoC. The DRAM is in separate chips that are in packages within the overall M1 package. So all RAM access is already off the SoC chip.

Moreover, even if the RAM needed to be moved out of the M1 package, it’s pretty easy to minimize the penalty in almost all cases by simply using a very large system cache.

 

[Unregistered 4U]

I think M2 represents…


M * 2 - Solve for M

 

[tubular]

You are wrong. There is no DRAM on the SoC. The DRAM is in separate chips that are in packages within the overall M1 package. So all RAM access is already off the SoC chip.

Moreover, even if the RAM needed to be moved out of the M1 package, it’s pretty easy to minimize the penalty in almost all cases by simply using a very large system cache.

I'm curious about this because it seems to me that if Apple is aiming at RAM sizes sufficient to replace the Mac Pro's 500GB+ upper end, whenever this decade that happens, then sooner or later you've outgrown any conceivable manufacturable package size. I don't see how it gets done without moving it at some point to a second package.

 

[cmaier]

I'm curious about this because it seems to me that if Apple is aiming at RAM sizes sufficient to replace the Mac Pro's 500GB+ upper end, whenever this decade that happens, then sooner or later you've outgrown any conceivable manufacturable package size. I don't see how it gets done without moving it at some point to a second package.

I don’t think they are all that worried about supporting 500GB+ anytime soon.

In any event, there is no problem moving the main RAM off the main package. Most of the time, what you want will be in the in-package cache. (Assuming you properly design the cache and give it the right size). If you have enough bandwidth between the highest-level of cache and RAM, you then the penalty that you pay for a (rare) cache miss isn’t too debilitating. There will always be workloads that suffer a bit because they randomly access memory in such a way that there are many cache misses, but we’re no worse off than we are today - there aren’t a lot of CPUs out there that keep all their RAM abutting the CPU.

 

[neilw]

I don’t think they are all that worried about supporting 500GB+ anytime soon.

In any event, there is no problem moving the main RAM off the main package. Most of the time, what you want will be in the in-package cache. (Assuming you properly design the cache and give it the right size). If you have enough bandwidth between the highest-level of cache and RAM, you then the penalty that you pay for a (rare) cache miss isn’t too debilitating. There will always be workloads that suffer a bit because they randomly access memory in such a way that there are many cache misses, but we’re no worse off than we are today - there aren’t a lot of CPUs out there that keep all their RAM abutting the CPU.

Furthermore, anything marketed as "Mac Pro" absolutely must have expandable memory, which precludes having it soldered to the CPU package. Although they could theoretically have a small bit of memory on-package, and then the rest on DIMMs or whatever. I'm not sure if that would make sense though.

In any case, it never seemed likely that they would maintain the integrated module approach all the way up through to the top end of the product line; it just doesn't scale up infinitely. It's fantastic for the low end, and *maybe* for the middle, although I have doubts if there's going to be a discrete GPU and much higher memory ceilings.

We really have almost no idea what Apple's going to do, much like we knew very little about the M1 itself in advance. That's what makes this fun.

 

[Fomalhaut]

Yes, that means it's M1X, not M2. That confirms it. Thank you.

Not necessarily. Apple (& other chip fabs) have been happy to use the same process node sizes for multiple generations. Look at Intel - 14nm since...for ever

A "refined 5nm process" could just as easily refer to the new A15/M2 micro-architecture.

 

[Fomalhaut]

Who knows what they'll call it, but for sake of our discussion I think it's fair to say the M1X would use the same CPU cores as M1, but simply have more of them (GPU approach is still TBD), while M2 would be something with updated CPU cores also to be used on the A15.

I would also expect an M1X to be built on the same 5nm process as the M1, but that is not guaranteed.

At this point it certainly seems possible they're going to skip the M1X and go straight to the M2... but if the M2 is a higher end chip designed for the larger MBP and the larger iMac, will there also be a "lite" version for the Air/13"/small iMac, or will they skip a year, or what. Until we go through a complete cycle with AS chip releases for all the Macs, we can't really assume how Apple is going to manage this whole thing. My current best guess is that Macs are handled like iPads and/or Apple Watches, where each model does not necessarily get a new CPU every year. The iPhone will be the only product that absolutely gets a new chip each year.

But we'll see. Exciting times.

That's a great summary, and I think it's likely to correct.

At this stage it's hard to know if the next release will be a beefed-up M1 or an M2 with core improvements. You would think that the next generation of Apple Silicon will have common design with the A15, which is why a WWDC announcement and July release in uncertain (if we assume a new iPhone with A15 would be launched in Sep/Oct). Maybe they are already ready with the M2 to go with the smaller production runs for the Macs?

Or maybe the reported M2 is actually an M1-Pro. Same M1 cores, but more of them, and hopefully improvements to I/O (more TB3 ports, fix USB speed issues, multiple external monitors), and more RAM options.

It's possible that the M1 is just a single-variant initial offering, and that the different size options will only start with M2.

Whenever "M2" arrives, I think it highly likely that it will have 2 or 3 tiers of performance, so there will an M2-entry level similar to the M1, a mid-level for MBP14/ medium iMac, and a high-performance level for MBP16/larger iMac.

As you say, we'll see!

 

[Bug-Creator]

Furthermore, anything marketed as "Mac Pro" absolutely must have expandable memory,

Anything that is marketed as a "MacPro" is marketed as a "MacPro" (case in point the TrashPro)

On a more serious note, SSDs are going faster and faster to the point that they come in modules that can be used in a RAM-socket (I joke you not) so IF you put somewhere between 64 to 256GB of RAM onboard (on package) and have several fast SSD running in parallel plus good SW support you will be down to noone caring except for the usual whiny kids (most of them wouldn't be able to buy anything "Pro") running highly artificial benchmarks.

Note: Not saying that this is what we will see next year, but thats where the tech is heading.

 

[Mimiron]

Buy the computer that meets your needs for the expected time of ownership.

If what's on the market today doesn't fully meet your (realistic) needs, then don't buy it.

The maximum of 16 GB non-upgradable RAM worries me the most.

 

[deeddawg]

The maximum of 16 GB non-upgradable RAM worries me the most.

AFAIK there’s nothing saying that limit will exist with the upper tier systems when they transition to Apple Silicon.

and the two port Intel MBA and MBP models had the same 16GB limit suggesting it was a design choice rather than a true limitation

 

[quarkysg]

AFAIK there’s nothing saying that limit will exist with the upper tier systems when they transition to Apple Silicon.

and the two port Intel MBA and MBP models had the same 16GB limit suggesting it was a design choice rather than a true limitation

Yes , I think it’s a design decision to limit the memory capacity to 16GB for the M1 Macs that’s announced/released as the model those are replacing have the same memory limit.

As for the M1, Apple could have put a lot of 8GB pair of the LPDDR4X and pass the data I/O lines via multiplexers and call it a day. Performance would sucks, but hey we will get gobs of memory 😬.

 

[neilw]

AFAIK there’s nothing saying absolutely no chance that limit will exist with the upper tier systems when they transition to Apple Silicon.

Fixed.

16 GB limit is absolutely, positively, 100% a design choice for the set of machines that use the M1.

 

[ericwn]

Fixed.

16 GB limit is absolutely, positively, 100% a design choice for the set of machines that use the M1.

Yes very likely that they will offer higher RAM ceilings with their more performance oriented chips further down the line and probably at spicy upgrade costs.

 

[deeddawg]

Fixed.

16 GB limit is absolutely, positively, 100% a design choice for the set of machines that use the M1.

While I fully agree with you, I avoid stating absolutes when it comes to Apple and their future plans.

 

[Fomalhaut]

AFAIK there’s nothing saying that limit will exist with the upper tier systems when they transition to Apple Silicon.

and the two port Intel MBA and MBP models had the same 16GB limit suggesting it was a design choice rather than a true limitation

^This. My assessment of the reason for the 16GB RAM limit on the M1 is that it is a combination of product positioning (M1 = entry-level -> 16GB should be enough this level), and the physical space on the SoC.

The M1 core architecture is 64-bit and should have no inherent limit to addressing more memory, given a sufficiently wide memory bus.

There may be 16GB chips the same size as the 8GB ones on the M1 package, which would have allowed up to 32GB, but I'd bet that they are more expensive/GB than the smaller chips, and Apple made the decision to go with 8 & 16GB options. It's enough for a the majority of Mac users.

What some people seem to be missing is that the next Apple Silicon SoCs are likely to be physically larger (to fit 2 or 3 times the number of performance cores, and maybe twice the number of GPU cores. There should be space for at least 32GB or RAM.

The on-package RAM approach isn't going to scale to support the 1.5TB that the Mac Pro can fit or even the current 128GB maximum on the iMac 27. It will be interesting to see how Apple addresses this. Maybe some kind of tiered memory model using the on-package RAM as an L3 cache.

 

[cmaier]

^This. My assessment of the reason for the 16GB RAM limit on the M1 is that it is a combination of product positioning (M1 = entry-level -> 16GB should be enough this level), and the physical space on the SoC.

The M1 core architecture is 64-bit and should have no inherent limit to addressing more memory, given a sufficiently wide memory bus.

There may be 16GB chips the same size as the 8GB ones on the M1 package, which would have allowed up to 32GB, but I'd bet that they are more expensive/GB than the smaller chips, and Apple made the decision to go with 8 & 16GB options. It's enough for a the majority of Mac users.

What some people seem to be missing is that the next Apple Silicon SoCs are likely to be physically larger (to fit 2 or 3 times the number of performance cores, and maybe twice the number of GPU cores. There should be space for at least 32GB or RAM.

The on-package RAM approach isn't going to scale to support the 1.5TB that the Mac Pro can fit or even the current 128GB maximum on the iMac 27. It will be interesting to see how Apple addresses this. Maybe some kind of tiered memory model using the on-package RAM as an L3 cache.

They could fit 128GB in a reasonable package size, as it turns out, but I agree with your sentiments. That said, I would not be surprised if the upcoming ”M2” machines announced at WWDC use a package with 8+4 cores, a separate GPU chip, and a choice of either 32GB or 64GB (all in the package). I think the “external memory” thing is further down the road, reserved for Mac Pro, and they will eventually offer up to 128GB in-package. That said, we’re all guessing.

 

[pdoherty]

Seems odd that M2 is coming already was expecting m1X as most

Considering the M1 is all cores functioning what would an M1X even be? Just a slightly-higher clockspeed?

 

[cmaier]

Considering the M1 is all cores functioning what would an M1X even be? Just a slightly-higher clockspeed?

People use “m1x” to refer to a variant with more cores, and potentially bigger cache, more I/O bandwidth, and more DRAM.

 

[Fomalhaut]

Seems odd that M2 is coming already was expecting m1X as most

The M2 would be a second generation version of Apple Silicon for Mac, whereas the "M1X" would be a larger variant of the existing M1, with more CPU & GPU cores. I would personally have expected a second generation M2 around the same time as the new iPhones which will use the same updated core technology (in the A15), but it's possible that the rumor is true and that it is already ready to start fabrication for the lower volumes of Mac SoCs

There may not be an "M1X" at all, and the M1 may just be a single variant generation, leading directly to an M2 that has both a new core micro-architecture (with some speed improvement) and support for more CPU performance cores and larger GPU.

It remains to be seen how Apple will name the chip generations and variants. They may not use letter prefixes at all (e.g. M1X, M1Z) to indicate size/power variants. It might just be 12-core M2, 16-core-M2 etc.

 

[cmaier]

The M2 would be a second generation version of Apple Silicon for Mac, whereas the "M1X" would be a larger variant of the existing M1, with more CPU & GPU cores. I would personally have expected a second generation M2 around the same time as the new iPhones which will use the same updated core technology (in the A15), but it's possible that the rumor is true and that it is already ready to start fabrication for the lower volumes of Mac SoCs

There may not be an "M1X" at all, and the M1 may just be a single variant generation, leading directly to an M2 that has both a new core micro-architecture (with some speed improvement) and support for more CPU performance cores and larger GPU.

It remains to be seen how Apple will name the chip generations and variants. They may not use letter prefixes at all (e.g. M1X, M1Z) to indicate size/power variants. It might just be 12-core M2, 16-core-M2 etc.

M1 (or variations) may lurk around for a few years in future Apple TV’s, non-pro-ipads, etc., even if M2 goes into machines announced at WWDC. Kind of like A12.

 

[Mick84]

Really wish these chips would run bootcamp/windows but they don't

 

[cmaier]

Really wish these chips would run bootcamp/windows but they don't

Weird. I wish intel chips could natively run Arm software. Oh, well.