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

 

[poke01]

View attachment 107810
M4 has the same GPU architecture as A18 with improved RT. However, Apple kept silent about this at the M4 iPad Pro event.

Is this the solar bay test?

 

[jdubs03]

View attachment 107811

This year, Apple has done a very good job of improving power efficiency.

Gotta be LN2 for that score. That's ridiculously high.

 

[mvprod123]

Is this the solar bay test?

It's a kind of internal test.

 

[Doug S]

This year, Apple has done a very good job of improving power efficiency.

I'm still of the opinion that the problematic N3B process was the reason for A17P's large increase in power draw. Apple was paying for "known good dies" and TSMC was probably insisting on cranking up the power a bit to include some dies Apple normally tosses (or saves to use in less power sensitive products like Apple TV) to deliver enough dies to meet Apple's needs.

N3E fixed the yield issues, so Apple was able to do their normal binning.

 

[jdubs03]

View attachment 107811

This year, Apple has done a very good job of improving power efficiency.

One thing I noticed about this chart, and then the one below is the A17 Pro values are different. I'm assuming just an accidental mix up.



Odd too that the spec2017 results don't go to the max frequency for the A17 Pro in his A18/P review. But in the previous videos he shows the same spec2017 int and fp values for the A17 Pro, but the correct max frequency. I'm assuming another accidental value there. The reason I mention it, is because he points out that the A18P sees a 4.5% IPC increase (integer), but in that other video we see 6.2%. And when the normal 3.78GHz value is used instead of 3.72GHz, the IPC increase shows as 6.1%.



What gives? 😵‍💫

 

[poke01]

One thing I noticed about this chart, and then the one below is the A17 Pro values are different. I'm assuming just an accidental mix up.

View attachment 107819

Odd too that the spec2017 results don't go to the max frequency for the A17 Pro in his A18/P review. But in the previous videos he shows the same spec2017 int and fp values for the A17 Pro, but the correct max frequency. I'm assuming another accidental value there. The reason I mention it, is because he points out that the A18P sees a 4.5% IPC increase (integer), but in that other video we see 6.2%. And when the normal 3.78GHz value is used instead of 3.72GHz, the IPC increase shows as 6.1%.

View attachment 107822
View attachment 107823
What gives? 😵‍💫

The frequency in the A18 Pro for the A17 Pro might be typo since it’s the exact same results as the A17 Pro review. A lot of typos, I guess it was rushed.

Hopefully for the 8G4 review we get a less error prone graphs.

 

[poke01]

Much faster H.265 encoders in A18 Pro. Bodes well for the M4 Pro/Max chips.

 

[FlameTail]

View attachment 107810
M4 has the same GPU architecture as A18 with improved RT. However, Apple kept silent about this at the M4 iPad Pro event.

Hmm. Interesting.

I'm still of the opinion that the problematic N3B process was the reason for A17P's large increase in power draw. Apple was paying for "known good dies" and TSMC was probably insisting on cranking up the power a bit to include some dies Apple normally tosses (or saves to use in less power sensitive products like Apple TV) to deliver enough dies to meet Apple's needs.

N3E fixed the yield issues, so Apple was able to do their normal binning.

That bodes well for upcoming Android SoCs.


Kinda incredible how iOS18 improves battery life so much.

But the Geekbench scores are also lowered.

 

[FlameTail]

When it comes to A18 Pro, Apple is still outspending Android SoC vendors in terms of cache capacity. But the A18's cache capacity is actually similar to current Android flagship CPUs.

Also something that the Geekerwan review revealed is that the reduced cache doesn't make a huge difference to the performance or efficiency. But there is a difference, yes;

 

[poke01]

GPU performance is improved by quite a bit

 

[FlameTail]

So A18 also uses the same P-core uarch, E-core uarch and GPU uarch as M4?

 

[Doug S]

Odd too that the spec2017 results don't go to the max frequency for the A17 Pro in his A18/P review. But in the previous videos he shows the same spec2017 int and fp values for the A17 Pro, but the correct max frequency. I'm assuming another accidental value there. The reason I mention it, is because he points out that the A18P sees a 4.5% IPC increase (integer), but in that other video we see 6.2%. And when the normal 3.78GHz value is used instead of 3.72GHz, the IPC increase shows as 6.1%.



What gives? 😵‍💫

I always ignore the listed frequencies because benchmarks always seem to have a hard time determining the frequency. They aren't actually running at "less" frequency, its just the frequency being calculated/reported that's different.

What stood out to me was the slightly higher power draw of the non-Pro A18s. I wonder if that's just due to the smaller cache (i.e. working the memory controller a little more) or if they're the same die and Apple binned on power draw, using ones that could operate on a bit lower voltage in the Pro models.

Looking forward to teardowns so we can answer this "is it one die or two" question...

 

[poke01]

So A18 also uses the same P-core uarch, E-core uarch and GPU uarch as M4?

Yes. I think we will know more in the Mac event.

 

[jdubs03]

Hmm. Interesting.

That bodes well for upcoming Android SoCs.

View attachment 107835
Kinda incredible how iOS18 improves battery life so much.
View attachment 107836
But the Geekbench scores are also lowered.

I noticed on one of the Apple OS youTubers I follow that geekbench scores were a bit lower in iOS 18 compared to iOS 17, particularly after iOS 18.0 beta 2-3. So when considering that, those 8100-8200 scores make much more sense, which means ones at 8400+ must just be super good bins and/or ideal cooling conditions.

One thing I’ve personally noticed for my 14 Pro is that in Safari and Firefox my Kraken 1.1 time was faster by about 13% and my Speedometer 3 score was higher by about 6%.

 

[jdubs03]

Figured I’d post this since it’s newsy. A19/P on TSMC N3P. A20P on N2.

Kuo: iPhone 17 to Use 3nm Chip Tech, iPhone 18 Pro Models to Use 2nm

Next year's iPhone 17 series will feature processors made using TSMC's enhanced N3P 3-nanometer chip technology, but only iPhone 18 Pro...

www.macrumors.com

Should be a good opportunity to upgrade the architecture for the Ecores. It’ll probably be an iteration on the Pcore.


Looks like the release cadence for Apple Silicon will be staggered a year behind HVM.

 

[Eug]

Apple poised to introduce self-developed 5G modem in iPhones by 2025

Contrary to its usual strategy, Apple appears to be breaking tradition by introducing new technology in the iPhone SE first. This decision aligns with the relatively low cost of 5G modems and the smaller shipment scale of the iPhone SE.

As Apple's in-house 5G modem has yet to incorporate mmWave technology, the company will still rely on Qualcomm's technical support. The key question is whether Apple can achieve a breakthrough in mmWave technology before its modem licensing agreement with Qualcomm expires in 2027.

Insiders from Apple's supply chain suggest that the company may introduce its in-house Wi-Fi chip in new iPads in 2025. Alternatively, Apple might choose to debut it in certain iPhone 18 models in 2026.

 

[Doug S]

I noticed on one of the Apple OS youTubers I follow that geekbench scores were a bit lower in iOS 18 compared to iOS 17, particularly after iOS 18.0 beta 2-3. So when considering that, those 8100-8200 scores make much more sense, which means ones at 8400+ must just be super good bins and/or ideal cooling conditions.

One thing I’ve personally noticed for my 14 Pro is that in Safari and Firefox my Kraken 1.1 time was faster by about 13% and my Speedometer 3 score was higher by about 6%.

No real reason for iOS 18 to have lower scores other than handling power management or temperature targets more conservatively. Since Geekbench is a peaky load if they changed how quickly it ramps up clock speed or promotes a load from E to P cores that might account for it. If it is temperature target related then you'd see it in all loads.

If this turns out to be a real thing and not just some weird measurement artifact for that one guy, I hope Geekerwan picks up on it and does testing between iOS 17 and 18 to figure out what changed.

 

[Doug S]

Apple poised to introduce self-developed 5G modem in iPhones by 2025

Contrary to its usual strategy, Apple appears to be breaking tradition by introducing new technology in the iPhone SE first. This decision aligns with the relatively low cost of 5G modems and the smaller shipment scale of the iPhone SE.

As Apple's in-house 5G modem has yet to incorporate mmWave technology, the company will still rely on Qualcomm's technical support. The key question is whether Apple can achieve a breakthrough in mmWave technology before its modem licensing agreement with Qualcomm expires in 2027.

Insiders from Apple's supply chain suggest that the company may introduce its in-house Wi-Fi chip in new iPads in 2025. Alternatively, Apple might choose to debut it in certain iPhone 18 models in 2026.

I doubt they need a "breakthrough" to implement mmwave. They just haven't had a reason to prioritize it, given that they plan to roll it out in the SE first and the current SE doesn't support mmwave, even though the iPhones introduced before it did - probably left it out for both cost and product segmentation reasons.

It is also pretty sparsely deployed, and always will be, so the lack of it shouldn't be a reason anyone would reject buying the new SE.

 

[The Hardcard]

No real reason for iOS 18 to have lower scores other than handling power management or temperature targets more conservatively. Since Geekbench is a peaky load if they changed how quickly it ramps up clock speed or promotes a load from E to P cores that might account for it. If it is temperature target related then you'd see it in all loads.

If this turns out to be a real thing and not just some weird measurement artifact for that one guy, I hope Geekerwan picks up on it and does testing between iOS 17 and 18 to figure out what changed.

He did identify the change in the delay before ramping to full frequency on the P cores in iOS 18. He spends a couple of minutes talking about it before the conclusions closeout of the review. He lists graphs of the frequency ramping, but unless I’m missing something, there are no quantifying numbers on the charts or in his audio. I wanted to see an expression in milliseconds.

He did note that it particularly affects Geekbench subtests that last less than one second.

 

[FlameTail]

If this turns out to be a real thing and not just some weird measurement artifact for that one guy, I hope Geekerwan picks up on it and does testing between iOS 17 and 18 to figure out what changed.

Geekewan mentioned it in his video.

 

[poke01]

If this boosting algorithm is applied to macOS 15 then Geekbench scores will be lower their too.

So it looks Cinebench is the best way to see ST performance now on the Mac.

 

[jdubs03]

If this boosting algorithm is applied to macOS 15 then Geekbench scores will be lower their too.

So it looks Cinebench is the best way to see ST performance now on the Mac.

Perhaps. My thinking is since it’s not a mobile device where sipping power is paramount (compared to a laptop) it probably doesn’t matter as much to them to implement the same delay. Will be interesting to see how that plays out.

 

[name99]

Hmm. Interesting.

That bodes well for upcoming Android SoCs.

View attachment 107835
Kinda incredible how iOS18 improves battery life so much.
View attachment 107836
But the Geekbench scores are also lowered.

This is well understood. With iOS18 Apple slightly lengthened how long the CPU runs at lower frequencie before it climbs to higher frequencies. Very good for battery life, bad for benchmarks that only take a second or so...

The basic point is that, IF YOU CARE PRIMARILY ABOUT THE USER EXPERIENCE not about winning benchmarks, you're going to keep diddling this curve every year, because "time spent at lower frequencies" should basically be tuned to human response time, which means it's going to grow as the CPU becomes faster (can do more at lower GHz). My guess is that someone at Apple realized this, after a few years of never bothering to tweak the DVFS curve, but maybe now it will be redone with each new CPU?

Also means that if the goal of your benchmark is to test peak performance, you need to do something like run a second or two of warmup code before kicking off the real, timed, code.

 

[Nothingness]

So it looks Cinebench is the best way to see ST performance now on the Mac.

You miswrote SPEC 😀

 

[Doug S]

Also means that if the goal of your benchmark is to test peak performance, you need to do something like run a second or two of warmup code before kicking off the real, timed, code.

Then the objection would be that the warmup code is hurting CPUs that are more aggressive about ramping up clock, because the "warmup" code would do as its name said and make the CPU more likely to throttle later in the benchmark run.

There isn't any way I can see to do this sort of thing and be "fair" to all. All we can do is be aware of the ever growing list of asterisks when comparing different platforms. It used to be so easy in the days when you were mostly testing single core systems that always ran at the same frequency and their only method of "power saving" was the off switch

But hey, now people complaining about Apple's SME boost in GB 6.3 can stop complaining, as they've handed back that boost with iOS 18!