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

 

[mikegg]

Oh so close to 4000/15000:

iPad16,4 - Geekbench

Benchmark results for an iPad16,4 with an ARM processor.

browser.geekbench.com

View attachment 98842

Less than 2% away from 4000, and only 0.5% away from 15000.
Maybe the M4 MacBook Pro in the coming year will achieve it.

BTW, Geekbench appears to have removed search access to the Geekbench 5 results.

I mean, that’s just ridiculous. It’s unbelievable really.

 

[Eug]

Heat dissipation is excellent. Didn't warm up much during his usage, except during charging.

Geekbench 6:



3D Mark:



Logic Pro export:



iMovie 30 minute video export:

 

[Eug]

Dave also says it stays cool. There is some throttling in looped 3DMark Wild Life Extreme Stress Test though.



It plateaus out at the second loop though and doesn't change after that. The 11" was ever so slightly lower performance than the 13".

Both lose about 25% performance or so, but then never go any lower.



Battery life remains good, but is lower than the M1 model in some tests.
The 11" M4 did better than the 13" for battery life, and in fact won the Genshin Loop test.

It should be noted that the 13" M4 got a slightly smaller battery compared to last generation, but the 11" M4 got a bigger battery.

 

[poke01]

Logic Pro export:

Any ideas on why Logic is 40% faster?

That’s a big jump

 

[Mopetar]

Any ideas on why Logic is 40% faster?

That’s a big jump

Probably improvements in some of the dedicated logic Apple bakes in to the hardware. They probably found a faster implementation that used more transistors that the node shrink gave them the room for. It might be something that was in M3.

 

[SteinFG]

Now that M4 jumped to 10 cores for bigger MT performance, do any of you expect iphone chips to jump from 6 to 7/8 soon? (this year or next)

 

[Nothingness]

I don't understand Chinese but here we go: https://www.bilibili.com/video/BV1ir421j7vR/?spm_id_from=333.999.0.0

At about 9:36 there are results for SPEC 2017 with power consumption:

INT A17 Pro: 9.32/3.62W vs M4 11.37/6.98W
FP A17 Pro: 13.88/4.40W vs M4 17.04/7.41W

I'm not sure how that is relevant given that he compares a phone and a tablet that are unlikely to have the same power settings.

 

[naukkis]

I don't understand Chinese but here we go: https://www.bilibili.com/video/BV1ir421j7vR/?spm_id_from=333.999.0.0

At about 9:36 there are results for SPEC 2017 with power consumption:

INT A17 Pro: 9.32/3.62W vs M4 11.37/6.98W
FP A17 Pro: 13.88/4.40W vs M4 17.04/7.41W

I'm not sure how that is relevant given that he compares a phone and a tablet that are unlikely to have the same power settings.

Those spec-results are excellent and pretty much in line with GB6 results.

 

[SarahKerrigan]

I don't understand Chinese but here we go: https://www.bilibili.com/video/BV1ir421j7vR/?spm_id_from=333.999.0.0

At about 9:36 there are results for SPEC 2017 with power consumption:

INT A17 Pro: 9.32/3.62W vs M4 11.37/6.98W
FP A17 Pro: 13.88/4.40W vs M4 17.04/7.41W

I'm not sure how that is relevant given that he compares a phone and a tablet that are unlikely to have the same power settings.

Looks like about a 5% iso-clock bump on int and +6% on FP, assuming both are running at their nominal frequencies.

 

[Eug]

iPad16,5 - Geekbench

Benchmark results for an iPad16,5 with an ARM processor.

browser.geekbench.com

Highest one so far?

Apparently that one was obtained in a freezer.

@amosliu137 posted the new highest single-core, obtained with liquid nitrogen.

Question - Liquid nitrogen cooling for M4

Maybe it is the first time to use Liquid nitrogen cooling ARM chip? The M4 is so difficult to gain 4.42Ghz in room temperature . The guys in Geekerwan use Liquid nitrogen to get the best performance of geekbench. M4 get 3977 for Geekbench 6 and 2754 for Geekbench 5.

forums.anandtech.com

iPad16,5 - Geekbench

Benchmark results for an iPad16,5 with an ARM processor.

browser.geekbench.com

New score is 3977, which is only 0.6% away from 4000.

 

[igor_kavinski]

Run it enough times and benchmark variance might just put it at or above 4000...

Maybe they are holding it wrong?

 

[Nothingness]

Looks like about a 5% iso-clock bump on int and +6% on FP, assuming both are running at their nominal frequencies.

Hmm what clocks did you pick? I picked 3.78 and 4.04 and I don't get the same result as you.

 

[SarahKerrigan]

Hmm what clocks did you pick? I picked 3.78 and 4.04 and I don't get the same result as you.

I put in 3.8 and 4.4 (3.78 is more precise, 4.4 is from Geekbench results for M4 - like this one.)

Isn't 4.05GHz the M3 clock speed? Or am I talking crazy? (This probably doesn't matter, I'm just wondering if it's where you got 4.04 maybe?)

 

[Nothingness]

I put in 3.8 and 4.4 (3.78 is more precise, 4.4 is from Geekbench results for M4 - like this one.)

Isn't 4.05GHz the M3 clock speed? Or am I talking crazy? (This probably doesn't matter, I'm just wondering if it's where you got 4.04 maybe?)

I'll blame the end of the day :|

 

[Nothingness]

So here we go, by benchmark for SPECint.

 

[SarahKerrigan]

So here we go, by benchmark for SPECint.

View attachment 98958

Thanks for posting the full subtests! The original video wasn't loading for me for whatever reason.

I don't really know what to make of it, honestly. I will say that xz is not my favorite subtest, for a lot of reasons, and it's the only regression. For the rest - this looks to me like microarchitecture changes, not merely uncore or memory controller changes, and I'm really curious what those changes are. It would be really nice if our friends at Apple did presentations at Hot Chips or ISSCC from time to time.

 

[Nothingness]

My favorite test is gcc and the improvement is not great on that one. I hope we'll soon have more data.

For xz, compression can have hard to predict branches (depending on the input files); so some changes to bpred beneficial to other tests might make it worse for xz.

 

[Doug S]

Run it enough times and benchmark variance might just put it at or above 4000...

Maybe they are holding it wrong?

Don't they know you have to point it at magnetic north (in the northern hemisphere, or magnetic south in the southern) so the magnetic flux across the FinFET transistors is aligned for optimal electron flow?

Of course the best possible performance would be achieved in the dead center of the southern magnetic anomaly during a geomagnetic storm like last weekend's.

 

[SarahKerrigan]

My favorite test is gcc and the improvement is not great on that one. I hope we'll soon have more data.

For xz, compression can have hard to predict branches (depending on the input files); so some changes to bpred beneficial to other tests might make it worse for xz.

I like gcc and perlbench - xalanc is often a good proxy for my workloads too. Not so much of a fan of xz, exchange2, or x264. Some of my reasons for loathing xz are not fully rational (it takes forever to run, and it's the only subtest that allows OpenMP parallelization in the official submission) but I've also just never seen a strong correlation between it and anything else.

 

[poke01]

More details from longhorn

https://twitter.com/x/status/1790580044550504468

 

[Eug]

9-core is doing very well

iPad16,4 - Geekbench

Benchmark results for an iPad16,4 with an ARM processor.

browser.geekbench.com

13411 multi-core

Freezer guy got 14924 multi-core for 10-core, and regular machines get about 14500 or so, meaning 10-core is only about 8-12% faster than 9-core for-multi core.

 

[roger_k]

Looks like about a 5% iso-clock bump on int and +6% on FP, assuming both are running at their nominal frequencies.

I wonder how they achieved such an increase on FP. I don't think there were any changes to the backend?

I'm not sure how that is relevant given that he compares a phone and a tablet that are unlikely to have the same power settings.

One should also keep in mind that they are testing that chip at the edge of its design capability. These are not intended to run with such aggressive cooling. The power consumption under usual circumstances will likely be lower (as will performance, judging from GB6 around 3-5%).

 

[Doug S]

My favorite test is gcc and the improvement is not great on that one. I hope we'll soon have more data.

For a CPU that's only 8 months further along the development timeline, and getting a bigger clock boost than the process improvement allowed for I don't think its that bad. I'm honestly mystified that Apple released the fastest CPU on the planet for single thread in a freaking 5mm tablet and people are underwhelmed. I guess they set the bar for improvement too high in the past, but diminishing returns are a thing. They had it a lot easier a decade ago when they were coming at x86 from way behind and large yearly gains were there for the taking.

 

[poke01]

I'm honestly mystified that Apple released the fastest CPU on the planet for single thread in a freaking 5mm tablet and people are underwhelmed. I guess they set the bar for improvement too high in the past, but diminishing returns are a thing.

I guess people want those big gains.

 

[igor_kavinski]

I'm honestly mystified that Apple released the fastest CPU on the planet for single thread in a freaking 5mm tablet and people are underwhelmed. I guess they set the bar for improvement too high in the past, but diminishing returns are a thing. They had it a lot easier a decade ago when they were coming at x86 from way behind and large yearly gains were there for the taking.

I'm sure that they are planning something big for the macbooks since their competition will be Zen 5 and Lunar Lake and even Strix Halo once the M4 Macbooks are released. They wouldn't like their shiny new CPU underperforming against those. I'm guessing they may do something extreme and even desperate (fancy cooling solution to hit high clocks etc.). The macbook M4 may not even be the same die as the iPad one. It may have bigger caches or other tweaks to really push the computing throughput. Starting from June, things are getting very, very interesting.

If they didn't have this mentality of wanting Apple users to own multiple Apple devices, they could've allowed the iPad Pro to at least run the normal MacOS or dual boot it with the iPadOS. All that power is simply wasted if it can't be used for desktop computing. At least the current M1/M2/M3 users would've gotten an immediate and overwhelming urge to buy the M4 iPad Pro and lighten their traveling load as well as enjoy a boost in their workloads. Apple, the company that refuses to do geeky stuff and instead tries to make computing exclusive and stylish, with obvious drawbacks like too much computing power in the hands of a lot of wealthy individuals with no idea of what to do with it other than mundane stuff. Ironically, despite packing the most computing power per square milimeter, Apple CPUs could be the most underutilized silicon in the ENTIRE history of computing. God bless corporate America and their brilliant business types!

The masses will have to wait maybe 5 years for these to be available on the used market and maybe then their true potential will finally be realized and they will most likely be used by people actually grateful for the computing power of these CPUs, using them for productive tasks that matter, rather than reading corporate emails, creating flashy but containing corporate speak gibberish presentations and planning corporate political strategies to one-up their colleagues and competitors and doing their part to make humanity miserable by destroying lives.