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

 

[mikegg]

yep

Welcome to the party.

 

[mikegg]

https://twitter.com/x/status/1883276535643455790

View attachment 115664


when M4 ultra?

Forget the Ultra. When 2x Ultra chip?

Give us 512GB of RAM at 2,184 TB/s bandwidth.

Heck, make a version with HBM for the Mac Pro since power consumption and space does not matter at that size and use case. I'm sure their custom AI chip efforts will already use HBM.

 

[Doug S]

This is why I still say the next Ultra will be M5 and not M4. This means M5 products will be available to ship about May thevsame timeframe as the earliest estimates of an Ultra launch, with some saying that it might be even be months later.

Just because they ship M5s in the spring for say iPad Pro or whatever doesn't mean that M5P/M5M let alone M5U will be ready within a few months of that. They could release an M5 based iPad Pro tomorrow and it wouldn't make me believe that a Mac Pro shipping this summer would be anything other than M4 based.

It is also the first IP that is specifically designed to power Apple Intelligence datacenters.

I'm still 100% confident Apple goes fully custom for that. They may have used some M2Us initially because that's all they had, but it would be stupid to waste a bunch of silicon area on stuff you don't need to shoehorn the same Apple Silicon you ship to customers for internal AI clouds. They'll want to swap out LPDDR for HBM, remove all the unnecessary cruft like display drivers and USB controllers, and ship cores designed specifically for their needs rather than making due with their GPU or NPU cores they ship to customers.

Now maybe this new fully custom thing isn't ready yet and they'll use some M4Us for that, but the plan will be to stop doing so as soon as the fully custom stuff is ready. If that design targets N2 they may have to make due with M4U for a few quarters.

Lastly, it will be the first IP to respond to how close Qualcomm is to Apple’s single threaded performance as well as Qualcomm’s GPU performance with 8 Elite.

I don't think Apple cares at all about "responding" to Qualcomm. They aren't in competition with Qualcomm, and Qualcomm's relative performance doesn't impact Apple any more than the relative performance of AMD/Intel x86 does. Which is to say, very little to not at all. Now sure their engineers are people, and people have pride etc. so they may have their own desires on staying ahead but as a corporate goal no, not a chance.

 

[soresu]

I don't think Apple cares at all about "responding" to Qualcomm. They aren't in competition with Qualcomm, and Qualcomm's relative performance doesn't impact Apple any more than the relative performance of AMD/Intel x86 does. Which is to say, very little to not at all. Now sure their engineers are people, and people have pride etc. so they may have their own desires on staying ahead but as a corporate goal no, not a chance.

IMHO it's never been a truly apples to apples (hur hur pun) comparison anyway, because none of the other players have full product stack control across both hardware and software.

 

[mikegg]

I don't think Apple cares at all about "responding" to Qualcomm. They aren't in competition with Qualcomm, and Qualcomm's relative performance doesn't impact Apple any more than the relative performance of AMD/Intel x86 does. Which is to say, very little to not at all. Now sure their engineers are people, and people have pride etc. so they may have their own desires on staying ahead but as a corporate goal no, not a chance.

Personally, I think Apple does want to keep the ST performance crown. They talk about it all the time in marketing, including executives during interviews. They're proud of it.

 

[mikegg]

To get there they alone will need probably $100 trillion, any clue how they'll get that money considering their models aren't profitable and they can't afford the hardware, have no skin in the development of said hardware (meaning others could leverage their profitability to get priority access to said hardware), and literally do not have a functional (aka one that makes more money than it costs to run it) business model?

Right now they're lucky as they have Microsoft willing to throw their data centers behind it, but it costs a lot of money to build and run data centers (just look at how Amazon and everyone is like "oh it costs us money to run voice assistants and we don't really get money from it"; AI is that but even more ridiculous disparity in the resources needed for the outputs) and they're starting to face pushback. At some point Microsoft is gonna have to justify the expenses to their stock holders. They're basically building towards a nuclear future, by which I mean Microsoft is building towards a future where new data center will be scrutinized as much as a nuclear reactor, as people are starting to turn against such (because they don't employ nearly as much people as claimed, especially long term, they steal resources like water which are gong to become even more significant, and their power needs are gonna require nuclear reactors to operate).

Sam Altman can say ChatGPT will "solve physics" all he wants, but that's nonsense and people need to start pointing this out. People act like there's a simple A->B route, but there simply isn't. The current AI is not designed to do such things, and to get to the next levels of even this form of AI requires orders of magnitude increase in processing capability, that is beyond the grasp of the combined current tech industry. To get "PhD AI" we're probably looking at orders of magnitude investment of the entire worlds' GDPs. How that can be justified when there's thus far been no tangible benefits to AI, and its quickly being used to do harm.

There are ways to get improvements without that processing, but they're fundamentally opposed to how AI is being marketed, and I don't know of any company actually even attempting to use it to augment human workers, but rather to replace them (and usually they then have to go and pay slave wages in India to human workers that do the actual work, as seen with Amazon's purchase recognition "AI" system). The only groups likely to be doing that are cybrcriminals who will be using AI to make botnets, worms, and ransomware even more prevalent. The damage that will cause will make the Crowdstrike fiasco seem pale in comparison.



This makes no sense, unless you genuinely believe biological transistors will happen in that timeframe. My point being, iPhone is hardware interface. LLMs are software (and ones using basic inputs because they're not really capable of much more yet). The attempts at making dedicated AI devices have been so absurdly atrocious that they likely singlehandedly are responsible for people starting to question AI hype. No matter the AI you're gonna need an interface unless you functionally build a brain or biological processor that can be added to our brains. You really seem to be listening to people like Sam Altman who have a nonsensical vision of what reality is. By the time that happens the traditional iPhone will be obsolete, because the timeline for when that happens is decades out. We'll be looking at major interface changes sooner than that.



Because such a thing doesn't exist and nothing they've built so far is even attempting to actually do that?

And? Those models are not capable of what you're talking about as they require precision and accuracy, which their models are fundamentally incapable of (because they're designed completely oppositionally to that, where they've been accepting lower and lower precision and accuracy in order to try and push out AI using current processing).

Ironically, that is one of the few legitimate useful things AI is good for, assisting human PhDs who can manipulate it to perform straightforward but computationally intensive tasks that would otherwise take them half a lifetime's worth of work. So they can then take the output and do something actually useful with it, because the AI can't because its not even built to be able to do that and they don't even know how to develop such.

Your argument is, again, nonsensical as they're selling AI to anyone willing to pay them money, that doesn't magically make it smarter or whatever argument you're trying to make there. Much like how companies tried jumping on blockchain (at direction of know-nothing executives that don't understand the technology, let alone how their company would actually use it), even more companies are jumping on the AI bandwagon with again, no real idea of what it actually does, what they could use it for, or anything else. Most of them are just making chatbots that are there basically just to frustrate their customers so they can layoff their human customer service. And also so they can micromanage their workforce so that they can lay off as many of them as they can to cut short term costs to boost stock prices.

This post is going to age very poorly.

 

[Doug S]

Personally, I think Apple does want to keep the ST performance crown. They talk about it all the time in marketing, including executives during interviews. They're proud of it.

But that wouldn't involve a "response" to Qualcomm specifically. They aren't any closer than Intel/AMD.

Anyway the idea that M5 is Apple's "response" is massively flawed, since it would have already been taped out by the time the first Nuvia cores appeared on the market. M7 would be the first design that could potentially "respond", if Apple was actually going to do so.

Apple wants to get the best ST performance they can, just like everyone else. It isn't like they are holding back. So they can't do anything more than what they're already doing in "response". I suppose they could crank up the clock and burn more power, but that would go against decades of Apple's belief that power efficiency should be prioritized over raw performance. Their decision to leave PPC and their decision to leave x86 were both heavily motivated by being able to achieve better power efficiency in the ISA they were migrating to.

 

[MS_AT]

Their decision to leave PPC and their decision to leave x86 were both heavily motivated by being able to achieve better power efficiency in the ISA they were migrating to.

Or maybe by the fact they could do as they see fit with the ISA they were migrating too, what would be impossible with x64 and PPC. My point being, if Apple wanted they could make an efficient x64 CPU, it's just it would not be possible due to licensing.

 

[soresu]

it's just it would not be possible due to licensing

At this point it wouldn't matter even if they had an architectural license giftwrapped for pennies per chip.

They have already fully transitioned MacOS and most of its major apps to ARM, and the move is largely synergistic to their iOS platform on top of that.

 

[oak8292]

Or maybe by the fact they could do as they see fit with the ISA they were migrating too, what would be impossible with x64 and PPC. My point being, if Apple wanted they could make an efficient x64 CPU, it's just it would not be possible due to licensing.

I would say that the transitions were about both efficiency and economics. Apple has core designs (large and small) that are ‘printed’ in the hundreds of millions if not billions. The engineering resource is relatively inexpensive and potentially less per core than Intel.

Apple might have been able to design efficient cores with x86 if they could clear out all the backwards compatibility. The ARM cores that Apple uses have already jettisoned 32 bit compatibility. Intel is still supporting 32 bit in x86. Apple has a lot of freedom and say with the ARM ISA that is highly unlikely with x86.

Apple keeps developers on their toes with OS and hardware changes. Intel apparently supports old software that has been abandoned by its developers. Is there really software that hasn’t been updated in the 20 years since 64 bit hardware showed up.

 

[Eug]

For those looking for a deal on the M4 Mac mini, if you can't get edu pricing for whatever reason, it's now on the refurb store.

They screwed up on the pricing of the refurb'd old models though. The M2 16 GB Mac mini similarly configured costs MORE than the M4 on the refurb store. The only cheaper M2 is 8 GB. (There is no 8 GB M4.)

 

[Raqia]

A bit of a non-sequitur, but here's a good read:

https://archive.computerhistory.org/resources/access/text/2024/10/102809019-05-01-acc.pdf

Keller goes into his time at Apple and some of its advantages when it came to the economics and designs of its SoCs.

 

[Gideon]

A bit of a non-sequitur, but here's a good read:

https://archive.computerhistory.org/resources/access/text/2024/10/102809019-05-01-acc.pdf

Keller goes into his time at Apple and some of its advantages when it came to the economics and designs of its SoCs.

There was a video on YT of the interview as well (no time to look it up currently unfortunately)

 

[moinmoin]

There was a video on YT of the interview as well (no time to look it up currently unfortunately)

which was posted at

Page 368 - Discussion - Apple Silicon SoC thread

Page 368 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

and

Page 33 - Question - Zen 6 Speculation Thread

Page 33 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

before.

In general Computer History Museum does good stuff, always worth it.

 

[Eug]

For those looking for a deal on the M4 Mac mini, if you can't get edu pricing for whatever reason, it's now on the refurb store.

They screwed up on the pricing of the refurb'd old models though. The M2 16 GB Mac mini similarly configured costs MORE than the M4 on the refurb store. The only cheaper M2 is 8 GB. (There is no 8 GB M4.)

They fixed it. They are now selling the refurb M2 8/256 Mac mini for US$319 (or CA$439/US$309 in Canada). Short on RAM and storage, but nonetheless, I don't think Apple has ever sold any Mac for this low a price, and I also think this is the last time we will see it this low.

 

[Thibsie]

lol, the refurb store for Belgium doesn' t look anything like that unfortunately.

 

[Eug]

M4 Mac mini NAND modules on an exact fit daughter board are now available, if you're willing to buy from China and install it yourself.

https://www.aliexpress.us/item/1005008351841618.html

One of the reviews states that the daughter board is thinner than the OEM one so not quite as robustly built, but it fits and works fine.

I wonder why the 2 TB Toshiba costs so much more than the 2 TB SanDisk.

EDIT:

This is an install and review of a similar module from a different seller:

 

[eek2121]

I'm still 100% confident Apple goes fully custom for that. They may have used some M2Us initially because that's all they had, but it would be stupid to waste a bunch of silicon area on stuff you don't need to shoehorn the same Apple Silicon you ship to customers for internal AI clouds. They'll want to swap out LPDDR for HBM, remove all the unnecessary cruft like display drivers and USB controllers, and ship cores designed specifically for their needs rather than making due with their GPU or NPU cores they ship to customers.

Now maybe this new fully custom thing isn't ready yet and they'll use some M4Us for that, but the plan will be to stop doing so as soon as the fully custom stuff is ready. If that design targets N2 they may have to make do with M4U for a few quarters.

I don’t know about that. I do agree that they probably won’t use them in end user devices, but one thing they can do is get back into the server market. That would help subsidize their own costs. I am baffled they haven’t done that yet. People use mac minis as servers all the time, there are hosting companies that ONLY do that.
Outside of that, they could make the Mac Pro more workstation oriented and use a chip with the design you mentioned.

Just a thought.

 

[igor_kavinski]

I wonder why the 2 TB Toshiba costs so much more than the 2 TB SanDisk.

Because Sandisk has a reputation for failing a lot.

 

[Thibsie]

Because Sandisk has a reputation for failing a lot.

Really ? (Real question))

 

[igor_kavinski]

Really ? (Real question))

Yep. I've had bad luck with Sandisk flash drives suddenly failing as far back as 2008 when a really expensive and fast 4GB drive failed and lost all data on it. Worked afterwards with a full format but never trusted it with important data again.

Western Digital responds to claims that SanDisk SSD failures have design flaw

SanDisk's parent company has been facing lawsuits over several failed SSDs. Now, researchers have discovered that is issue is not in the firmware.

www.digitaltrends.com

SanDisk SSD Issues Caused by Major Hardware Weaknesses: Report

Bad solder and poorly-sized components.

petapixel.com

SanDisk Extreme SSDs are “worthless,” multiple lawsuits against WD say

Ars cited in two SanDisk SSD failure lawsuits filed yesterday.

arstechnica.com

 

[poke01]

I’m curious to see apples 1st modem this week, I’m sure it will lack in a big way.

 

[joshua95]

I’m curious to see apples 1st modem this week, I’m sure it will lack in a big way.

Not that release was confirmed? Why do you believe it will lack?

 

[poke01]

Not that release was confirmed? Why do you believe it will lack?

It will lack mmWave and it won’t be as fast and the call reception won’t be as good Qualcomms in the first iteration.

Apple had a hard time with developing modems

 

[Mopetar]

It will lack mmWave and it won’t be as fast and the call reception won’t be as good Qualcomms in the first iteration.

Apple had a hard time with developing modems

Didn't they just buy Intel's modem team last time? It wasn't like they were doing great work until they were acquired by Apple.

My understanding is that rather than just buying some other group that hadn't delivered anything remarkable, that this time they built their own team from scratch and have a group that won't be a batch of dysfunctional misfits coasting on a successful main business.

I'd be surprised if their first product were at the same level as Qualcomm, but Apple's chip team wasn't the giant they are now when they first started either.