Discussion Apple Silicon SoC thread

[Eug]

M1
5 nm
Unified memory architecture - LP-DDR4
16 billion transistors

8-core CPU

4 high-performance cores
192 KB instruction cache
128 KB data cache
Shared 12 MB L2 cache

4 high-efficiency cores
128 KB instruction cache
64 KB data cache
Shared 4 MB L2 cache
(Apple claims the 4 high-effiency cores alone perform like a dual-core Intel MacBook Air)

8-core iGPU (but there is a 7-core variant, likely with one inactive core)
128 execution units
Up to 24576 concurrent threads
2.6 Teraflops
82 Gigatexels/s
41 gigapixels/s

16-core neural engine
Secure Enclave
USB 4

Products:
$999 ($899 edu) 13" MacBook Air (fanless) - 18 hour video playback battery life
$699 Mac mini (with fan)
$1299 ($1199 edu) 13" MacBook Pro (with fan) - 20 hour video playback battery life

Memory options 8 GB and 16 GB. No 32 GB option (unless you go Intel).

It should be noted that the M1 chip in these three Macs is the same (aside from GPU core number). Basically, Apple is taking the same approach which these chips as they do the iPhones and iPads. Just one SKU (excluding the X variants), which is the same across all iDevices (aside from maybe slight clock speed differences occasionally).

EDIT:



M1 Pro 8-core CPU (6+2), 14-core GPU
M1 Pro 10-core CPU (8+2), 14-core GPU
M1 Pro 10-core CPU (8+2), 16-core GPU
M1 Max 10-core CPU (8+2), 24-core GPU
M1 Max 10-core CPU (8+2), 32-core GPU

M1 Pro and M1 Max discussion here:

Page 78 - Discussion - Apple Silicon SoC thread

Page 78 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

M1 Ultra discussion here:

Page 109 - Discussion - Apple Silicon SoC thread

Page 109 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

M2 discussion here:

Page 127 - Discussion - Apple Silicon SoC thread

Page 127 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Second Generation 5 nm
Unified memory architecture - LPDDR5, up to 24 GB and 100 GB/s
20 billion transistors

8-core CPU

4 high-performance cores
192 KB instruction cache
128 KB data cache
Shared 16 MB L2 cache

4 high-efficiency cores
128 KB instruction cache
64 KB data cache
Shared 4 MB L2 cache

10-core iGPU (but there is an 8-core variant)
3.6 Teraflops

16-core neural engine
Secure Enclave
USB 4

Hardware acceleration for 8K h.264, h.264, ProRes

M3 Family discussion here:

Page 215 - Discussion - Apple Silicon SoC thread

Page 215 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

M4 Family discussion here:

Page 263 - Discussion - Apple Silicon SoC thread

Page 263 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[A///]

Expect AMD and Intel to do wonders when they move to 3nm.

2024 for amd and when is the magical date for intel with their intel co branding to minimize their incompetence?

 

[Doug S]

Yeah, it is 2P + 3E and still 5-core GPU

tvOS Overview - Apple Developer

I thought the earlier article had suggested otherwise, but your link is an unimpeachable source!

I too wonder how many dies with one bad little core they could possibly have, one little core is what 2% of the chip area? N5P was already pretty mature when Apple started selling phones with A15, so the yields couldn't have been that bad. Some rough assumptions and back of the envelope calculations indicate they've maybe built up an inventory of 500K to 750K A15s with one defective core. They'll run out of those in a month, but I guess that's better than tossing them.

I suppose they could find a use for A15s with a bad big core, using them in something like Homepod. Currently that uses an A8 but there's supposedly a refresh on the way. They might have some stock of something older like A12 to use in it though, we'll see. Either way it shouldn't need that much CPU power.

 

[poke01]

2024 for amd and when is the magical date for intel with their intel co branding to minimize their incompetence?

Intel will also use TSMC 3nm in 2024 with Arrow Lake.

 

[Roland00Address]

I thought the earlier article had suggested otherwise, but your link is an unimpeachable source!

I too wonder how many dies with one bad little core they could possibly have, one little core is what 2% of the chip area? N5P was already pretty mature when Apple started selling phones with A15, so the yields couldn't have been that bad. Some rough assumptions and back of the envelope calculations indicate they've maybe built up an inventory of 500K to 750K A15s with one defective core. They'll run out of those in a month, but I guess that's better than tossing them.

I suppose they could find a use for A15s with a bad big core, using them in something like Homepod. Currently that uses an A8 but there's supposedly a refresh on the way. They might have some stock of something older like A12 to use in it though, we'll see. Either way it shouldn't need that much CPU power.

but there is 4 different small cores to choose from, so if it’s 2% of area it is 4x2% salvaged chips if they had defects in those specific parts of the chip. Likewise the GPU which is same GPU but different configurations in the iPhone 13 and iPhone 14, so for $100 more (since Apple is selling both) you get the better GPU and any other differences between the models.

Apple ships about 240 million iPhones per year, plus other devices like iPad Minis, Apple TVs, the Apple Studio Display, and perhaps some other device I am forgetting.

The a15 has been in use for 1 year now, and if they are shipping it for 3 years as a main silicon in Apple devices I would not be surprised we are talking 600 perhaps 800 million chips sold. Being able to “reclaim” the small amount of chips with defects with CPU or GPU defects on a single cluster, will still be double digit millions of chips. It just makes “sense“ inside of my head.

 

[Doug S]

but there is 4 different small cores to choose from, so if it’s 2% of area it is 4x2% salvaged chips if they had defects in those specific parts of the chip. Likewise the GPU which is same GPU but different configurations in the iPhone 13 and iPhone 14, so for $100 more (since Apple is selling both) you get the better GPU and any other differences between the models.

Apple ships about 240 million iPhones per year, plus other devices like iPad Minis, Apple TVs, the Apple Studio Display, and perhaps some other device I am forgetting.

The a15 has been in use for 1 year now, and if they are shipping it for 3 years as a main silicon in Apple devices I would not be surprised we are talking 600 perhaps 800 million chips sold. Being able to “reclaim” the small amount of chips with defects with CPU or GPU defects on a single cluster, will still be double digit millions of chips. It just makes “sense“ inside of my head.

No chance they ship 600 or 800 million of them. That would be the total in all devices sold over three years, as if there is never an A16 or A17, or was an A13 or A14 previously.

I don't think anyone knows the exact product mix Apple sells, but that doesn't matter for estimating the number of each SoC Apple sells. Whatever the mix of "latest and greatest" vs "last year's model" vs "two year's ago model" they would end up shipping around the total of one year's iPhone sales over the three years. Now A15 will ship more because of the Pro / non Pro split (though indications are that the Pro models are much more popular so the non Pro might be only say 1/3 of the total of iPhone 14 sales)

So between that, and iPhone SE they might get to 350 million A15s. If you conservatively assume N5P has a 90% yield of defect free parts (I'm guessing it is higher, TSMC has reason to tune processes for yield over performance while Intel has reason to go the other way) then 10% of parts have some defect, and we'll say 8% of those have one bad E core but are otherwise OK. Let's call it 1% of the overall A15 production for simplicity. If they will sell 350 million A15s during the product life, they'll end up with around 3.5 million Apple TV candidates.

I couldn't find anything recent but this article claims Apple sold 13 million Apple TVs in a single QUARTER in 2019 and sells that many pretty much every quarter. Not all of them are the latest model, but 3.5 million chips would last them at best two or three months. Sure that saves them a lot of money getting 3.5 million chips for "free" but the overwhelming majority of 2022 Apple TVs will end up shipping with a fully functional A15.

 

[Eug]

Apple appears to be first in line for TSMC's Arizona chip manufacturing capacity (paywalls).

Bloomberg - Are you a robot?

Basically Tim said they'll start buying there in 2024 (which is when it's supposed to be completed). However, TSMC had previously said that first Arizona location would be N5 and N4. TSMC is also building a second location for N3, but I'm not sure when that will be completed.

WSJ News Exclusive | Chip-Making Juggernaut TSMC Eyes Multibillion-Dollar Arizona Factory Expansion

Taiwan Semiconductor Manufacturing, the world’s largest contract chip maker, plans to announce a second multibillion-dollar factory in Arizona.

www.wsj.com

 

[Roland00Address]

Doug I agree the numbers can be much lower than I outlined. We have no true insight of the product mix of Apple, both of us agree on that, for the reasons you outlined.

So cut those numbers in half and the point still stands. Apple is not wasting money with lost chips, and is getting a higher utilisation rate if they are still supply / wafer constrained and they are competing with other customers. An Apple TV does not need 4 efficiency cores over 3, it is an ideal situation where if one of those A15 devices had to lose a cpu core it makes since for one is not CPU bound with an Apple TV 4K while an iPhone SE or other device may want that extra cpu core for power saving purposes.

I would still argue that A15 utilisation is higher compared to previous generations over multiple years since the A16 is only on Pro models costing $1000 and $1000+ (before carrier deals), while all iPhones of this cycle from $429 to $899 is using the A15 aka last years chip.

Likewise very little A14s for that process of TSMC is getting all those M1 focused derivatives. A13s (aka a 7nm) part are still being used for the $329 iPad and the studio display. But Tim Cook for the most part has moved all the silicon to 5nm derivative TSMC foundries such as what we are calling 4nm and the two older 5nm foundry processes.

But yeah, maybe I am wrong and over estimated the final lifetime uses of A15s over several years. Maybe it is half that number over 3+ years, the point is still the same.

 

[Doug S]

Basically Tim said they'll start buying there in 2024 (which is when it's supposed to be completed). However, TSMC had previously said that first Arizona location would be N5 and N4. TSMC is also building a second location for N3, but I'm not sure when that will be completed.

This is just PR for TSMC and Apple. Sure they will still be using some N4/N5 stuff in 2024 - if nothing else for the 2022 Apple TV which will likely still be sold then - but 90% of their needs will have moved to N3. And by the time the N3 fab that's recently been talked about comes online (assuming it happens) 90% of Apple's needs will have moved to N2.

TSMC has been clear that their state of the art production will remain exclusive to Taiwan, and that's Apple's bread and butter. The only way Apple sources a meaningful portion of its chip needs from the US is if they someday switch to Intel's foundry.

 

[richardskrad]

The entire PC industry shipped 340 M units in 2021. The iPhone alone sold 238 M units. It’s kind of crazy how big Apple has become. It’s a behemoth.

 

[FlameTail]

If you conservatively assume N5P has a 90% yield of defect free parts (I'm guessing it is higher, TSMC has reason to tune processes for yield over performance while Intel has reason to go the other way) then 10% of parts have some defect, and we'll say 8% of those have one bad E core but are otherwise OK.

I can only imagine how bad the yields are for chips made on Samsung's nodes. No wonder Qualcomm ditched Samsung and moved to TSMC for their flagship chips.

 

[uzzi38]

Well I can see why. It's still based on 5mm. Until Apple moves to a new node like 3nm Pref/w will suffer.

Apples chips are not magic. Expect AMD and Intel to do wonders when they move to 3nm.

We should see funny stuff before that.

Reminder: At least for AMD a large amount of their product stack in 2024 is still going to be N4 based nodes.

Unrelated to that comment (I mean it), but I look forwards to seeing how Phoenix compares to M1/M2 Max (not like there'll be much of a difference between the two for efficiency if the M2 is any sign).

 

[jpiniero]

This is just PR for TSMC and Apple. Sure they will still be using some N4/N5 stuff in 2024 - if nothing else for the 2022 Apple TV which will likely still be sold then - but 90% of their needs will have moved to N3. And by the time the N3 fab that's recently been talked about comes online (assuming it happens) 90% of Apple's needs will have moved to N2.

They are going to sell plenty of the 2023 iPhone in 2024 I'm guessing. And that's not even including if they were to say put the 2023 Pro SoC in the 2024 base model.

 

[Doug S]

They are going to sell plenty of the 2023 iPhone in 2024 I'm guessing. And that's not even including if they were to say put the 2023 Pro SoC in the 2024 base model.

The 2023 iPhone will use N3/N3E, we'll have to see if the "non Pro gets last year's SoC" is the new normal or was just a one time thing caused by the price difference between LPDDR5 and LPDDR4x and had nothing really to do with the SoC. Everyone was expecting higher prices with inflation etc. but Apple surprised by holding the line. They may have had to make a few compromises to keep the BOM in line with their margin targets.

We could see all iPhone 15s with A17 as the whole line moves to LPDDR5.

 

[jpiniero]

The 2023 iPhone will use N3/N3E

It has to be using N4. They aren't going to use N3 for it, and N3E is too late.

 

[poke01]

It has to be using N4. They aren't going to use N3 for it, and N3E is too late.

N4P. That means one more year of slow upgrades.

 

[Doug S]

It has to be using N4. They aren't going to use N3 for it, and N3E is too late.

I don't buy that, when has Apple ever not used the latest and great process? Even if N3E isn't ready in time (I think it will be) and even if N3 didn't give any better performance or power the area gains alone are too massive considering the number of chips they need. But I guess we'll see.

Even if that's true it would be one year blip. TSMC has been clear that the leading edge will always remain in Taiwan, so the US fabs will always get the N+1 process that's probably fine for AMD but too late for the overwhelming majority of Apple's needs.

 

[uzzi38]

N4P. That means one more year of slow upgrades.

N4P is a similar improvement to power/perf to N3E. If we only get a small upgrade, it'll say a lot about Apple.

 

[uzzi38]

I don't buy that, when has Apple ever not used the latest and great process? Even if N3E isn't ready in time (I think it will be) and even if N3 didn't give any better performance or power the area gains alone are too massive considering the number of chips they need. But I guess we'll see.

Even if that's true it would be one year blip. TSMC has been clear that the leading edge will always remain in Taiwan, so the US fabs will always get the N+1 process that's probably fine for AMD but too late for the overwhelming majority of Apple's needs.

Except from a power/perf and cost effectiveness point of view N3 is not the "greatest" process. It's just straight up inferior.

 

[poke01]

N4P is a similar improvement to power/perf to N3E. If we only get a small upgrade, it'll say a lot about Apple.

Yes, if A17 is small upgrade then it will tell us that what made the A7 to A12 excellent was Gerald Williams who now works at Qualcomm.

 

[Nothingness]

Yes, if A17 is small upgrade then it will tell us that what made the A7 to A12 excellent was Gerald Williams who now works at Qualcomm.

Or rather it will tell us that the improvements to a uarch tend to bring less and less speedups with time. At some point you have to restart from scratch and evolve from there. But even that will show limits because the current uarch already is very good.

No single man, be it Gerald Williams or Jim Keller, makes a micro-architecture all alone. They are brilliant but I can guarantee all CPU companies have brilliant engineers with ideas.

 

[jpiniero]

N4P is a similar improvement to power/perf to N3E. If we only get a small upgrade, it'll say a lot about Apple.

N4P isn't that good. I'd expect Apple to move to N3E once it becomes available. But not the original N3.

 

[Eug]

It has to be using N4. They aren't going to use N3 for it, and N3E is too late.

I'm curious, where are you reading N3E is too late for the 2023 iPhones? FWIW, Nikkei disagrees with you.

Apple to use TSMC's next 3-nm chip tech in iPhones, Macs next year

Move will mark industry's first adoption of updated production technology

asia.nikkei.com

TAIPEI -- Apple aims to be the first company to use an updated version of Taiwan Semiconductor Manufacturing Co.'s latest chipmaking technology next year, with plans to adopt it for some of its iPhones and Mac computers, sources briefed on the matter told Nikkei Asia.

The A17 mobile processor currently under development will be mass-produced using TSMC's N3E chipmaking tech, expected to be available in the second half of next year, according to three people familiar with the matter. The A17 will be used in the premium entry in the iPhone lineup slated for release in 2023, they said.

---

My undereducated assumption was that N3 would be utilized for M2 Pro and M2 Max/Ultra/Ultrax2.

BTW, my (faint) hope was that the reason the Mac mini was delayed was so that they could release an N3 M2 Pro Mac mini at around the same time as the M2 Pro MacBook Pro (+/- N3 M2 Max Mac Studio). However, my expectation now after reading all the tea leaves is that there will be no M2 Pro Mac mini at all. They'll probably stick with N5 M2 for the Mac mini, and the delay in release is a combination of marketing decisions and supply chain constraints for chips other than the M-series SoC. (They'd rather redirect those components to the MacBook Air and 13" MacBook Pro, and of course the iPhone 14 series, since those are the cash cows.)

 

[Roland00Address]

So are we not getting an Arm Mac Pro?
(they promised 2 year transition of all their lines after the M1 introduced.)

 

[Eug]

So are we not getting an Arm Mac Pro?
(they promised 2 year transition of all their lines after the M1 introduced.)

When they released Mac Studio, Apple actually made a point of saying the Mac Pro is still coming (which is unusual for Apple), although they also conspicuously refrained from mentioning the 2 year time frame this time around. They also said M1 series was complete so we know it isn't based on M1, which kind of makes sense since M1 series (and the OS support) doesn't seem to be designed to go past M1 Max x 2, as stated by Hector Martin of Asahi Linux fame. (ie. Hector Martin predicted M1 Ultra before it was released, but at the same predicted that there would be nothing past M1 Ultra in the M1 series.)

Since then there have been rumours of a development box which has PCIe slots but with no discrete GPU support, and which stays with the unified memory design with no external RAM slots. This comes from the same person who leaked the Mac Studio details before its launch, including its actual name and the SoC configuration. (Nobody believed him of course, but then it got released. )

So, we know that Mac Pro is indeed coming, but we don't know the specifics besides rumours of course. It seems the consensus guess is Mac Pro will be N3-something in 2023, based on M2 series.

 

[LightningZ71]

It seems odd that it won't have dGPU support. I get that their iGPUs are plenty powerful, but, that seems short sided as there are some very powerful pro series cards with specific purposes. However, I can see the other uses for PCIe slots, such as additional NVME capacity, using CXL for RAM expansion, and additional I/O capabilities.