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

 

[smalM]

I always thought the Pro/Max would go chiplett and it would look like this:
"i/o die" - E-cluster, NPU, USB, TB, en-/decoder, display controller etc.
"compute die" - MI, SLC, P-cluster, GPU-cluster
Pro 1x i/o + 1x compute
Max 1x i/o + 2x compute
Ultra 1x i/o + 4x compute + 1x i/o

And along came the M3 Pro and the M3 Max didn't include UltraFusion.
So much for my ability to predict anything Apple....

 

[FlameTail]

I always thought the Pro/Max would go chiplett and it would look like this:
"i/o die" - E-cluster, NPU, USB, TB, en-/decoder, display controller etc.
"compute die" - MI, SLC, P-cluster, GPU-cluster
Pro 1x i/o + 1x compute
Max 1x i/o + 2x compute
Ultra 1x i/o + 4x compute + 1x i/o

There's power and latency penalties when using so many tiles. That's why Intel reduced the number of tiles in Lunar Lake to two.

 

[Doug S]

I thought about chop, but the separate design for the M3 Pro makes me think otherwise. I don’t see how they could include CPU resources in the chop and I also don’t see how the Pro and Max go back to having the same CPU. 12 P cores is way too many for the Pro, yet how could they give the Max any less?

Who says they have to chop it in the same way as M1? The "chop" portion doesn't have to just include more GPU cores, it could include another P core cluster, etc.

 

[Doug S]

There's power and latency penalties when using so many tiles. That's why Intel reduced the number of tiles in Lunar Lake to two.

I think a lot of that is due to Intel's ring bus, since communication between chiplets will often have to pass through other chiplets. Having eight chiplets on something like that is gonna be a problem. If all dies were directly connected (like Apple's patent for a four chip Ultra like solution) or everything passed through an intermediate switch die (maybe that's what AMD's IOD is but I'm not sure) the power and latency issues are reduced.

 

[mvprod123]

Binned for GPU cores and I wonder if it will also be binned for clock speed.

iPad mini (A17 Pro)

iPad16,2 - Geekbench

Benchmark results for an iPad16,2 with an ARM processor.

browser.geekbench.com

iPad16,2 - Geekbench

Benchmark results for an iPad16,2 with an ARM processor.

browser.geekbench.com

 

[Eug]

iPad mini (A17 Pro)

iPad16,2 - Geekbench

Benchmark results for an iPad16,2 with an ARM processor.

browser.geekbench.com

iPad16,2 - Geekbench

Benchmark results for an iPad16,2 with an ARM processor.

browser.geekbench.com

View attachment 109642
View attachment 109643

Clock speed is unchanged. The multi-core CPU score is somewhat low but that could just be due to random variation. We shall see once more scores come in.

As expected, the Metal score is lower than average, but again we shall see if those low scores are consistent.

For reference, this is what A17 Pro can do under ideal conditions:

iPhone 15 Pro Max - Geekbench

Benchmark results for an iPhone 15 Pro Max with an Apple A17 Pro processor.

browser.geekbench.com

And in the freezer?

iPhone16,1 - Geekbench

Benchmark results for an iPhone16,1 with an ARM processor.

browser.geekbench.com

 

[mvprod123]

Clock speed is unchanged. The multi-core CPU score is somewhat low but that could just be due to random variation. We shall see once more scores come in.

As expected, the Metal score is lower than average, but again we shall see if those low scores are consistent.

For reference, this is what A17 Pro can do under ideal conditions:

iPhone 15 Pro Max - Geekbench

Benchmark results for an iPhone 15 Pro Max with an Apple A17 Pro processor.

browser.geekbench.com

And in the freezer?

iPhone16,1 - Geekbench

Benchmark results for an iPhone16,1 with an ARM processor.

browser.geekbench.com

Interestingly, the scores were obtained on version 18.0.

 

[igor_kavinski]

Bloomberg - Are you a robot?

In fact, some at Apple believe that its generative AI technology — at least, so far — is more than two years behind the industry leaders.

 

[okoroezenwa]

Bloomberg - Are you a robot?

Nice.

 

[mikegg]

Density will jump again with BSPDN. And we still have plenty of tricks on hand for increasing IPC.

People have been saying Apple hit peak performance since about the A8. Like most extrapolations based purely on drawing graphs, as opposed to understanding the drivers of change, these predictions have not aged well.

A16 is the first TSMC node with backside delivery and the density improvement is actually even worse than N3 to N2.

 

[FlameTail]

A16 is the first TSMC node with backside delivery and the density improvement is actually even worse than N3 to N2.

View attachment 109847

A16 is not a new node, iirc. It's basically N2 but with BSPDN added to it. So with that in mind, density scaling of A16 seems actually good.

 

[name99]

A16 is the first TSMC node with backside delivery and the density improvement is actually even worse than N3 to N2.

View attachment 109847

*SRAM* density
Not generic logic density.

And, as FlameTail points out, that's a bump you get with (as far as we know) no lithography improvements, JUST as a result of BSPDN.

 

[Doug S]

A16 is not a new node, iirc. It's basically N2 but with BSPDN added to it. So with that in mind, density scaling of A16 seems actually good.

Yep, listing the same mass production time as N2P should be the clue.

TSMC originally said that they would do two versions of N2, the standard N2 then six months later "N2+BSPDN" would began mass production. At some point they changed that plan and pushed it back six months to be concurrent with N2P, but unlike the (apparent) original plan they've given it a different name to better differentiate it.

 

[mvprod123]

Finally

https://twitter.com/x/status/1848778814371028993

 

[igor_kavinski]

So looking like it's gonna be Oct 30.

So my wishlist (which isn't getting fulfilled until after I'm gone ):

First ever Apple made desktop mobo
First ever Apple pluggable/upgradable CPU
First ever Apple authorized and certified DDR5-10,000 dual channel dual slot CUDIMMs
First ever Apple authorized and certified heatsink/fan
Compatibility with Windows ARM including day one drivers.

Combo deal starting price $1500.

 

[Mopetar]

I don't think any of those will be happening. Apple doesn't sell CPUs so they have no reason to make a socket or their own motherboard. They typically don't go for the fastest or most exotic memory configurations either. Having to produce at volume means needing components at volume as well.

Windows ARM compatibility isn't solely on Apple either. If you wanted to run Windows on ARM why buy a Mac when there are other vendors/solutions. Similarly, why buy a Mac just to run Windows? Some part of the reason of buying a Mac is so you don't have to run Windows. If you absolutely need to, virtualization is always an option.

 

[poke01]

I don't think any of those will be happening. Apple doesn't sell CPUs so they have no reason to make a socket or their own motherboard. They typically don't go for the fastest or most exotic memory configurations either. Having to produce at volume means needing components at volume as well.

Windows ARM compatibility isn't solely on Apple either. If you wanted to run Windows on ARM why buy a Mac when there are other vendors/solutions. Similarly, why buy a Mac just to run Windows? Some part of the reason of buying a Mac is so you don't have to run Windows. If you absolutely need to, virtualization is always an option.

I think Igor is being sarcastic, he knows it it’s not happening lol

 

[Eug]

First ever Apple pluggable/upgradable CPU

Sarcastic or not, FWIW, I swapped out the Xeons in my old Mac Pro, from 2 x dual-core to 2 x quad-core. Interestingly, the Xeons I used had never shipped in a Mac ever, but worked fine nonetheless.

Originals were 2 x 2.66 GHz dual-core Xeon 5150. New ones were 2 x quad-core 2.33 GHz Xeon 5345. Cost me US$20 for the pair. (I was bored during the COVID lockdown.)

I also swapped out the CPU from my G4 450 MHz Cube to a G4 1.7 GHz, but that required replacement of the entire CPU daughterboard with one made by Sonnet.

---

The number 1 thing on my wish list is a new entry level iPad, with 8 GB RAM and 128 GB storage hopefully. However, rumours have it coming some time in 2025, perhaps Q2-Q4. I guess my kid will just have to continue using her iPad Air 2 for up to another year. Luckily the battery is still usable, despite that machine being purchased 9 years ago, in 2015. (The battery in our other iPad Air 2 purchased in 2014 was totally shot several years ago, but that iPad Air 2 saw a lot more use.)

8 GB in an entry level iPad would represent a doubling of RAM, but I'm optimistic for that since Apple seems to be spec'ing a minimum of 8 GB RAM for AI. I don't really care much at this point about AI on an entry level iPad, but I do care about having more than 4 GB RAM. I'm less optimistic about 128 GB storage, but since the Air and mini have gone 128 GB in 2024, and the entry level iPad often has as much storage as the entry level Air and mini, I am still hopeful.

The new Mac minis will be interesting, but truthfully I have no actual need to upgrade from my M1 which is still running splendidly.

 

[igor_kavinski]

Similarly, why buy a Mac just to run Windows?

To get the Mac SoC power efficient experience without struggling with the awful Mac UI.

 

[igor_kavinski]

but truthfully I have no actual need to upgrade from my M1 which is still running splendidly.

I know. There really isn't much wrong with that SoC, other than price (should be put into the public domain so everyone's base experience becomes M1) and being in the WRONG company's computer

 

[The Hardcard]

To get the Mac SoC power efficient experience without struggling with the awful Mac UI.

You still have never said why Snapdragon X Elite doesn’t answer your call. The upcoming version sounds like everything you are asking for, let alone Lunar Lake.

Then there’s Mediatek coming with yet another innovative large cache top-level out of order powered SOC with Nvidia graphics. It seems like a great time to not worry about Apple Silicon since you don’t like the business model, the hardware or the OS. Why aren’t you swooning to the powerful, power efficient sweet siren song of Snapdragon.

Oh how you hurt the feelings of Gerard Williams and John Bruno. All the strain and sweat to lead the development of performant, efficient chips that run Windows. Then they come onto Anandtech forums to relax and unwind only to see you forsake their efforts and continue to hope for Apple to completely upend its multi trillion dollar business model so you can get M chips under different conditions.

The day after they present what they must feel are tremendous achievements, no less. Your posts here today must be like shivs in their hearts.

 

[poke01]

I feel like Mac UI is good for laptops, especially with the trackpad. On desktops with multiple monitors it does struggle.

 

[pj-]

The thing with macs and macos is it's the best computing platform in the world if you grew up with macs or if your usage happens to align exactly to how apple thinks you should use a computer and you operate entirely within their ecosystem.

If you deviate from that even an inch, the options are one of:

• you used to be able to change that with a terminal command but they removed it in 10.whatever
• you can do that, there's a $14.99 app, and it has a fun name referencing food that has no relation to its function
• no you can't do that and you should feel bad for trying

 

[poke01]

you can do that, there's a $14.99 app, and it has a fun name referencing food that has no relation to its function

There are open source alternatives. Example is Magnet and Rectangle. Magnet is paid and Rectangle is open source and free. They both do the same thing.

It’s just YouTubers recommend the paid versions most of the time.

you operate entirely within their ecosystem.

ehh. I use an Android phone and a MacBook plus a Windows PC. It’s totally doable. There obviously use cases Macs cannot do a well as PCs, gaming is one and other niche uses cases like MST support.

 

[pj-]

There are open source alternatives. Example is Magnet and Rectangle. Magnet is paid and Rectangle is open source and free. They both do the same thing.

It’s just YouTubers recommend the paid versions most of the time.

ehh. I use an Android phone and a MacBook plus a Windows PC. It’s totally doable. There obviously use cases Macs cannot do a well as PCs, gaming is one and other niche uses cases like MST support.

Not gonna hate on the assumption that a mac user would solely rely on youtube for software recommendations, but that is not me. I'm a sw engineer by education and trade, I use windows and linux professionally, and mac/windows/linux/ipados/android personally. I know how to find software. Also I feel like I should get some benefit of the doubt for being a forum user of any sort in the year 2024, let alone anand.

I was exaggerating a bit but getting my m3 mbp working as I would expect with two non-"hidpi" monitors (mirroring dummy higher resolution displays to trick the OS into doing font smoothing) and using non-apple peripherals (3rd party apps to override default mouse acceleration, scroll wheel behavior) was a major pain.