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

 

[FlameTail]

Wow. I am impressed how there is a sprawling discussion about AI models in the Apple Silicon thread, while Apple Silicon itself is being discussed in other threads... Quite the irony.

Anyways,

Mark Gurman dropped some info:

Apple preparing release of M4 MacBook Pro this fall - 9to5Mac

M4 MacBook Pro, Mac mini, and iMac remain on track to get a notable refresh later this year with the M4 chip & more, according to a new report

9to5mac.com

M4 is coming to Mac Mini, Macbook Pro., iMac in 2024.

In 2025, it will go into the Macbook Air, Mac Studio, Mac Pro.

Could M5 be released in 2025 then?

 

[johnsonwax]

Why do you say they're not selling that?

In part because none of these guy are reporting revenues consistent with them selling that. You seem to think 'trying to sell something' is the same thing as actually selling something. It's a wild claim for you to make in the middle of tech sell-off following a peak in the Gartner Hype Cycle for AI. A LOT of investors followed your reasoning and they're looking at earnings reports and wondering why nobody is buying this stuff.

 

[johnsonwax]

Wow. I am impressed how there is a sprawling discussion about AI models in the Apple Silicon thread, while Apple Silicon itself is being discussed in other threads... Quite the irony.

Not really. Intel imploding is forcing a lot of narrowly tech focused people to step back and try and make sense of the larger market dynamics, as a lot of people wanted to believe that Apple had no direct impact on Intel's business. So, of course those discussions are happening anywhere but here - we're the least likely to need the lesson.

Let's ask the big question: can x86 remain competitive given Microsoft's shift in priority to ARM and Apple's increased ability to outspend ARM/Intel? And what does that mean for Apple Silicon?

Anyway, Gurman suggesting M4 Pro/Max will be this year. Ok.

 

[mvprod123]

Not really. Intel imploding is forcing a lot of narrowly tech focused people to step back and try and make sense of the larger market dynamics, as a lot of people wanted to believe that Apple had no direct impact on Intel's business. So, of course those discussions are happening anywhere but here - we're the least likely to need the lesson.

Let's ask the big question: can x86 remain competitive given Microsoft's shift in priority to ARM and Apple's increased ability to outspend ARM/Intel? And what does that mean for Apple Silicon?

Anyway, Gurman suggesting M4 Pro/Max will be this year. Ok.

It's much more interesting to see what the M4 Ultra will offer. It will outperform the top desktop solutions from Intel/AMD.

 

[name99]

Wow. I am impressed how there is a sprawling discussion about AI models in the Apple Silicon thread, while Apple Silicon itself is being discussed in other threads... Quite the irony.

Anyways,

Mark Gurman dropped some info:

Apple preparing release of M4 MacBook Pro this fall - 9to5Mac

M4 MacBook Pro, Mac mini, and iMac remain on track to get a notable refresh later this year with the M4 chip & more, according to a new report

9to5mac.com

M4 is coming to Mac Mini, Macbook Pro., iMac in 2024.

In 2025, it will go into the Macbook Air, Mac Studio, Mac Pro.

Could M5 be released in 2025 then?

Your timeline makes zero sense.

The obvious hypothesis is that M4 on Mac is waiting for Sequoia, not least because some of the new functionality or performance is best showed off by the new OS (and of course that's most likely in the AI space).
This means that the hardware is being delayed until ~end September/beginning October - MAYBE they ship with 15.0 but demo with 15.1 in the hardware event, to show how great it is; or maybe they delay the hardware until an event two weeks or so before 15.1 is ready?

Either way, delaying the MBA (aka the best selling Mac) makes no sense. And if MacBook Pro is ready that means Max is ready. And people have become antsy about the Studio/Pro delay...
I can see a reason to delay Pro/Max vs base M, but we have that delay already baked into the M4 release a few months ago.

Maybe something like at the iPhone event they announce the new Macs (up to Pro/Max levels) then maybe December or so they announce the next step - superficially an event for the new Ultra's so Studios and Mac Pro, but the "one more thing" is announcing the Extreme?

On the one hand, sales of Studio's and Pro's don't care much about Christmas, so that can be announced anytime. On the other hand, I don't think Studio/Pro base is big enough to justify an event, but announcing the Extreme just via Press Release seems unlikely!
So what's something else they can throw in, an event some time around Dec..Feb that can accommodate something for the masses, but also include Studio/Pro?

Seems way too soon for a new Vision Pro model?
Maybe some sort of combo everything new but not too exciting event - new aTV, new AirPods, new HomePods, maybe some sort of Services updates like a deal with MGM for movies? Consumer front-end, Studio+Pro backend?

I'm trying to think through from the event backwards to what sort of hardware we might see on what sort of schedule.

 

[jdubs03]

That info is from Mark Gurman, whom if I’m not mistaken has the best track record for Apple leaks/info.

 

[johnsonwax]

Your timeline makes zero sense.

The obvious hypothesis is that M4 on Mac is waiting for Sequoia, not least because some of the new functionality or performance is best showed off by the new OS (and of course that's most likely in the AI space).
This means that the hardware is being delayed until ~end September/beginning October - MAYBE they ship with 15.0 but demo with 15.1 in the hardware event, to show how great it is; or maybe they delay the hardware until an event two weeks or so before 15.1 is ready?

Either way, delaying the MBA (aka the best selling Mac) makes no sense. And if MacBook Pro is ready that means Max is ready. And people have become antsy about the Studio/Pro delay...
I can see a reason to delay Pro/Max vs base M, but we have that delay already baked into the M4 release a few months ago.

Maybe something like at the iPhone event they announce the new Macs (up to Pro/Max levels) then maybe December or so they announce the next step - superficially an event for the new Ultra's so Studios and Mac Pro, but the "one more thing" is announcing the Extreme?

On the one hand, sales of Studio's and Pro's don't care much about Christmas, so that can be announced anytime. On the other hand, I don't think Studio/Pro base is big enough to justify an event, but announcing the Extreme just via Press Release seems unlikely!
So what's something else they can throw in, an event some time around Dec..Feb that can accommodate something for the masses, but also include Studio/Pro?

Seems way too soon for a new Vision Pro model?
Maybe some sort of combo everything new but not too exciting event - new aTV, new AirPods, new HomePods, maybe some sort of Services updates like a deal with MGM for movies? Consumer front-end, Studio+Pro backend?

I'm trying to think through from the event backwards to what sort of hardware we might see on what sort of schedule.

Delaying MBA makes sense if my previous hypothesis that Apple launched M4 in iPad first (talk about not making sense) because TSMC had capacity ahead of A series production, Apple chose a product that fit that capacity number knowing they'd have to switch off of M series to A series, because Apple isn't splitting the phones between current and previous nodes, they're having to launch all models on the current node, increasing demand for the current node by at least ⅓ over last year (80M chips, roughly). If that bumps M series production for a while, how do you manage this rollout?

Put iPad up front because it avoids the awkward 'why no other Mac for 8 months' if you start with MBA, puts the Macs behind Sequoia as you note, M series priority shifts to Pro/Max because MBA may be the best selling, but MBP is the most important, mini is small, iMac isn't huge, and MBA catches up when TSMC has capacity.

The problem Apple has to manage here is that they're drinking all of their products through the same straw. They could invest in more TSMC capacity but they'd have to pay for that entirely themselves, because it's not clear there are other customers that can absorb that kind of volume at the price Apple can afford to pay (Intel sure as shit can't now). So their best scenario is to integrate under the product curve, divide by the total product roll-out timetable, and work out the cross-sectional area of that straw, pay for that much capacity, and then rearrange the products to fit through there in the necessary timeline, understanding that the iPhone is the most important product and is on the most rigid timeline, so everything builds around A series production.

This all presumes that Pro/Max is on the same N3E node, and isn't being pushed back to an N3P one as I also previously wondered. If the rumor is correct that seems to rule that idea out since it doesn't sound as though there's any way to get a N3P product out in 2024, even under the most optimistic scenario. So, indeed it's all the same straw.

 

[johnsonwax]

Maybe something like at the iPhone event they announce the new Macs (up to Pro/Max levels) then maybe December or so they announce the next step - superficially an event for the new Ultra's so Studios and Mac Pro, but the "one more thing" is announcing the Extreme?

On the one hand, sales of Studio's and Pro's don't care much about Christmas, so that can be announced anytime. On the other hand, I don't think Studio/Pro base is big enough to justify an event, but announcing the Extreme just via Press Release seems unlikely!
So what's something else they can throw in, an event some time around Dec..Feb that can accommodate something for the masses, but also include Studio/Pro?

Seems way too soon for a new Vision Pro model?
Maybe some sort of combo everything new but not too exciting event - new aTV, new AirPods, new HomePods, maybe some sort of Services updates like a deal with MGM for movies? Consumer front-end, Studio+Pro backend?

I'm trying to think through from the event backwards to what sort of hardware we might see on what sort of schedule.

Also, Apple doesn't announce in December. November they do. Doesn't have to do with those sales corresponding to holidays, rather that they already have their supply chain hands full with the stuff that does correspond to holidays and there's no reason to complicate that further - best to delay it to Jan/Feb.

2 more years for a new Vision Pro. I'm seeing signs Apple has more balls in the air than they're comfortable with given the AI stuff. My guess is they go easy on new stuff for the next year until things settle down.

 

[yottabit]

Delaying MBA makes sense if my previous hypothesis that Apple launched M4 in iPad first (talk about not making sense) because TSMC had capacity ahead of A series production, Apple chose a product that fit that capacity number knowing they'd have to switch off of M series to A series, because Apple isn't splitting the phones between current and previous nodes, they're having to launch all models on the current node, increasing demand for the current node by at least ⅓ over last year (80M chips, roughly). If that bumps M series production for a while, how do you manage this rollout?

Put iPad up front because it avoids the awkward 'why no other Mac for 8 months' if you start with MBA, puts the Macs behind Sequoia as you note, M series priority shifts to Pro/Max because MBA may be the best selling, but MBP is the most important, mini is small, iMac isn't huge, and MBA catches up when TSMC has capacity.

The problem Apple has to manage here is that they're drinking all of their products through the same straw. They could invest in more TSMC capacity but they'd have to pay for that entirely themselves, because it's not clear there are other customers that can absorb that kind of volume at the price Apple can afford to pay (Intel sure as shit can't now). So their best scenario is to integrate under the product curve, divide by the total product roll-out timetable, and work out the cross-sectional area of that straw, pay for that much capacity, and then rearrange the products to fit through there in the necessary timeline, understanding that the iPhone is the most important product and is on the most rigid timeline, so everything builds around A series production.

This all presumes that Pro/Max is on the same N3E node, and isn't being pushed back to an N3P one as I also previously wondered. If the rumor is correct that seems to rule that idea out since it doesn't sound as though there's any way to get a N3P product out in 2024, even under the most optimistic scenario. So, indeed it's all the same straw.

Man, I'd hate to be on their purchasing team. That makes my head spin!

 

[johnsonwax]

Man, I'd hate to be on their purchasing team. That makes my head spin!

Eh. Best supply chain management in the world. And for a company their size they have VERY few products (lot of quantity, but not a lot of different parts to juggle), so it shouldn't really be that hard for them.

 

[FlameTail]

Your timeline makes zero sense.

It's not mine. It's Gurman's.

 

[poke01]

I think the timeline is accurate. M4 Pro and Max will come this year in the MBP. shouldn’t surprise anyone.

The bigger Mac desktops will come by mid 2025.

Apple wants an annual release, let’s see if that’s true. I think it is.

 

[johnsonwax]

I think the timeline is accurate. M4 Pro and Max will come this year in the MBP. shouldn’t surprise anyone.

The bigger Mac desktops will come by mid 2025.

Apple wants an annual release, let’s see if that’s true. I think it is.

I suspect annual release will prove unsustainable.

 

[poke01]

I suspect annual release will prove unsustainable.

Now that Apple is also using the M-series for Cloud compute, it might be sustainable.

I could see them releasing the Ultra every other gen. The IP architecture updates are going to be annual thanks to iPhone, as long as Apple doesn’t stop those then scaling the architecture in future M series products shouldn’t be too costly.

 

[jdubs03]

They have competition. They’re not going to want to cede any ground to their CPU or OEM competitors.

 

[okoroezenwa]

Annual cycle for laptops? I can see it. Annual for desktops? Probably not, and we can already see that now.

 

[FlameTail]

Annual cycle for laptops? I can see it. Annual for desktops? Probably not, and we can already see that now.

Annual cycle for iPads too.

 

[mikegg]

In part because none of these guy are reporting revenues consistent with them selling that. You seem to think 'trying to sell something' is the same thing as actually selling something. It's a wild claim for you to make in the middle of tech sell-off following a peak in the Gartner Hype Cycle for AI. A LOT of investors followed your reasoning and they're looking at earnings reports and wondering why nobody is buying this stuff.

OpenAI's revenue is estimated to be $4 billion/year.

My girlfriend works at PwC. They just launched an internal "ChatPwC" based on GPT4. Thus far, they were barred from using the public ChatGPT. It didn't stop the employees from using the public ChatGPT anyways because it was so useful for them. Now they have an internal one.

In my own company, we use GPT4's API for real business use cases.

I feel like people here are living under a rock.

Edit: Girlfriend literally sent me this today

 

[Gideon]

OpenAI's revenue is estimated to be $4 billion/year.

Your post will age veeery poorly:

https://twitter.com/x/status/1820627845393379756

Obviously LLMa are here to stay and are useful tools for many fields (mine included). Also many AI models (unrelated to transformers) were already very useful long before LLMs

Also it's very likely that at some point someone will invent something better than the transformers we curently have, revitalizing the industry and proving new applications for "AI" (god I hate that umbrella term, as it's way more confusing to ithe average joe than the predecessors. MUCH worse than "cloud", etc)

But both these facts have almost no relevance to the current AI bubble (that's about to burst) taking billions of investment with it.

Internet also continued to be relevant before and after the dotcom boom.

The LLM Cloud AI part has been addressed by many previous posts. "AI PCs" and the like are just desparate bandwagoning, on the same hypetrain by MS and OEMs to sell more laptops in a post covid "hangover" when most people just upgraded and have little need for new PCs

Don't believe me? Watch this interview with the CEO of Framework Laptops essentially saying the same thing starting from 15:45 (excellent interview all round btw):

 

[johnsonwax]

OpenAI's revenue is estimated to be $4 billion/year.

And they're estimated to lose $5B this year. Intel has a lot of revenue. But it's not enough revenue for the business that's beneath them.

AI models like ChatGPT right now (and maybe this will change) are diminishing returns for accelerating cost of training. That's not necessarily a problem if you're sitting on a profitable business and can mange things in the numerator, but if you're still chasing profitability you have to do that in the denominator.

Amazon just admitted they've spent $25B on Alexa and had no path to profitability. They were doing the same thing - if we dump increasing amounts of money in the denominator eventually the path to profitability will reveal itself. It never did. And that's not a business that has accelerating costs.

We just has a story of a Pharma CIO who dumped Copilot (ChatGPT under the hood) because the output was kind of garbage and not worth the money. He was a customer, he bought the product and in hindsight he said it wasn't worth it. We don't know where that sweet spot will be for each business. You seem to think they have it because you now have it on your desk. The Pharma CIO put it on the desk of his people and then got rid of it when it didn't produce results.

 

[johnsonwax]

To sweep this back on topic a bit:

One of the benefits of Apple's approach to doing this on device is that all of the marginal costs have either already been paid for by the user or will be paid for by the users. The user already bought the iPhone or whatever and the user will pay for the electricity to run the model and do so subconsciously - nobody is going to look at their electric bill and say 'damn, that Apple Intelligence is costing me a lot of money'.

However, where Apple will need to manage things is in battery life because the continual training/tuning and the ease of use of a Siri query (especially if users find it to be more useful) will meaningfully cut into that. It likely won't matter on the Mac, but it will matter on the iPhone. Battery life is not perceived by users as a continuous function. Nobody gives a shit if they have 10% battery life or 80% battery life at the end of the day - they plug in the phone and the next day they're at 100%. What matters is if the phone gets to the end of the day or not.

That's likely going to get a lot harder for Apple to manage, especially if they are leaning into the kinds of convenience AI that they're doing with Siri which has more utility on iPhone. The 'rewrite your term paper' AI (let's call these task intensive models) is really geared toward the Mac.

How they manage power draw, increase in battery capacity, and also what kinds of AI models they roll out in the future (convenience or task intensive) is going to be interesting to observe. Also, how Watch picks up some of these features because it becomes an additional source of convenience AI that is hitting the same iPhone battery. I would expect battery life to become a thing again.

 

[Doug S]

To sweep this back on topic a bit:

One of the benefits of Apple's approach to doing this on device is that all of the marginal costs have either already been paid for by the user or will be paid for by the users. The user already bought the iPhone or whatever and the user will pay for the electricity to run the model and do so subconsciously - nobody is going to look at their electric bill and say 'damn, that Apple Intelligence is costing me a lot of money'.

However, where Apple will need to manage things is in battery life because the continual training/tuning and the ease of use of a Siri query (especially if users find it to be more useful) will meaningfully cut into that. It likely won't matter on the Mac, but it will matter on the iPhone. Battery life is not perceived by users as a continuous function. Nobody gives a shit if they have 10% battery life or 80% battery life at the end of the day - they plug in the phone and the next day they're at 100%. What matters is if the phone gets to the end of the day or not.

That's likely going to get a lot harder for Apple to manage, especially if they are leaning into the kinds of convenience AI that they're doing with Siri which has more utility on iPhone. The 'rewrite your term paper' AI (let's call these task intensive models) is really geared toward the Mac.

How they manage power draw, increase in battery capacity, and also what kinds of AI models they roll out in the future (convenience or task intensive) is going to be interesting to observe. Also, how Watch picks up some of these features because it becomes an additional source of convenience AI that is hitting the same iPhone battery. I would expect battery life to become a thing again.

Why do you think it would be an issue for battery life? How much "AI" do you think people will be using? Do you think it is going to be running in the background all the time or something?

You might ask Siri to do something for a couple dozen times a day, max. You aren't going to be in AI utilizing conversions with your phone for several hours straight like one might in a social media doomscroll. Maybe someday if AI advances beyond LLMs and into true AGI territory that might happen, but it isn't something Apple or anyone else needs to worry about in 2025!

 

[johnsonwax]

Why do you think it would be an issue for battery life? How much "AI" do you think people will be using? Do you think it is going to be running in the background all the time or something?

You might ask Siri to do something for a couple dozen times a day, max. You aren't going to be in AI utilizing conversions with your phone for several hours straight like one might in a social media doomscroll. Maybe someday if AI advances beyond LLMs and into true AGI territory that might happen, but it isn't something Apple or anyone else needs to worry about in 2025!

So, there's two things it needs to do:
1) respond to prompts which is going to be proportional to how much utility people find in it. AI is designed to add value, to add utility, so Apple wants that to go up. If AI really adds a lot of utility to Siri, it might go up a lot on iPhone. Success has a cost in this case.
2) the device has to do the ongoing work of including your content in the model - consuming your emails and texts and photos and all that and that's pretty expensive even if you aren't using AI much. Now, if Apple can shuffle all of that to the charging cycle - then no problem. If they need to do that more incrementally so that AI has knowledge of things that have happened today, then they won't be able to do that as much. Not sure what'll be possible here.

As for several hours straight - I know people that now NEVER type out texts - they are 100% Siri dictated (how much AI gets added to dictation I don't know). I think for Watch wearers Siri becomes the primary interface if its utility increases, so I think it can be a lot.

 

[Doug S]

So, there's two things it needs to do:
1) respond to prompts which is going to be proportional to how much utility people find in it. AI is designed to add value, to add utility, so Apple wants that to go up. If AI really adds a lot of utility to Siri, it might go up a lot on iPhone. Success has a cost in this case.
2) the device has to do the ongoing work of including your content in the model - consuming your emails and texts and photos and all that and that's pretty expensive even if you aren't using AI much. Now, if Apple can shuffle all of that to the charging cycle - then no problem. If they need to do that more incrementally so that AI has knowledge of things that have happened today, then they won't be able to do that as much. Not sure what'll be possible here.

As for several hours straight - I know people that now NEVER type out texts - they are 100% Siri dictated (how much AI gets added to dictation I don't know). I think for Watch wearers Siri becomes the primary interface if its utility increases, so I think it can be a lot.

Consuming your emails and photos is a batch process that happens once. It doesn't need to reread them after it has seen them once. Even if you're taking hundreds of photos a day or receiving hundreds of emails a day (of a type it cares about, it isn't going to bother with spam or marketing) that's what a second for each one or 5-10 minutes of load per day. Yikes that might run 3-4% off your battery, its the end of the world!

 

[mikegg]

Your post will age veeery poorly:

And they're estimated to lose $5B this year. Intel has a lot of revenue. But it's not enough revenue for the business that's beneath them.

Both clickbait articles.

Now I know your sources. I mean, this is like a bystander who reads clickbait articles and think he knows everything.

Literally every enterprise company I have inside info from friends and ex coworkers is making a private GPT.

But go ahead. Believe some idiotic claim about how OpenAI is losing $5 billion/year when they have zero insider information on their financials.

And what’s even more laughable is making an equivalence to Alexa, which is a voice AI that has been shown time and time again to not have a viable business model. How the hell is Alexa even remotely comparable to GPT4? The only similarity they have is that they both train on GPUs. That’s it. You might as well compare GPT4 to an orange.