A16 bionic supposedly on 5nm process. What does this mean for M3?

[Technerd108]

I just read at a few sites that the a16 bionic was based on 5nm process as the third chip to be made on the same process node, technically second as M1 wasn't "enhanced" but still all 5nm chips.

Since usually the M series chips are based on the a series chips does this mean that M3 will still be on 5nm? M2 was based on a15 if I am correct. And if M3 will be based on a16 then it will be a 5nm chip too. So we may have to wait for M4 chip to be on a smaller process node some time in 2024?

Does that sound right? Please correct my information if I am wrong because I am trying to learn and be accurate. Please don't spout what you think but something based on facts and hopefully with links.

Thank you

 

[Citizen45]

I think we won’t be seeing the M3 until the end of 2023.

By then, the A17 will be out and I believe it will be based on the 3nm process.

So I believe that the M3 will be based on the A17, not the A16.

 

[senttoschool]

It means one of these possibilities:

1. Apple has redesigned the A15 to use 3nm. Apple will release M2 Pro/Max as 3nm A15 SoCs in early 2023.
2. Apple has designed A16 to be 5nm. In conjunction, Apple's team has taken the A16 design and shrunk it to use 3nm for the M3 in early/mid 2023.
3. M3 will use A16 and be based on the same 5nm.
4. M3 will use A17 and be based on 3nm and will be released in fall/winter 2023.

I believe #2 is most likely.

 

[deconstruct60]

Since usually the M series chips are based on the a series chips does this mean that M3 will still be on 5nm? M2 was based on a15 if I am correct. And if M3 will be based on a16 then it will be a 5nm chip too. So we may have to wait for M4 chip to be on a smaller process node some time in 2024?

There are two presumptions here that don't have to hold at all.


1. Apple "has to" roll out new cores from smallest SoC to bigger SoC.
There is an equally viable path where they roll it it from bigger SoC to smaller one.

For example

M2 Extreme (Quad) on N3 Jan-Feb '23 ( smaller M2 arch that is easier to package)
M2 Ultra on N3 March '23
M2 Max on N3 April '23
[ Skip putting the M2 Pro on N3 and keep it on N5p high volume , low risk ramp ]
A17 on N3 Sept '23 (pilfer the shrunk P and E cores from above , slap some new uncore element ( better ProRes , camera ISP , etc. and call it A17. That is a bit more 'honest' than yet another serving of the same thing. )

Apple would only be putting the Mac Pro at risk eariy in the N3 cycle and those system pricing will be average well over $6K so the SoC have a high market up attached.



Using a high profit margin , low volume die works about as well to 'pipe clean' as new fab process as ' low margin , very high volume ' die. Either use a chunk of the large margins to pay for the higher than desired defect rate. Or 'duck'/avoid the defects by being a smaller chip and lots more dies coming out of a wafer so can afford to amortize in smaller chunks those defective dies.


50 dies per wafer that generate $1000 profit per die generates as much money as
525 dies per wafer that generate $96 profit per die.


Also rolling out the Mac Pro on N5P up against N5P AMD and N4 Nvidia GPUs is just asking for trouble. The iPhone SoC have a larger lead over the competition. At the high end of the worktation market ... M-series doesn't.


2. The M3 does not have to be based on the A16 at all. There is presumption there that the M3 has to come out in 2023. (e.g,, Apple has to put the MBA on some yearly cadence of updates just like the iPhone). Apple doesn't have to do that at all. In fact, they did not do it for the M1 -> M2 transition and things worked out just fine.

A17 on N3 Sept '23
M3 on N3 Feb '24 . ( about ~18 months after M2 )

The A14 and M1 showed up around the same time.

If the A16 a yet another N5P retread with just about the same microarchitecture then, that is indicative that Apple spent most of their time and money doing something else. Perhaps that is a new AR/VR SoC ( on N3 ?) or some other part of the line up to N3 sooner rather ( Spring '23 ) than later ( Fall '23). Or the cellular modem is a black hole of resources and a distraction.

TSMC N3 comes with some new Technology (FinFlex) which should have a higher learning curve on it. So taking an architecture they have worked the bugs out of and then mainly concentrate on making it small would be a good way to manage risk.

 

[deconstruct60]

It means one of these possibilities:

1. Apple has redesigned the A15 to use 3nm. Apple will release M2 Pro/Max as 3nm A15 SoCs in early 2023.

Not really. The P cores block complex and E cores block complex (and GPU basic block groups ) are not solely attach to a single SoC package. They get used into They are just blocks that an be combined with uncore elements to make a specific SoC. The A15 isn't necessarily 'first' to completion before the any other SoC design work can start. These are all SoC with multiple year lead times and basic building blocks get developed and dropped into each of them relatively concurrently.

It is akin to saying the building defines the bricks as opposed the bricks being used to make the building.

2. Apple has designed A16 to be 5nm. In conjunction, Apple's team has taken the A16 design and shrunk it to use 3nm for the M3 in early/mid 2023.

Again it is the P/E building blocks are are at issue rather than SoC containers.


Anything that Apple releases after the start of 2023 is unlikely to be on some going on two year old N5P process at that point. Once into release windows inside of 2023 N3 or at worse N4P are completely viable selections for Apple. Even 4 years ago N3 was "bad timing" for 2022 products. So it is not surprising if the A16 skips it.



If A16 is the third iteration on N5P , then it likely a pretty bloated chip. Apple would rather that be smaller. ( just in terms of better wafer allocation usage).

 

[Boil]

I just read at a few sites...

Please don't spout what you think but something based on facts and hopefully with links.

So you are basing your post on some things you read "at a few sites" (no sources or links given), but you will only accept replies "based on facts and hopefully with links"...?

You do realize anyone with actual knowledge of the Apple SoC roadmap is most likely under NDA in regards to dissemination of that knowledge...?

I THINK Apple will have the M3 series of SoCs on 3nm, sometime within 2023...

I HOPE Apple will give us the M2 Pro SoC in the Mac mini platform sometime this year...

I EXPECT Apple to integrate some form of hardware ray tracing in the M3 series of SoCs...

I am EXCITED to see the progress being made on a Metalified Blender...

 

[Sydde]

TSMC is only just starting to ramp up N3, so keeping the high-volume iPhone parts on N5P means they can get enough units for the iPhone lines and use the limited N3 line for the lower-volume Mac lines.

 

[Boil]

TSMC is only just starting to ramp up N3, so keeping the high-volume iPhone parts on N5P means they can get enough units for the iPhone lines and use the limited N3 line for the lower-volume Mac lines.

No lower volume than the SoCs (M2 Ultra / M2 Extreme) for the forthcoming ASi Mac Pro...!

 

[kvic]

I have no horse in this race. Another likely scenario:

• A16 designed on 3nm; back ported to 5nm. M3-series SoCs will remain as planned on a more advanced variant of 3nm the least
• M2 designed on 3nm; back ported to 5nm. Dropped 3nm
• M2 Pro, M2 Max designed on 3nm; will remain on 3nm. To be released around 2Q23 the earliest
• M2 Extreme designed on 3nm; remain on 3nm. To be released in 4Q22 the earliest. Same for M2 Ultra perhaps.

 

[Sydde]

No lower volume than the SoCs (M2 Ultra / M2 Extreme) for the forthcoming ASi Mac Pro...!

As it stands, Mac Studio Ultra embarrasses 2019 Mac Pro in many comparions. To overpower the Studio in GPU tasks, a Pro would have to cost at least 4 times as much. I am skeptical that Apple would seriously be breaking a sweat over the absence of an ASi Pro and would rather take their time to build and present a great masterwork amid great fanfare instead of rushing some thing out for the sake of needing a Mac Pro.

 

[Zest28]

M3 will be based on 3nm A17.

 

[leman]

A16 bionic supposedly on 5nm process. What does this mean for M3?​

Probably nothing.

Please don't spout what you think but something based on facts and hopefully with links.

Which facts? Which links? All of these things are conjecture, speculations or potentially invalid industry leaks. Apple never confirmed what their plans are. The only fact we know is that TSMC has some new N3 processes announced for 2023, roars all.

 

[diamond.g]

As it stands, Mac Studio Ultra embarrasses 2019 Mac Pro in many comparions. To overpower the Studio in GPU tasks, a Pro would have to cost at least 4 times as much. I am skeptical that Apple would seriously be breaking a sweat over the absence of an ASi Pro and would rather take their time to build and present a great masterwork amid great fanfare instead of rushing some thing out for the sake of needing a Mac Pro.

Isn't the point of the MP is to have PCIe cards to use? I am curious how they would handle the splitting of PCIe lanes across cores (I guess connecting through the interposer) and if they are going to give up Thunderbolt ports to add PCIe.

 

[gpat]

I'm hoping for a Macbook Air refresh on October 2023 at the latest, featuring 3nm M3.
We really have to change a 2013 Macbook Air by then in our household, and I'm not going to recommend the current M2 iteration.
On the other end, I can wait much longer for my 16" M1 Pro replacement. In fact, the longer the better.

 

[thenewperson]

We really have to change a 2013 Macbook Air by then in our household, and I'm not going to recommend the current M2 iteration.

Don't want to go with an M1?

 

[Retskrad]

I think whatever Apple bases the A16 on will be extremely close to the 3nm in performance and efficiency. We have to remember that 7nm, 5nm, 3nm are just marketing names. Enhanced versions of 5nm can get you extremely close to 3nm.

 

[gpat]

Don't want to go with an M1?

If I wanted an M1, I would have bought one in 2020.
I have a personal rule of not buying products in the later part of their life cycle.

 

[LinkRS]

I just read at a few sites that the a16 bionic was based on 5nm process as the third chip to be made on the same process node, technically second as M1 wasn't "enhanced" but still all 5nm chips.

Since usually the M series chips are based on the a series chips does this mean that M3 will still be on 5nm? M2 was based on a15 if I am correct. And if M3 will be based on a16 then it will be a 5nm chip too. So we may have to wait for M4 chip to be on a smaller process node some time in 2024?

Does that sound right? Please correct my information if I am wrong because I am trying to learn and be accurate. Please don't spout what you think but something based on facts and hopefully with links.

Thank you

Howdy Technerd108,

Nobody can answer your questions "based on facts and hopefully with links" as the only people with such knowledge, likely do not read this forum. This is a rumors site, where fans get together to discuss things. Plus, you do realize that the node size has nothing to do with the capabilities of CPU, and Apple can (and has) done some phenomenal work based off a node size of 5nm. The primary benefit a smaller node size will offer, is that you can fit more transistors (or cache) into the same space as a larger node size. In some cases, the smaller node is just used to make the current silicon smaller, increasing the number of "chips" per die, potentially decreasing overall fabrication costs. In many cases, the shift to a smaller node is done in tandem to a change in microarchitecture and/or design. So to some folks going to a smaller node means a more capable chip and/or SOC, however it doesn't have to be. In some cases a more mature "larger" node will be chosen as the yields on these will be higher than a "new" node, decreasing overall costs.

Now, if the current microarchitecture is so feature rich that it is producing too much heat and causing cooling problems, the switch to a smaller node would be beneficial as the distance between transistors decrease, so does current leak and overall temperature. However, the current M1/M2 SOCs don't seem to be having a problem with this, on the current node size. So, you probably shouldn't be concerned when Apple switches to 3nm, just what their SOCs can do

Good luck!

Rich S.

 

[Technerd108]

Howdy Technerd108,

Nobody can answer your questions "based on facts and hopefully with links" as the only people with such knowledge, likely do not read this forum. This is a rumors site, where fans get together to discuss things. Plus, you do realize that the node size has nothing to do with the capabilities of CPU, and Apple can (and has) done some phenomenal work based off a node size of 5nm. The primary benefit a smaller node size will offer, is that you can fit more transistors (or cache) into the same space as a larger node size. In some cases, the smaller node is just used to make the current silicon smaller, increasing the number of "chips" per die, potentially decreasing overall fabrication costs. In many cases, the shift to a smaller node is done in tandem to a change in microarchitecture and/or design. So to some folks going to a smaller node means a more capable chip and/or SOC, however it doesn't have to be. In some cases a more mature "larger" node will be chosen as the yields on these will be higher than a "new" node, decreasing overall costs.

Now, if the current microarchitecture is so feature rich that it is producing too much heat and causing cooling problems, the switch to a smaller node would be beneficial as the distance between transistors decrease, so does current leak and overall temperature. However, the current M1/M2 SOCs don't seem to be having a problem with this, on the current node size. So, you probably shouldn't be concerned when Apple switches to 3nm, just what their SOCs can do

Good luck!

Rich S.

But the only metric that matters is node size. If it isn’t 3nm it is terrible and just an overclocked M1. Lol

Reminds me of Snapdragon 8 gen 1 on a smaller node 4nm but still has heat issues and not a huge improvement on multi core.

I agree with you.

 

[deconstruct60]

I think whatever Apple bases the A16 on will be extremely close to the 3nm in performance and efficiency. We have to remember that 7nm, 5nm, 3nm are just marketing names. Enhanced versions of 5nm can get you extremely close to 3nm.

Extremely close N5P and N3 ? Not really.

"...

​

TSMC​
'​	N5 vs N7​	N5P vs N5​	N5HPC vs N5​	N4 vs N5​	N4P vs N5​	N4p vs N4​	N4X vs N5​	N4X vs N4P​	N3 vs N5
Power	-30%​	-10%​	?​	lower​	-22%​	-​	?​	?​	-25-30%​
Performance	+15%​	+5%​	+7%​	higher​	+11%​	+6%​	+15% or more​	+4% or more​	+10-15%​
Logic Area Reduction % (Density)	0.55x -45% (1.8x)​	-​	-​	0.94x -6% 1.06x​	0.94x -6% 1.06x​	-​	?​	?​	0.58x -42% (1.7x)​

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

Perhaps a bit wishy-washy on some stats when compare N4P to 'lowest improvement range' N3. But it sounds like to get highest wafer volume early in in April-June that Apple went with N5P. Even with N4P to N3 there is a huge gap in logic area density. That can indirectly bring better performance of a different architecture rather than if only primarily just make things smaller. (e.g., performance uplift due to bigger, on-die caches. )

But for the far range of N3 ( -30% power) there is a 8% gap. That isn't pragmatically in the "extremely close" zone. The key question would be whether Apple could very effectively using the "FinFlex" abilities of N3 to apply the different density/efficiencies variations on the same die to the appropriate areas. ( for example the densest Fin variation to P cores , medium variation to E cores and video decoders and mainstream 'uncore' , and largest Fin flex to the very large L3 (or L2) areas. ). N3 has a wide spread of uplift on power/performance because there is range of design can choose while still staying inside of the base "N3" wafer. The N5 family ( N5 - N4 and alphabet adjectives) don't allow for that on a single die.

If just try to use N3 as a mono-FlnFet baseline then probably will minimize the benefit if using it over N4P. That would be 'close' , but what Apple has is a die with lots of different function units on it that really could potentially use different FinFet implementations on a single die. It would be alot more complicated than anything they ever did previously , but the gains would probably be very real even if don't push the density lever to the maximum to crank up the core counts by a large amount. Take the M2 incremental core count uplift and just make a smaller , more power efficient chip on the first iteration on N3.



P.S. A16 on N4P might work in a corner case where the iPhone Pro phones pragmatically just didn't arrive anywhere close to when the plain iPhones arrived. Another reason to split the phones processors ( beside upselling to higher profit margins) was to reduct the volume by alot so a lower capacity N4P could supply stuff by November or so. N4P would be about as good as a consistently bad N3 implementation that only used the worse FinFet option uniformly across the whole die. N4P might have an edge there in that it manufactures quicker ( that 'bake' times on N3 are substantively longer due to the built in FlexFin complexity. )

Decent chance though that an A16 on N5P is more expensive to make than a A15 and so to save margins Apple is only rolling it out the the more expensive phones ( hoping to sell a higher fraction of more expensive phones with fatter margins to cover the cost increase. )

 

[theorist9]

I don't think it means anything. As I've mentioned a couple of times earlier, the Macs have much lower volume than the iPhones, so new chip processes may initially be seen on the Macs, and later appear on the iPhones as the chip production volume gears up. So while any Macs they release this year will probably also be on N5P, those released early next year could be on N3 (irrespective of whether they use the M2 or M3 microarchitecture).

 

[deconstruct60]

I don't think it means anything. A15 is on N5, but M2 (which is based on A15) is on a different process: N5P.

The A15 is on N5P.

"..

Apple A15 TMMU71 N5P SoC GPU N5P SoC Design Analysis ... "​

Apple A15 TMMU71 N5P SoC GPU N5P SoC Design Analysis | TechInsights

Apple A15 TMMU71 N5P SoC GPU N5P SoC Design Analysis

www.techinsights.com

That is why yet another N5P is rather weak after already moved up off of baseline N5 before. If it was Intel, some folks would screaming about lame 5nm++++ updates.

As I've mentioned a couple of times earlier, the Macs have much lower volume than the iPhones, so new chip processes may initially be seen on the Macs, and later appear on the iPhones as the chip production volume gears up.

The leading, premiere Phones have a problem in that they have to appear to a non technical rigid schedule of every 12 months. TSMC new fab process is not on a every 12 month schedule. Moore's Law was closer to 18 months and if slide off if that it will be longer, not quicker. But highly unlikely to settle back into an exact multiple of 12 months ( e.g., 24 month schedule.)

At times the iPhone is just going to be a very bad sync with what TSMC is rolling out. Basically inevitable because the iPhone is on a marketing date and the TSMC is more so on a technical (engineering ) date.

So while any Macs they release this year will probably also be on N5P, those released early next year could be on N3 (irrespective of whether they use the M2 or M3 microarchitecture).

There may have been Macs that Apple though they could potentially eek out at the very end of 2022 in their plans from 4+ years ago. If TSMC has actually started High volume production at the absolutely earliest date of their roadmap estimate ( July 1, 2022) then a N3 Mac in 2022 would have been viable. TSMC had a range of 2H '22 (July to December). They are going to hit about the middle of that. Which isn't bad in big picture, but very bad for a Dec '22 launch for a modest volume product that pragmatically uses large dies. . And extremely bad for a product that had anything in the over 100K units/quarter run rate.

 

[theorist9]

The A15 is on N5P.

Thanks for the correction, I'll update my post accordingly.

There may have been Macs that Apple though they could potentially eek out at the very end of 2022 in their plans from 4+ years ago. If TSMC has actually started High volume production at the absolutely earliest date of their roadmap estimate ( July 1, 2022) then a N3 Mac in 2022 would have been viable.

In TSMC's July 14, 2022 Earnings Call, they said revenue contribution (which means shipping to customers) for N3 is expected to start in the first half of 2023. Now maybe that's not so hard and fast that it doesn't preclude the possbility of shipping in low volume to Apple in late 2022 (especially if things go better than they anticipated). I just don't know.

https://investor.tsmc.com/english/e...99cda9eeecc65315449c/TSMC 2Q22 Transcript.pdf

 

[Sydde]

I think whatever Apple bases the A16 on will be extremely close to the 3nm in performance and efficiency. We have to remember that 7nm, 5nm, 3nm are just marketing names. Enhanced versions of 5nm can get you extremely close to 3nm.

While it is true that Nx is largely marketing jargon that does not actually mean x nanometer gate size, there is a very big difference between N4 and N3, in that N3 will give designers much finer control over individual gates ("FinFlex"), allowing for greater efficiency improvements.

 

[deconstruct60]

In TSMC's July 14, 2022 Earnings Call, they said revenue contribution (which means shipping to customers) for N3 is expected to start in the first half of 2023. Now maybe that's not so hard and fast that it doesn't preclude the possbility of shipping in low volume to Apple in late 2022 (especially if things go better than they anticipated). I just don't know.

Before N3 (and other TSMC fab tech) goes into "High volume production" it is in a 'at risk' production status. Customers can order up wafers from that process that would be relatively (to something like a MBP 14") very low volume. That 'at risk' money that TSMC will bring in 2022 just is not interesting enough to report to Wall St. analyst. If Apple gets 200 wafers in 2022 that they have to pay for in 2022 , it isn't going to relatively matter. At $19K per wafer that is $3.8M . For a small company that could be a lot of money that they would chat up Wall St analysts with. For Apple/TSMC that is kind of the 'petty cash fund' or even 'round off error' money.

If Apple intends to ship a product to customers in Spring '23 they had to have made 100's of SoCs for testing , prototypes , etc. They probably had to pay for those. So if going through gyrations as to whether TSMC has gotten more than zero money out of N3 in 2022 . Yeah, but that doesn't get you a Mac/iPhone product in 2022.

N3 fab process takes many weeks to bake. Then they have to build these multiple chip packages from multiple dies into a single packge. Then take that package and layer it on the package subsystem with the memory. it is not a microwave popcorn like timeline for production. That is why HVM starting in Sept doesn't' lead to money until Jan-Feb. The real core issue here is that HVM likely has not started yet. No HVM , no product.



Apple is rather shameless about doing product launches with lower stock than the initial demand curve suggests , but even for something for the Mac Pro would have to run a factory for about a month before they could ship into just "awkwardly long" shipping times.

P.S. do suspect that Apple will do a dog-and-pony show in 2022 though. A "look but don't touch" demo at least.