Starting to look like M2 might be based A16?

[senttoschool]

• M2 Pro/Max is reported to have 12 CPU (8/4) cores instead of 10 (8/2), which hints at a newer architecture than A15. A15 did not have an increase in CPU cores.
• Usage of 4nm, which is what the A16 is expected to use. This means both M2 and A16 will be using the exact same node, which probably helps in development and testing.
• 2-year update cadence to M series seems likely at this point. Basically, M series will always skip one generation of the A series. M3 would be based on A18. Makes a ton of sense because these chips are much more complicated to design and manufacture than the A series. Pretty hard to design a new M Ultra every year.
• M1 and A14 were released within a month of each other. This suggests that Apple develops the base M and the A in tandem. A15 has already come and gone. It could be that Apple is developing the A16 and M2 in tandem instead of the A15.
• All rumors point to a second-half 2022 launch for the new M2 Macbook Air which aligns with the 2-year cadence.
• The A series could be adopting a cycle of a big performance increase generation followed by an efficiency generation. A14 was a big perf increase. A15 was mostly an efficiency increase + higher GPU core count. M series might always be based on the big perf increase generation.

Also note that the M2 GPU cores are rumored to increase by 2, following A15. This likely means the A16 will not have an increase in GPU cores. However, we could see the A16 add more efficiency cores because of the rumored additional 2 efficiency cores of the M2 Pro/Max.

If this is true, then the M2 will be a very good upgrade over the M1. It will combine the efficiency improvements and higher GPU cores from A15, and maybe the perf and higher CPU cores from A16. Plus, it will be using a slightly improved node.

However, I'll probably skip and upgrade my M1 Pro when the M3 hits, because it'll likely use 3nm, which is a proper node shrink from 5nm.

 

[UBS28]

M2 will be based on A15. Apple is not going to shift iPhone production to Mac, as Apple makes very little money on Mac's in comparison to iPhone's.

 

[thenewperson]

M2 Pro/Max is reported to have 12 CPU (8/4) cores instead of 10 (8/2), which hints at a newer architecture than A15. A15 did not have an increase in CPU cores.

Eh, this is weak. Seems more like a design decision to go with 2 more efficiency cores in the larger chips to match the lower ones than proof that there’ll be an increase in E cores in the A16.

• Usage of 4nm, which is what the A16 is expected to use. This means both M2 and A16 will be using the exact same node, which probably helps in development and testing.

True, but it wouldn’t be the first time the larger version of a certain A chip uses a newer process (A10 on 16nm vs A10X on 10nm)

 

[deconstruct60]

Eh, this is weak. Seems more like a design decision to go with 2 more efficiency cores in the larger chips to match the lower ones than proof that there’ll be an increase in E cores in the A16.

There isn't much to indicate there is a E core count increase in A16 here. The A15 didn't add CPU cores in part because there wasn't a bug transitory budget increase. The A15 was just a larger die than the A14. The ProRes video and additional GPU grew the die larger. If Apple had added more P or even E cores the die would have just grown even larger still.


It is likely more "low transistor budget increase" than it is weak. Going from 5nm to 4nm doesn't really 'buy' much in increased transistor density. So adding two P cores would make the overall die bigger ( which costs more in die consumption and defects ). Going to a 'full' 4 E core complex would cost a between a 1/2 to 1/4 of that space increase. ( already a 2 E core complex so have the shared L2 already. Not really anything new since not chopping down the standard 4E complex). Also not as much of a memory bandwidth pressure increase.

Pretty good chance that Apple has spent most of the 5->4 small density increase on something else. Very similar to the A14->A15 transition but with a slightly bigger transistor budget growth. ( GPU upgrades like caching. NPU upgrades , AV1 decoder (both Intel and AMD will have them by end of the year. ). Pretty likely that the A16 GPU and A/V de/encoding has had some upgrades.

Going to N4/N4P would get Apple some frequency headroom to push through some clock frequency tweas all around, so the P (and E) could leverage that. Probably some incrementaly internal architecture they could do inside the P and E cores also. Make 'better' cores not more of them. That isn't a 'weakness' move.


The Pro and Max are already relatively big chips. Unless there is some big "win" in making them bigger dies , there is more upside for Apple to wait for M3 (and 'full node' shrink to some TSMC N3 variant ) for any P core count increases (or just 'bigger P' cores complexes and the count stays the same. More registers , L1 , and L2 cache. bigger AMX ... etc. ).

True, but it wouldn’t be the first time the larger version of a certain A chip uses a newer process (A10 on 16nm vs A10X on 10nm)

Unlikely that A16 is still around on N5/N5P ... Apple has already done a coule of iterations of A series on N5. N4/N4P is a N5 dervative but it gets some incremental density improvements ( ~5-8%). That is enough to push the A15's ~100mm^2 die back done to the 'normal' plain A series range of 85-95mm^2. There isn't a big upside for Apple to keep making a > 100mm^2 for longer than necessary.

[ the plain iPhone 14 is suppose to keep the A15 but they could easily just switch to the full GPU ones as the "upgrade". It is a bigger die but it also mostly paid for development at this point also. ]

 

[deconstruct60]

M2 will be based on A15. Apple is not going to shift iPhone production to Mac, as Apple makes very little money on Mac's in comparison to iPhone's.

1. Rumors are that the iPhone 14 non "Pro" versions are still on the A15. If that is the case there is lower demand for TSMC N4 A16 as they are only going into the Pro versions. ( The Pro versions are significant fraction of leading edge iPhones sales, but not most of the sales. )

2. Apple doesn't have to do all of the M2's all at once. IF they spread the TSMC N4 M2 SoC roll out over a wider number of months then the number of M2's needed would go way down also.

e.g.


scenario 1 Mac Mini M2 WWDC ( note apple registered three systems this spring and only did the studio). Lead with not the most high volume plain M1 model upgrade.

M2 MBP 13" October
M2 MBA Novemeber

scenario 2 Mac Mini "Pro" M1 Pro WWDC (drop another Intel model off the line up)
MBA M2 October
MBP 13" Janurary.


If they start the run of the M2 production before the A16 starts ( as a short pipe cleaner) and keep a fraction of the N4 production on M2 while the bulk when the phone then would not be much of a problem. Even more so if tamp done the iPhone 14 consumption by doing a ( binning "hand me done" to the plain iPhone version with just A15 (on N5/N5P). )

 

[deconstruct60]

• https://www.macrumors.com/2022/04/12/unreleased-mac-mini-studio-display-firmware/All rumors point to a second-half 2022 launch for the new M2 Macbook Air which aligns with the 2-year cadence.

Not all rumors.


"... I’m told there are two new Macs coming around the middle of the year or early in the second half. One of those is likely to be the new MacBook Air. ..."

Gurman: Apple could announce two new Macs at WWDC 2022 - 9to5Mac

After announcing the new Mac Studio and Studio Display last month, Apple still has a few more Macs to introduce in 2022.

9to5mac.com

Also a M2 Mini doesn't "have to" wait until Fall either.

Unreleased Mac Mini Spotted in Latest Studio Display Firmware

An unreleased Mac mini has been spotted within code of the latest Studio Display firmware, adding credence to rumors that Apple is working on an...

www.macrumors.com

If Apple can't finish off the Mac Pro transition until December 2022 or into 2023 due to some delays then trotting out the M2 at WWDC would be a good "diversion" from that. They wouldn't have crossed of the Mac Pro but would have gotten to version 2 of the M-series after two years. The Mac development kit was a Mini. M2 Mini would be two generations past that fledgling transition kit. .... isn't that lots of progress? Apple will say they are making decent progress.

• The A series could be adopting a cycle of a big performance increase generation followed by an efficiency generation. A14 was a big perf increase. A15 was mostly an efficiency increase + higher GPU core count. M series might always be based on the big perf increase generation.

Also note that the M2 GPU cores are rumored to increase by 2, following A15. This likely means the A16 will not have an increase in GPU cores. However, we could see the A16 add more efficiency cores because of the rumored additional 2 efficiency cores of the M2 Pro/Max.

The A16 doesn't have to increase GPU count because there was die size bloat from A14 -> A15. Going from N5/5P -> N4 would allow Apple to take the die bloat back out and get back to the normal plain A-series die size. Same thing for M2. The 5-8% density increase of N4 would allow M2 to pic up two GPU cores (and likely at least the 'small' ProRes de/encoder from the A series. ) and not drift out of the A10Z-A12Z/M1 'normal' die size range.

Same with the M1 Pro / Max ... if get small density boost.. Take smallish core count bumps and call it day until a full node shrink rolls around ( some TSMC N3 varaint ). Two E cores is pretty small. And two GPU cores don't necessarily need new P core complex (L2 and other "overhead").

If this is true, then the M2 will be a very good upgrade over the M1. It will combine the efficiency improvements a nd higher GPU cores from A15, and maybe the perf and higher CPU cores from A16. Plus, it will be using a slightly improved node.

Not sure that the M1's GPUs ( or at least the M1 Pro and Max GPUs) were that far behind the A15 GPUs. A very limited desnity gain would allow Apple to puts some speed bumps in there. But suspect selling this as some "huge gain" upgrade is probably overselling it. It doubt this would drive most M1 users to toss what the have and upgreade because there is amazing new performance. M2 is probably more grounded in getting the "don't buy version 1" buyers, than the early adopters.

For the Pro and Max the GPU core count is already relatively high ( versus A series or M1 ). Two more cores per GPU cluster really isn't likely a huge performance gap. Something, but not huge.
Decent chance Apple will leverage that increase for just more BTO binning to make the pricing upgrade ladder taller ( and more profits. )

 

[senttoschool]

M2 will be based on A15. Apple is not going to shift iPhone production to Mac, as Apple makes very little money on Mac's in comparison to iPhone's.

So what do you think M3 will be based on? A16?

Apple will have to skip some generations of A series so that the M can catch up yo the latest core. I doubt that they’re doing an M series for every A series.

 

[BootLoxes]

I hope M3 is based on A18

 

[MrGunny94]

We don't really know do we? Let's wait and see, most likely could be A15 or A16. The thing is we are not 100% clear on how the Apple Silicon roadmap is like yet because we are still seeing chips release like the M1 Ultra which are obviously based on the latter.

 

[deconstruct60]

So what do you think M3 will be based on? A16?

Apple will have to skip some generations of A series so that the M can catch up yo the latest core. I doubt that they’re doing an M series for every A series.

There is a pretty decent chance that some of the larger die members of the M series also skip generations of the 'plain' M series.

For the iPad Pro A--X models, Apple got into a pattern of skipping intervals were didn't get a substantive process shrink. So it was more a Moore's Law frequency uptick ( e.g., around 18-24 months; not 12 ).

The 'plain' , entry M series SoC is going into iPads . Neither Macs nor iPads are pragmatically 'hardlocked' to every 12 month cycles like the iPhone is now.

And Apple could make a M2 ( maybe a M2 Pro) without making a M2 Ultra or M2 Double Ultra.

The unit volumes of SoCs produced per year is going to be substantively different across the M-series scale. Updating them all at the same rate doesn't much much economic sense. The iPhones generate tons of unit volume and revenue so can 'throw money' at them each year with deeply overlapping development pipelines. Once up in the highest end Macs now down to two (maybe three) orders of magnitude lower run rates. (yes they are far more expensive end users and have large profit margins... but obsoleting them on a 12 months basis would be wasteful. )


There is a good chance that it is a bit flawed to label the M3 based on A16 or A16 based on M3. There is probably a common foundation design ( code name whatever, 'foobar' ) that is the baseline design for years where there is an shared overlap. Whether the M-series or A-series version comes out first would hinge on when the TSMC fab process it is based on go into high volume mode . High volume mode comes early in the Spring ( Feb-April) or early Winter ( Nov-Jan ) then Apple can run off M series before going into peak mode for the new A series. If TSMC high volume mode matches the pre Fall ramp for the iPhone then the A-series could go first. When TSMC's roll out windows vary by 18 months or so then it won't match the rigid 12 month cycle of the A-series. Dogma of A series has to come before M-series or vice versa makes no sense when the underlying manufacturing factors are not on exact 12 month cycles.

 

[UBS28]

So what do you think M3 will be based on? A16?

Apple will have to skip some generations of A series so that the M can catch up yo the latest core. I doubt that they’re doing an M series for every A series.

The M1 is basically an iPad chip, so Apple will do the same approach they have been doing for years as with the iPad.

The iPhone will have the latest and greatest and the iPad and Mac will use old generation tech.

 

[deconstruct60]

The M1 is basically an iPad chip, so Apple will do the same approach they have been doing for years as with the iPad.

The iPhone will have the latest and greatest and the iPmad and Mac will use old generation tech.

Technically no. First the M1 is basically an iPad Pro chip. Not the low-mid range iPad SoC.


A10 - 16nm Sept 2016

A10X - 10nm June 2017

[ https://en.wikipedia.org/wiki/Apple_A10

https://en.wikipedia.org/wiki/Apple_A10X
]
iPad Pro chip uses a more advanced fab tech process than the A10. Not a "hand me down" at all.

A11 -- 10nm Sept 2017 ( So iPad Pro chip came out before the A11 )
[ no A11x . as essentially nothing smaller to jump to. ]

A12 -- 7 nm Sept 2018
A12X -- 7nm October 2018

Basically a 'tie' in terms of release. A process node shrink so new iPad Pro A--X version done.

A13 -- 7nm Sept 2019
[ no A13X as essentially nothing smaller to jump to. A12Z is a rebadged A12X. Extra GPU core was always there. It was just binned off in X release. ]

A14 -- 5nm Sept 2020
M1 -- 5nm Nov 2020

Again basically a 'tie' in terms of release. Again a full process node shrink to jump to so taken.


Apple has iterated the plain A-series more often but the iPad Pro SoC has not been a trailing edge SoC at all. The low-Mid range iPads... yes those are mainly "hand me down" SoC from the iPhone. However, that has not been try with the serious iPad Pro sequence at all.

Apple is now tossing the plain M-series into the mid range iPad Air. There is a pretty decent change that this will take on the "hand me down" capture statue , but switching from inheriting from iPhone to inheriting from the iPad Pro. ( in 2022 iPad Air gets M1 and later in 2022 iPad Pro goes to M2 . The Air will probably continue on M1 until the iPad Pro is close to transitioning onto the M3. Rise and repeat of selling "already paid for R&D" M-series in cheaper system. [ Not sure if it fits but wouldn't be surprising for Apple to stuff an 'second half of lifecycle' M1 into a iPad Mini also. The entry level iPad stays on iPhone A-series hand me downs. The A15 gave the iPad Mini decent USB-C port support. This will rotate down to the lowest end iPad over time. ]


P.S. The M2 jumping in on a TSMC N4 process upgrade wouldn't be a full node jump. but it would allow for incremental upgrades without any substantive die size bloat.

The main difference has been is that Apple has been willing to 'bloat' the A-series die a bit on "in between die shrink" iterations. The larger die has skipped those. That doesn't mean it is "behind' the A series. It is just 'passing' on some A series generations as not worthy enough.


Once full flushed out the M-series line up of dies will be much broader than the A-series. Probably not a full set released on each generation. Same reasons why bigger A--X/M1 skipping generations is at play with the biggest M-series die relative to the smallest one in the series. So there is a decent chance some M-series updates will have limited subsets feature tweaks that are shared across the 'largely skipped' A-series 'filler' generations between major fab process updates.

And once the cellular radio die ( or subset subsystem ) are weaved into the A-series dies will likely see more sub feature mixing and matching that only overlap with part of the M-series line up.

 

[EPO75]

The M1 is basically an iPad chip, so Apple will do the same approach they have been doing for years as with the iPad.

The iPhone will have the latest and greatest and the iPad and Mac will use old generation tech.

Because you know how? You write things like it is a 100% certainty... Things did in the past doesn't mean they will do in the future. Your iPad approach doesn't make any sense at all to, let just see what happens. Biggest issue now is the issue with the chips & China etc and also TSMC. So this could mean newer tech will be delayed.

 

[senttoschool]

@deconstruct60 So what's your guess? M2 based on A15 or A16?

 

[jdb8167]

@deconstruct60 So what's your guess? M2 based on A15 or A16?

I think it depends on if the A16 is going to be 3 nm. Right now the expectation is that 3 nm isn't ready until 2023 but I keep reading that TSMC is ready to go with 3 nm in the second half of 2022. If that date is for is non-volume risk production then I think it is possible that the M2 is based on a 4-5 nm A16. But if the A16 is going to be 3 nm, then it is unlikely that the M2 generation will be based on the A16.

 

[leman]

Plot twist: what it A16 is based on M2 instead?

 

[altaic]

Plot twist: what it A16 is based on M2 instead?

I’ve heard that the M2 is based on the M1. If I had to guess, I’d say the A16 is based on the A15. ?

 

[deconstruct60]

@deconstruct60 So what's your guess? M2 based on A15 or A16?

I think it is based on a TSMC N4 variant. There is no good reused to do another M-series iteration/generation on N5. (the A15 was a bloated bigger die compared to the A14. Apple probably doesn't want to pay more for the already larger M series dies. ). If Apple can make some small additions to the M2 and keep the die size the same then I think they would "take the money" ( instead of raisin costs).

It isn't based on the A16 . They probably have a shared design. My guess that the M2 was (back in 2019 plans) was suppose to come out around the same time (+ 1-3 months). as the Ultra (Studio). And that the timeline is messed up a bit.

Apple ramping on N4-Apple earlier in Spring would allow for a limited M2 launch first and A16 ramp over the Summer months. The order of release doesn't mean one is based on the other. It is when the designs get finished, not when they crank the production. Those are not necessarily the same time. ( validations on bigger chips typically take a bit longer. But the M2 isn't very large relative to the normal A series die ( 120-140mm^ versus. 80-95mm^2) )


[ I also suspect that the A15 got some minor "cross pollination" with M1 Max/Pro with the scaled down ProRes engine. And since that 'small' sharing was done, there is little reason to do a larger block. ]

 

[altaic]

I think it is based on a TSMC N4 variant. There is no good reused to do another M-series iteration/generation on N5. (the A15 was a bloated bigger die compared to the A14. Apple probably doesn't want to pay more for the already larger M series dies. ). If Apple can make some small additions to the M2 and keep the die size the same then I think they would "take the money" ( instead of raisin costs).

It isn't based on the A16 . They probably have a shared design. My guess that the M2 was (back in 2019 plans) was suppose to come out around the same time (+ 1-3 months). as the Ultra (Studio). And that the timeline is messed up a bit.

Apple ramping on N4-Apple earlier in Spring would allow for a limited M2 launch first and A16 ramp over the Summer months. The order of release doesn't mean one is based on the other. It is when the designs get finished, not when they crank the production. Those are not necessarily the same time. ( validations on bigger chips typically take a bit longer. But the M2 isn't very large relative to the normal A series die ( 120-140mm^ versus. 80-95mm^2) )


[ I also suspect that the A15 got some minor "cross pollination" with M1 Max/Pro with the scaled down ProRes engine. And since that 'small' sharing was done, there is little reason to do a larger block. ]

I’m (tangentially) curious, do you know if TSMC adds fab capacity with a new node, or do they upgrade existing equipment? Or is there a non-node-related production bottleneck?

 

[deconstruct60]

I think it depends on if the A16 is going to be 3 nm.

No. There is a TSMC N4 process node that Apple can use that is on a shorter availability timeline than N3 is.
TSMC N3 is a little late. ( not huge slide bit big enough would have been obvious years ago that it wouldn't finish on time for Spring '22 ).

Apple's Fall line up release pragmatically needs 1H of year 20xx mass volume status to lower risk to reasonable levels.

There is a chart here that shows that N4 gets some incremental performance and density uptick and is ready in 2022.

https://www.anandtech.com/show/17123/tsmc-unveils-n4x-node-high-voltages-for-high-clocks

[ Some folks have said Apple will hold out for N4P which has same "2H22 issues" as N3. Probably not. That is same 'timestamp' word, but likely not the same actual time of arrival. It is a tweak to 4. that is lower production complexity increase than doing N3 which has substantive differences. If Apple got 'early access' N4P then could be an option although would be tough to hit WWDC with something that ships quickly.

Or there is a N4-Apple that a special sauce variant that designed to have a quicker ramp than the probably more long term N4P ]

Right now the expectation is that 3 nm isn't ready until 2023 but I keep reading that TSMC is ready to go with 3 nm in the second half of 2022.

2H of 2022 is technically December. Which is too late for both the A16 and M2. iPhone die production has to start in April-June to finish the making/test/package/ship-to-Foxconn long logistics train by September. 2H22 means it always was a non starter for the 2022 iPhone. It is not a 1H start.

TSMC N3 is in "at risk production" now. That means folks who are not super sensitive to yields can make stuff but fine tuning isn't done.

Back in 2021 TSMC revealed on a conference call that they would not be recognizing any revenue from N3 until 2023. That means finished chips that they can get money for primarily aren't arriving until 2023. The implication there is that this is deep into Q4 2022 that high volume is starting. Wafer starts are not finished wafers. This isn't a microwave popcorn process ( a couple minutes ... ding it is done) . It takes a while for finished usable chips to come out the other side.

For a Feb-April product N3 is fine if have already purchased the wafer start slots at end of Q4 2022.

TSMC likely will start in Q4 so can claim they technically didn't miss the schedule prediction. However, the Q3 subset of that "2H22" prediction appears to be a total bust. So M2 and A16 would be a miss. (even if want to do M2 in Q4 if can't even start until Q4 then it would be a 'miss').

If that date is for is non-volume risk production then I think it is possible that the M2 is based on a 4-5 nm A16. But if the A16 is going to be 3 nm, then it is unlikely that the M2 generation will be based on the A16.

Some 4 variant is just way less risky. N3 is in at risk already. ( Apple could be making 'hundreds' of M3 chips now to throw into a large , concurrent testing trails. Possibly put a 'look but don't touch or use' prototype on stage in June if they wanted to. )

 

[deconstruct60]

I’m (tangentially) curious, do you know if TSMC adds fab capacity with a new node, or do they upgrade existing equipment? Or is there a non-node-related production bottleneck?

It is a mix. Wafers are generally fabricated in layers. So let us just use some even numbers for examples and say the mainstream TSMC N5 takes 16 layers. 8 layers could be done with EUV and 8 layers with DUV. ( those wafers will run though different machines ). N4 could be switching the mix to 10 layers and 6 layers with DUV. Still have 16 but spending more time in EUV fab tool(s) than in DUV tool(s).

So yes, they would generally need more EUV tools if wafers spend more time in them. Even more so if volume of the "older stuff" also using EUV steps isn't going down at all.

For same size wafer and same size scribe line between the dies then could use a common back end 'die cutter'. So some parts of the tool chain could remain the same. There are also additional tools besides the ASML "printers" in the process which may shift/change from node to node.

Since Covid demand for old and new stuff has gone up, so there is relatively little "now going idle" stuff in fabs anymore where can do load shifting between lines. If there were more brief lulls , then could do some shifting to fill empty idle capacity if the fab has flexible enough wafer routing and delivery lay out.

As Apple rotates out of N5 , then AMD/Intel/Qualcomm/etc roll in and use even more in aggregate. Nvidia is rolling into a N4 variant by end of year. Intel will likely be small scale ramping on N3 also in 6-10 months also. (even if Apple skips N3 until April-June it would be consumed by someone. )

 

[altaic]

It is a mix. Wafers are generally fabricated in layers. So let us just use some even numbers for examples and say the mainstream TSMC N5 takes 16 layers. 8 layers could be done with EUV and 8 layers with DUV. ( those wafers will run though different machines ). N4 could be switching the mix to 10 layers and 6 layers with DUV. Still have 16 but spending more time in EUV fab tool(s) than in DUV tool(s).

So yes, they would generally need more EUV tools if wafers spend more time in them. Even more so if volume of the "older stuff" also using EUV steps isn't going down at all.

For same size wafer and same size scribe line between the dies then could use a common back end 'die cutter'. So some parts of the tool chain could remain the same. There are also additional tools besides the ASML "printers" in the process which may shift/change from node to node.

Since Covid demand for old and new stuff has gone up, so there is relatively little "now going idle" stuff in fabs anymore where can do load shifting between lines. If there were more brief lulls , then could do some shifting to fill empty idle capacity if the fab has flexible enough wafer routing and delivery lay out.

As Apple rotates out of N5 , then AMD/Intel/Qualcomm/etc roll in and use even more in aggregate. Nvidia is rolling into a N4 variant by end of year. Intel will likely be small scale ramping on N3 also in 6-10 months also. (even if Apple skips N3 until April-June it would be consumed by someone. )

That’s really interesting, thanks!