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

 

[Eug]

If they're paying for 6 GB modules and not using any 4 GB modules, just give people 12 GB of RAM. No one who wants the top model is going to decide not to purchase it because the base model has 12 GB of RAM.

I'm now pretty sure this is what they are going to do with the M4 MacBook Pro, M4 MacBook Air, and M4 Mac mini. (I'm not convinced there will be an M4 iMac at all.)

As @poke01 says, putting 12 GB in a base iPad Pro that costs $999 a few months after the more expensive MacBook Air and even more expensive MacBook Pro, both of which start at 8 GB, would annoy people a lot more.

I'm thinking both the MBA and MBP will start at 12 GB with M4, and the M4 Pro will continue to start at 18 GB.

 

[The Hardcard]

If they're paying for 6 GB modules and not using any 4 GB modules, just give people 12 GB of RAM. No one who wants the top model is going to decide not to purchase it because the base model has 12 GB of RAM.

I don’t see how it is possible that Tim “The Supply Chain Margin Man” Cook would just pay straight up for higher spec parts. Whatever the reason, Apple is only paying for 8 GBs of RAM.

 

[Mopetar]

At some point the baseline RAM will increase beyond 8 GB. Now isn't any worse than later.

Artificially crippling memory modules is entirely different than them binning their CPUs with a core disabled for performance reasons.

 

[dr1337]

Whatever the reason, Apple is only paying for 8 GBs of RAM.

Or he's just passing the cost onto the customer. Apple is probably using these same modules in other products like the higher specc'd 36gb macbook pro, and maybe even on the m3 max 96gb sku.

 

[Eug]

At some point the baseline RAM will increase beyond 8 GB. Now isn't any worse than later.

Artificially crippling memory modules is entirely different than them binning their CPUs with a core disabled for performance reasons.

Now is worse for marketing reasons. I agree those marketing reasons suck from our point of view as consumers, but they're still reasons, and it was my expectation:

I expected M3 Macs to start at 8 GB, and they did.

I expected the 2024 iPad Pro to also start at 8 GB, not just because I thought it would be M3 (not M4), but also because of the timing vs. the M3 Macs, and also because Apple has a long history of rationing its upgrades. Because Apple was providing OLED and a new form factor, I figured they'd hold it back in some way, and the most obvious choice would be to hold it back in terms of memory at 8 GB.

Then in 2026 they'd release a new iPad Pro at 12 GB. I typically try to wait for memory upgrades before buying my iDevices, but in some cases I won't, when there are other big improvements (like OLED), or when I just need a machine relatively urgently. This year I fall into both categories, with the latter because I'm just using an A8X iPad Air 2 temporarily.

On another note, speaking of OLED, it seems people are complaining about graininess of the OLED displays at low brightness levels. I had a look at some of the pictures and it looks like the consistency of the brightness of adjacent individual pixels is not great when at low brightness levels. This makes for a somewhat grainy appearance on darker backgrounds if you look at it very closely - pixel peeping. Comparison LCD pictures show the pixel brightness to be more consistent at low brightness levels, but with the drawback of no true black levels.

 

[Glo.]

Or he's just passing the cost onto the customer. Apple is probably using these same modules in other products like the higher specc'd 36gb macbook pro, and maybe even on the m3 max 96gb sku.

Yes, HOWEVER, in MBPs you have 6400 LPDDR5.

In M4 iPad Pro you have 7500 LPDDR5X.

 

[Nothingness]

It's not a silly guess and Maynard is right.

That was not the point Maynard contested in what I wrote. I only talked about SSVE. Not what SME or SME2 allows you to do. I think my guess was pretty clear about that.

SME does have vector instructions that put the results in the ZA storage and usage them give you roughly 8x improvement over base SSVE versions. It's the nature of the AMX unit — the ALU is associated with the matrix storage, not vector storage. It is likely that the instructions that target Z registers as destination incur double data movement cost.

Again, yes SME(2).

Using FMLA that writes to ZA tile I get 250 GFLOPS.

That's SME2, not SSVE Is this that code? Nice work!

The vector ZA instructions have been added in SME2, so that's probably why you might have been not aware of them.

I'm not a vector programmer by far (interestingly writing simulators does not make you any good at writing code for simulated target), but I know the differences between SME(2) instructions and SSVE

What we know at this point is that running SVE code in streaming mode will likely give bad performance and so SVE use should be limited until we know more.

 

[The Hardcard]

Or he's just passing the cost onto the customer. Apple is probably using these same modules in other products like the higher specc'd 36gb macbook pro, and maybe even on the m3 max 96gb sku.

Cook doesn’t charge his customers extra money so he can give it to somebody else. The low end iPad Pro is 8 GBs because Apple is paying for 8 GBs. Any business decision or technical reason for a supplier to give a higher spec best believe is 100% on them.

Companies usually have to agree to give Apple decided upon volumes of components. The most likely explanation is that for whatever reason they weren’t fully ready to supply the needed volumes of 8 GBs of RAM. So they make up the difference with higher spec parts at no extra charge.

 

[The Hardcard]

Mistake post

 

[dr1337]

Cook doesn’t charge his customers extra money so he can give it to somebody else. The low end iPad is 8 GBs because Apple is paying for 8 GBs. Any business decision or technical reason for a supplier to give a higher spec best believe is 100% on them.

Companies usually have to agree to give Apple decided upon volumes of components. The most likely explanation is that for whatever reason they weren’t fully ready to supply the needed volumes of 8 GBs of RAM. So they make up the difference with higher spec parts at no extra charge.

If the DRAM company were cutting apple a special deal, they surely would have rebranded the dies or had them blanked over, especially if it were a deal for millions of chips.

And if every model has the 6gb modules, then there is no hypothetical 4gb shortage from the suppliers because they aren't used anyways. So why give apple a special price? The memory fabs would rather sell less silicon for a higher profit, obviously, why would they cut apple a special deal at all?

The fabs don't say, look we know you ordered 4gb modules, but oops we only made 6gb ones, here, have them for the same price!

Like yeah they could do that but lol that makes no sense in any functional regard that I can foresee currently.

 

[Eug]

If the DRAM company were cutting apple a special deal, they surely would have rebranded the dies or had them blanked over, especially if it were a deal for millions of chips.

And if every model has the 6gb modules, then there is no hypothetical 4gb shortage from the suppliers because they aren't used anyways. So why give apple a special price? The memory fabs would rather sell less silicon for a higher profit, obviously, why would they cut apple a special deal at all?

The fabs don't say, look we know you ordered 4gb modules, but oops we only made 6gb ones, here, have them for the same price!

Like yeah they could do that but lol that makes no sense in any functional regard that I can foresee currently.

FWIW, once I ordered Crucial RAM direct from Crucial for my Mac, and got Samsung RAM from them. That confused the heck out of me, since Crucial is owned by Micron.

Anyhow Apple gets special pricing from everyone it deals with... except maybe Qualcomm...

 

[Glo.]

FWIW, once I ordered Crucial RAM direct from Crucial for my Mac, and got Samsung RAM from them. That confused the heck out of me, since Crucial is owned by Micron.

Which must mean that Samsung owns Micron.

Right?

 

[Eug]

Which must mean that Samsung owns Micron.

Right?

Ha. At that time I think I may have even wondered that myself.

But for the record just in case somebody else reading this might get confused, no Micron is not owned by Samsung. I guess when Crucial runs out of stock of their own Micron memory, instead of just stopping selling the product, they're happy to sell someone else's. Or at least they were, once upon a time. (This was many years ago, when you could actually upgrade the memory in Macs.)

 

[The Hardcard]

If the DRAM company were cutting apple a special deal, they surely would have rebranded the dies or had them blanked over, especially if it were a deal for millions of chips.

And if every model has the 6gb modules, then there is no hypothetical 4gb shortage from the suppliers because they aren't used anyways. So why give apple a special price? The memory fabs would rather sell less silicon for a higher profit, obviously, why would they cut apple a special deal at all?

The fabs don't say, look we know you ordered 4gb modules, but oops we only made 6gb ones, here, have them for the same price!

Like yeah they could do that but lol that makes no sense in any functional regard that I can foresee currently.

It makes many orders of magnitude more sense than Tim Cook paying more than necessary for anything. Cook’s control of supply chain costs is just as legendary as Job’s vision and hyperbole, Wozniak’s technical prowess, or Ive’s thin and minimalist design.

All speculation must pass through the hallowed gates of Cook’s renowned refusal to pay more than necessary. Unless there is hard evidence to the contrary, always assume the supplier is fully eating the costs of any and all discrepancies and inexplicables.

If someone got Cook to pay for something he wasn’t planning on using, he deserves a multi-million dollar bonus and a vice presidency.

 

[name99]

OK, so

That was not the point Maynard contested in what I wrote. I only talked about SSVE. Not what SME or SME2 allows you to do. I think my guess was pretty clear about that.


Again, yes SME(2).


That's SME2, not SSVE Is this that code? Nice work!


I'm not a vector programmer by far (interestingly writing simulators does not make you any good at writing code for simulated target), but I know the differences between SME(2) instructions and SSVE

What we know at this point is that running SVE code in streaming mode will likely give bad performance and so SVE use should be limited until we know more.

OK. so your concern is with scoring debating points, not with either
- understanding hardware or
- understanding how to write OPTIMAL code for that hardware

Good to know, going forward.

It's minds like this that answer the question "Why doesn't Apple tell us that they have added SME (or SSVE, or ...) to the chip?"
Because they don't want to deal with the torrent of mindless BS that results from people thinking they have scored important gotcha's by parsing those words in the most legalistic (and useless) fashion imaginable.


Sure, for your legalistic concerns of writing an SSVE emulator, you care about elements of the spec. But can you not step outside that for just one minute to see how utterly uninteresting and irrelevant that is to everyone else concerned with these performance numbers?

 

[Doug S]

It makes many orders of magnitude more sense than Tim Cook paying more than necessary for anything. Cook’s control of supply chain costs is just as legendary as Job’s vision and hyperbole, Wozniak’s technical prowess, or Ive’s thin and minimalist design.

All speculation must pass through the hallowed gates of Cook’s renowned refusal to pay more than necessary. Unless there is hard evidence to the contrary, always assume the supplier is fully eating the costs of any and all discrepancies and inexplicables.

If someone got Cook to pay for something he wasn’t planning on using, he deserves a multi-million dollar bonus and a vice presidency.

The way DRAM suppliers have progressed from 16 Gb to 24 Gb to 32 Gb chips is by using improved processes that make denser chips. At some point it becomes economically inefficient to keep older production lines making lower density chips operating versus upgrading them to make higher density chips.

So if a DRAM supplier has a long term contract with Apple to supply 16 Gb DRAM chips for 8 GB packages but towards the end of that contract determines keeping a line open that's making 16 Gb chips will cost them money versus upgrading it to make 32 Gb chips they can sell for more than twice as much money. So they close the line and send Apple packages made using their now-smallest 24 Gb chips.

So why isn't Apple selling them as 12 GB iPads? Maybe that supplier is in the process of shutting down the 16 Gb lines, so couldn't/wouldn't guarantee they would all be 12 GB, and the last thing Apple wants is people buying/returning iPad Pros if they get one with 8 GB instead of 12 GB. Perhaps there is some rider in the contract that lets them deliver higher capacity packages but Apple can only use them as the capacity they paid for? i.e. an incentive to prevent customers from wanting very long contracts knowing they'll get "free stuff" in the last couple years.

 

[roger_k]

That was not the point Maynard contested in what I wrote. I only talked about SSVE. Not what SME or SME2 allows you to do. I think my guess was pretty clear about that.

I would say that the distinction between SME and SSVE (or between SME and SME2) are purely technical. What’s important to me specifically is that I have a bunch of instructions available in the streaming mode, and those variants that either accumulate to Z registers or to ZA slices. The instructions accumulating to Z registers are slow on Apple hardware.

That’s at least how I interpret what Maynard said.

Is this that code? Nice work!

Thanks!

but I know the differences between SME(2) instructions and SSVE

I understand where you are coming from, and as I write above, I don’t find this distinction to be particularly interesting. I have a bunch of different instructions at my disposal, I need to choose which ones to use. It’s not that important which part of the spec they come from. Personally, I refer to all of them as “streaming mode SVE”, simply because they are only available in the streaming mode.

 

[Eug]

Microsoft is advertising its new Copilot Plus machines are faster than the M3 MacBook Air.

Inside Microsoft’s mission to take down the MacBook Air

Microsoft might be finally ready to transition to Arm chips.

www.theverge.com

Microsoft representatives set out new Surface devices equipped with Qualcomm’s Snapdragon X Elite chips inside and compared them directly to Apple’s category-leading laptop. I witnessed an hour of demos and benchmarks that started with Geekbench and Cinebench comparisons, then moved on to apps and compatibility.

“You’re going to have the most powerful PC ever,” says Yusuf Mehdi, executive vice president and consumer chief marketing officer at Microsoft, during the briefing. “In fact, it’s going to outperform any device out there, including a MacBook Air with an M3 processor, by over 50 percent on sustained performance.”

This is probably old news to you guys, but for those who may not be aware, the top end 12-core Snapdragon X Elite gets around 15000 in Geekbench 6 which is even faster than M4, but single-core is less than 3000, and it uses way more power. M3 Pro is a bit faster multi-core though, at around 15900. Still, it's nice to see Qualcomm in the game, and I hope Microsoft sees some success in this space. It's always good to have some competition.

---

On another topic, I mentioned background services on Macs doing photo processing. This is on a slow Intel Mac so it's slower to complete but nonetheless it illustrates that photo workloads may be important to a lot of mainstream users, even when not actively running photo applications. I had previously launched Photos, but when I took the screengrab I had already exited the program.

 

[Eug]

Light on details, but I guess it's more a preview than a full launch.

 

[SpudLobby]

You know what those people should really complain about: Apple’s SSD and RAM price upgrades. Where’s the media/social media outrage about that?

Where’s the outrage on Apple’s stupid anti-repair tactics?

Those are evil, to put bluntly.

It’s costs Apple like $30 for a LPDRR5 module.

$200 for an 8GB RAM upgrade makes me puke.

He’s right tbh

 

[Nothingness]

I understand where you are coming from, and as I write above, I don’t find this distinction to be particularly interesting. I have a bunch of different instructions at my disposal, I need to choose which ones to use. It’s not that important which part of the spec they come from. Personally, I refer to all of them as “streaming mode SVE”, simply because they are only available in the streaming mode.

Understood On my side I was considering SSVE as a way to port existing SVE code or help with the creation of SVE code that could run on other CPUs that don't have SME; at this point SVE is, to say the least, not a success, but having dev use it on Apple could have helped demonstrate (or not) its usefulness.

 

[Eug]

Heh. Tom’s is reporting the iPad Pro M4 12 GB memory thing.

https://www.tomshardware.com/tablet...-12gb-teardowns-reveal-possible-apple-hijinks

 

[Nothingness]

A nice technical write up about SME on M4: https://github.com/tzakharko/m4-sme-exploration
Thanks to the writer

It's interesting to see the SME block has support for FP64 and can achieve about double the performance of 4 P cores running NEON FMLA.

 

[Mopetar]

Heh. Tom’s is reporting the iPad Pro M4 12 GB memory thing.

https://www.tomshardware.com/tablet...-12gb-teardowns-reveal-possible-apple-hijinks

Now that we've got a little blood in the water, let's see how fast the sharks start to circle.

 

[junjie1475]

【苹果M4性能分析：尽力了，但人类科技快到头了！】 https://www.bilibili.com/video/BV1N...eb&vd_source=a38447d92dbac66e202c0251155453d6