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]

Never I my wildest dreams did I imagine that the iPhone could run AAA that are current at any capacity.

Really looking forward to AC: Mirage and RE:4 Remake on my phone. I don't care if it's 720p 30fps, just the thought it's not the usual game mobile crap makes me happy.

Cyberpunk 2077 2.0 might actually be playable in 2-3 years time with even better GPUs on the iPhone.

 

[GC2:CS]

One thing we noted in the Arm thread is that Geekbench 6.1 & 6.2 scores are not comparable with Geekbench 6.0 scores.

OK, looking at the iPhone 15 Pro (A17 Pro) and using the top scores in the first two pages of results (to make sure it's GB6.1 / 6.2)

Apple A14: 2201 @ 2.99 GHz (736 pts / GHz - 100.0%)
Apple A15: 2387 @ 3.23 GHz (739 pts / GHz - 100.4%)
Apple A16: 2683 @ 3.46 GHz (775 pts / GHz - 105.3%)
Apple A17P: 2914 @ 3.77 GHz (773 pts / GHz - 105.0%)

So literally 0% IPC uplift / category A. So the 🏆 goes to @poke01. For now three generations, Apple has barely increased IPC by 5%. Efficiency: that would be interesting to test, but basically no one does serious testing on smartphone CPU power consumption. The A15, while basically no IPC, actually improved power consumption, even with notably higher clocks (5.5% lower avg power to 4.11W, 16.5% less energy, and 13-15% faster performance)

//

So even with improved branch prediction, wider decode, and wider execution: 0% IPC? On what programs / tests can you change those in a core and get zero IPC uplift? Unless clocks are wrong (which they rarely are), the CPU uArch seems like a clone of the past three uArches or those improvements are making up for some other deficiency not yet mentioned.

How did you choose those score numbers ?

 

[Nothingness]

How did you choose those score numbers ?

That's a good question indeed. Many of these numbers are significantly higher than what is there: https://browser.geekbench.com/ios-benchmarks

If I pick scores from the Web page and add some iPhone 15 I get this:


To be taken with a huge grain of salt due to various roundings (frequency for instance).

 

[ikjadoon]

How did you choose those score numbers ?

Do you mean more this? Just a simple check.

One thing we noted in the Arm thread is that Geekbench 6.1 & 6.2 scores are not comparable with Geekbench 6.0 scores.

OK, looking at the iPhone 15 Pro (A17 Pro) and using the top scores in the first two pages of results (to make sure it's GB6.1 / 6.2)

Apple A14: 2201 @ 2.99 GHz (736 pts / GHz - 100.0%)
Apple A15: 2387 @ 3.23 GHz (739 pts / GHz - 100.4%)
Apple A16: 2683 @ 3.46 GHz (775 pts / GHz - 105.3%)
Apple A17P: 2914 @ 3.77 GHz (773 pts / GHz - 105.0%)

That's a good question indeed. Many of these numbers are significantly higher than what is there: https://browser.geekbench.com/ios-benchmarks

If I pick scores from the Web page and add some iPhone 15 I get this:
View attachment 85869

To be taken with a huge grain of salt due to various roundings (frequency for instance).

Geekbench’s “official lists” are very messy, public averages and shouldn’t be relied on, IMO. Different battery percentages, different ambient temps, different battery health, etc. Apple iPhone SoCs do not boost, so Geekbench’s official list averages will always and only lower the peak / actual score.

So the actual SoC performance, it's just the highest Geekbench score listed (hopefully without active cooling, but these SoCs rarely thermal throttle on short 1T tests).

Ideally, we’d have reviewers’ numbers, but they will all be GB6.1/6.2, so you can’t compare with iPhone 12 / 13 / 14 in their original reviews. This is just a quick check, so just a bit more accurate than using the official lists. Ideally, one reviewer will re-test all devices on GB6.1/6.2 with temp / battery controls.

 

[thunng8]

That's a good question indeed. Many of these numbers are significantly higher than what is there: https://browser.geekbench.com/ios-benchmarks

If I pick scores from the Web page and add some iPhone 15 I get this:
View attachment 85869

To be taken with a huge grain of salt due to various roundings (frequency for instance).

A few substantially higher numbers have now been posted for the a17. Generally the first public leaks are lower because they are new phones that might be running a lot of background tasks as reviewers load up all their applications.

eg

iPhone16,1 - Geekbench

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

browser.geekbench.com

So close to 3000 for Geekbench sc.

 

[Doug S]

To get the actual SoC performance, you need to find the highest Geekbench score possible (hopefully without active cooling, but these SoCs rarely throttle on 1T testing due to heat).

What kind of active cooling could you do with an iPhone? Run the benchmark while standing in a walk in freezer?

 

[ikjadoon]

What kind of active cooling could you do with an iPhone? Run the benchmark while standing in a walk in freezer?

Sometimes, literally yes, haha, but more typically with a 140mm fan. It's how Andrei tested smartphone SoCs while at AnandTech.

When tracking the average frequencies under SPEC, benchmarking the Exynos 2100 S21 Ultra under my typical peak performance conditions where I place the phone over a 140mm fan to keep it cool, the X1 cores were still throttling quite significantly even though the phone was only luke-warm.

The following are precise mean frequencies for the SPEC workloads, both under my usual fan-cooled conditions, as well as putting the S21 Ultra in my freezer:

While it's not noted in each review, Andrei confirmed (IIRC in reddit comment that I can't find atm) that all his SPEC smartphone SoC testing used active cooling, but if he used the freezer, it will be noted. SPEC does take a few hours on a smartphone, too, so it might make sense with those long 1T tests.

But, Geekbench, hopefully, is so quick that ideally there's no heat throttling.

 

[poke01]

The A17 is a pretty good chip

 

[Doug S]

Sometimes, literally yes, haha, but more typically with a 140mm fan. It's how Andrei tested smartphone SoCs while at AnandTech.



While it's not noted in each review, Andrei confirmed (IIRC in reddit comment that I can't find atm) that all his SPEC smartphone SoC testing used active cooling, but if he used the freezer, it will be noted. SPEC does take a few hours on a smartphone, too, so it might make sense with those long 1T tests.

But, Geekbench, hopefully, is so quick that ideally there's no heat throttling.

Running full SPEC on a smartphone has always been a bit of a silly exercise, since no one uses a phone to perform long running calculations that load the CPU at 100% for hours. It could only tell you how fast its CPU would be in a PC form factor.

That made it quite useful running SPEC on iPhones annually though, as it provided a point of comparison to see how much Apple's SoCs were closing the gap with x86 to year to year back when everyone was wondering when Apple would be ready to announce the worst kept secret in the world.

 

[Mopetar]

Running full SPEC on a smartphone has always been a bit of a silly exercise, since no one uses a phone to perform long running calculations that load the CPU at 100% for hours. It could only tell you how fast its CPU would be in a PC form factor.

It's hardly silly. SPEC tests are chosen to be representative of a variety of different workloads. They do an excellent job of pointing out how well a chip performs in a variety of situations and are varied enough to show strengths and weaknesses in specific areas rather than reducing the score to a single number.

Phone users may not run anything for multiple hours outside of videos or games, but a CPU will have the same performance capability whether the task is minutes or microseconds long. The added benefit of a longer run is that it eliminates boost or attempted gaming of the results by trying to run at a higher frequency for a short duration.

 

[eek2121]

Never I my wildest dreams did I imagine that the iPhone could run AAA that are current at any capacity.

Really looking forward to AC: Mirage and RE:4 Remake on my phone. I don't care if it's 720p 30fps, just the thought it's not the usual game mobile crap makes me happy.

Cyberpunk 2077 2.0 might actually be playable in 2-3 years time with even better GPUs on the iPhone.

The iPhone's GPU is pretty amazing. Sometimes PC ports come along (like Dota Underlords) and I am surprised that the graphics are actually as high of quality as they are.

Kids are more likely to play silly games which most mobile games are.

Serious games like Death Stranding and RE4 Remake running on Apple silicon signal a shift in Apple's gaming strategy. Optimization is key, though. They absolutely need every game to run at minimum 30fps on the base M1 hardware, otherwise there will never be a market big enough for the publishers to care much about.

There are quite a few 'non-mobile' games on iOS. Some are GPU demanding, some aren't.

A few of the games ported from PC:

• Dota Underlords
• Stardew Valley
• Bloons TD 6
• Among Us
• A bunch of Squaresoft games
• Minecraft
• Civilization 6
• GTA: San Andreas
• Terraria

If I were Microsoft and Sony, I'd be a bit concerned about Apple and gaming. If Apple starts to attract bigger games to their platforms, it could pull market share away from the consoles.

Apple should release an updated Apple TV with an M2 chip (preferably a Max or Ultra variant). If they did this, they would have a "console" with more power than current gen consoles.

 

[Lodix]

Sometimes, literally yes, haha, but more typically with a 140mm fan. It's how Andrei tested smartphone SoCs while at AnandTech.



While it's not noted in each review, Andrei confirmed (IIRC in reddit comment that I can't find atm) that all his SPEC smartphone SoC testing used active cooling, but if he used the freezer, it will be noted. SPEC does take a few hours on a smartphone, too, so it might make sense with those long 1T tests.

But, Geekbench, hopefully, is so quick that ideally there's no heat throttling.

There is. That's why there is such a discrepancy between scores for smartphones in Geekbench.

 

[mikegg]

A few substantially higher numbers have now been posted for the a17. Generally the first public leaks are lower because they are new phones that might be running a lot of background tasks as reviewers load up all their applications.

eg

iPhone16,1 - Geekbench

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

browser.geekbench.com

So close to 3000 for Geekbench sc.

3000? If this is real, then it's ridiculous for a phone to have faster SC than every CPU out there except for the Intel Core i9-13900KS.

If it's 3000, then M3 should be around 3,400.

 

[Doug S]

It's hardly silly. SPEC tests are chosen to be representative of a variety of different workloads. They do an excellent job of pointing out how well a chip performs in a variety of situations and are varied enough to show strengths and weaknesses in specific areas rather than reducing the score to a single number.

Phone users may not run anything for multiple hours outside of videos or games, but a CPU will have the same performance capability whether the task is minutes or microseconds long. The added benefit of a longer run is that it eliminates boost or attempted gaming of the results by trying to run at a higher frequency for a short duration.

I think SPEC benchmarks are great, when used as intended. They are were designed to benchmark workstations, running the sorts of stuff people buy workstations to run. Running them on a phone - then twisting the results into uselessness by actively cooling the phone - is silly. It doesn't represent what the phone is capable of, at best it represents what the SoC in the phone would be capable of if it was installed in a PC with active cooling.

I'm not aware of any phones that have a turbo mode, so who is gaming results by boosting the clock? If they did, actively cooling the phone would defeat the purpose by helping them run at a higher clock than they are capable of in real world use. If you want SPEC results to be useful in the way you say, you'd want the phones benchmarked with any 140mm fan or freezer helping. Stick it in a 20C room with no moving air on top of 2" of styrofoam, fully charged but not connected to the charger, let it sit for an hour to reach ambient, then kick off SPEC.

 

[Eug]

I'm not aware of any phones that have a turbo mode, so who is gaming results by boosting the clock?

Samsung and others were caught throttling applications… except for benchmarks. So, effectively benchmarks were getting Turbo mode whereas other applications could not make use of that speed.

In fact, Geekbench banned Samsung and OnePlus phones for this reason.

 

[Mopetar]

I think SPEC benchmarks are great, when used as intended. They are were designed to benchmark workstations, running the sorts of stuff people buy workstations to run. Running them on a phone - then twisting the results into uselessness by actively cooling the phone - is silly.

I think you misunderstand what SPEC tests. It's just a series of small benchmarks that try to isolate various performance characteristics that may be representative of some specialized workloads. For example the 525.x264 benchmark tests video compression. Another is essentially just running Blender which is obviously a 3D rendering test, but if you have anything that behaves similarly to that, it's a useful test.

I'm not aware of any phones that have a turbo mode, so who is gaming results by boosting the clock? If they did, actively cooling the phone would defeat the purpose by helping them run at a higher clock than they are capable of in real world use.

Companies cheating on benchmarks has always been a problem. AT had had some articles covering it. Here's a recent one from 2020. There was a similar scandal almost a decade ago where a large number of companies were found to be cheating on benchmarks by detecting when one was being run and changing thermal limits or other performance settings to get better results.

Really the only downside of the SPEC suite for mobile devices is that some of the benchmarks are written in FORTRAN and can't be compiled for iOS (and probably Android) so they can't be used.

 

[eek2121]

3000? If this is real, then it's ridiculous for a phone to have faster SC than every CPU out there except for the Intel Core i9-13900KS.

If it's 3000, then M3 should be around 3,400.

It definitely isn’t overkill if there are apps that can use it. Also, the faster the core is (assuming no perf/watt regression), the faster a task can be completed and the cores can go back to sleep.

Even the GPU gains can seem absurd, but again…

 

[FlameTail]

So does Apple A17 use ARMv9 now?

 

[Doug S]

Samsung and others were caught throttling applications… except for benchmarks. So, effectively benchmarks were getting Turbo mode whereas other applications could not make use of that speed.

In fact, Geekbench banned Samsung and OnePlus phones for this reason.

Providing active cooling during benchmark runs would only serve to further incentivize such cheating.

 

[Mopetar]

So does Apple A17 use ARMv9 now?

That's what I was lead to believe. I've seen other sources posting that they've moved to ARMv9, but I've not seen anything official to corroborate that. I'm assuming it's on a data sheet that someone could publish in a review.

Providing active cooling during benchmark runs would only serve to further incentivize such cheating.

What are you going on about? I can't think of any reviewers that benchmark mobile devices with any kind of active cooling or in any way other than the device as it ships. Maybe some have done it as part of a case review, but that's a separate matter than a review of a phone or the SoC in that phone.

 

[Eug]

What are you going on about? I can't think of any reviewers that benchmark mobile devices with any kind of active cooling or in any way other than the device as it ships. Maybe some have done it as part of a case review, but that's a separate matter than a review of a phone or the SoC in that phone.

Just to nitpick: Occasionally you'll see reviewers that benchmark the phones while they're sitting inside a freezer. I haven't seen that for several years now though.

FWIW, the few times I've tested stuff like software video decoding, I've done comparisons with the case on and the case off, and if it's a tough codec that the SoC struggles with, the phone will perform much better with the case off than with the case on.

 

[Doug S]

What are you going on about? I can't think of any reviewers that benchmark mobile devices with any kind of active cooling or in any way other than the device as it ships. Maybe some have done it as part of a case review, but that's a separate matter than a review of a phone or the SoC in that phone.

ikjadoon literally posted a link two days ago showing that Andrei was doing this for smartphone SoC benchmarks while at Anandtech.

 

[Mopetar]

Just to nitpick: Occasionally you'll see reviewers that benchmark the phones while they're sitting inside a freezer. I haven't seen that for several years now though.

Unless it's being done for specific purposes it rather defeats the point of the review. I had assumed that in the review posted above it was only done because the reviewer suspected that the phone wasn't letting the chip reach full potential and wanted to investigate whether that was the case.

ikjadoon literally posted a link two days ago showing that Andrei was doing this for smartphone SoC benchmarks while at Anandtech.

I had initially misread that and assumed it was just being done to see if that particular phone was holding back the chip. I don't recall reading that was part of the testing methodology in any of the Apple SoC reviews (see this one for example), but perhaps I was skipping over it even if it was there.

Frankly if it's a review for the phone, I don't think he should use any kind of additional cooling. Even if he's just trying to do a deep dive on the SoC the other article shows that you're still testing the phone that's encasing it as well.

SPEC is still a good benchmark for determining CPU capabilities, but for a device review, showing results in the base state should be the baseline. Including additional results while the device is on a fan or in a freezer may be interesting from an academic perspective, but it's not particularly useful to me as a consumer. One could argue that AT had put all of that in separate articles related to the SoC more than the device itself, but compared to most reviews that layout the testing methodology in great detail, those articles are quite opaque when it comes to disclosing some rather important details.

 

[poke01]

Apple Silicon peaked

https://twitter.com/x/status/1704126470032114147

GPU is slower than Snapdragon 8 Gen 2. A17 is also less efficient CPU wise than A16

 

[Henry swagger]

Apple Silicon peaked

https://twitter.com/x/status/1704126470032114147

GPU is slower than Snapdragon 8 Gen 2. A17 is also less efficient CPU wise than A16

Tsmc N3 node seems to be terrible.. hope intel use n4p as backup gor arrow lake if yields d'not improve