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

 

[Markfw]

In laptop's where efficiency is the most important. Apple, AMD and Intel use 8P cores in high end laptops. But what Intel and AMD do is boost for even for the tiny things to get maximum pref as this is the only way that x86 makers designed it and this greatly effects battery life.

If we compare the M1 Pro vs the 6800U where they both have 8P cores. The 14" M1 Pro gets over 4hrs more than the Asus 6800U. This cannot be only due to node difference. We can limit the peaks on AMD/Intel but this in turn effects the single core.


View attachment 68861
Take a look at this video, it's in English.

Well, I have a 6850 thinkpad that has a 12 hour battery life that I got for $1138 That $5999 Zenbook numer is crap. Hence I don't believe any of that.

Try again

 

[poke01]

Well, I have a 6850 thinkpad that has a 12 hour battery life that I got for $1138 That $5999 Zenbook numer is crap. Hence I don't believe any of that.

Try again

Its 5999 Yen. In brackets it says $888 in US dollars. Please take a look at the pic again.

 

[poke01]

Battery life varies between apps used, brightness and battery capacity and that's why you can't compare between yours and others in videos.

I have another battery test this time R5 4500U vs M1.

Both set at the same brightness, running a video playback. The R5 4500U lasted 13hrs and the M1 lasted over 20hrs.
The video starts at batterty testing.

AMD is no where near Apple's efficiency in laptops.

 

[Markfw]

Battery life varies between apps used, brightness and battery capacity and that's why you can't compare between yours and others in videos.

I have another battery test this time R5 4500U vs M1.

Both set at the same brightness, running a video playback. The R5 4500U lasted 13hrs and the M1 lasted over 20hrs.
The video starts at batterty testing.

AMD is no where near Apple's efficiency in laptops.

OK, that kind of convincing. But what apple does and what windows does are a little different. Apple may be the best in efficiency, but at least AMD is getting close. And 12 hours battery life ? unheard of in years past from Either AMD or Intel.

 

[poke01]

Apple may be the best in efficiency, but at least AMD is getting close. And 12 hours battery life ? unheard of in years past from Either AMD or Intel.

I wouldn't call AMD close to Apple as Apple gets is around 7hrs ahead of AMD in Linus's test but in 2022 AMD is much better than Intel.
Intel is not getting 12+ hrs of battery life like AMD is so Intel needs time.

But as you said Apple's lead is because years of experience in smartphone devices where pref/w is very important.

 

[Doug S]

I think you misunderstand how the process works. There are two version of N7: DUV and EUV; EUV version has greatly reduced the mask layer to 4 layers compared to DUV version. But after N7+ (official term for EUV), every increment process like N6 will cost more to make because they are adding more mask layers. N5 can add up to 14 layers and N4 will cost more cause the process adding one more layers.

N3E and N3 are different stories. At first, TSMC introduces N3 process which can go up to 25 layers thus too expensive and also the gap is too wide even for Apple. Thus N3E is introduced with 20 layers mask. Therefore N3E is cheaper than N3 but still more expensive than N4

I think it is you who misunderstand. There is more going on than just EUV vs DUV process steps.

See the below article. TSMC here is talking about N4P versus N5, saying N4P features a "simplified, and therefore cheaper manufacturing process, requiring fewer masks and less wafer turnaround time".

This is what TSMC said about N6, what they say above about N4 (N4P in this case but the 'P' is just the performance enhanced version of N4) and what they say about N3E. The follow ons are always refinements of the main node designed to reduce cost and in the case of 'P' nodes increase performance.

https://www.tomshardware.com/news/tsmc-announces-n4p-process-a-refined-n4-chip-node

 

[Tigerick]

I think it is you who misunderstand. There is more going on than just EUV vs DUV process steps.

See the below article. TSMC here is talking about N4P versus N5, saying N4P features a "simplified, and therefore cheaper manufacturing process, requiring fewer masks and less wafer turnaround time".

This is what TSMC said about N6, what they say above about N4 (N4P in this case but the 'P' is just the performance enhanced version of N4) and what they say about N3E. The follow ons are always refinements of the main node designed to reduce cost and in the case of 'P' nodes increase performance.

https://www.tomshardware.com/news/tsmc-announces-n4p-process-a-refined-n4-chip-node

Don't mislead by TSMC PR talk, at least refer to Anandtech article to get more insight information

TSMC Readies Five 3nm Process Technologies, Adds FinFlex For Design Flexibility

www.anandtech.com

N4P only improve 6% in performance but nothing else. The main improvement already done with N4 which improve upon N5 with 1.06X logic density.

Here is interesting part the article did not say: How TSMC do it? The answer is: by adding one more EUV layer total 15 layers. And you may wonder why TSMC keep mention about reducing mask layers: that is actually compared to past DUV version which with multi-patterning can go up to 50 layers. I don't have pricing info on DUV with insane layers but I believe N7+ with 4 layers of EUV mask mark new beginning of lithography process.

So no, the price of advanced process will be keep on going up by adding more EUV mask layers. It is simple as that. That's why Jensen said Moore Law is dead. And that is why AMD is using chiplet technique to prolong it.

Only when TSMC begin new N2 process with GAA we might see big changes in EUV mask layers...

 

[mmaenpaa]

Battery life varies between apps used, brightness and battery capacity and that's why you can't compare between yours and others in videos.

I have another battery test this time R5 4500U vs M1.

Both set at the same brightness, running a video playback. The R5 4500U lasted 13hrs and the M1 lasted over 20hrs.
The video starts at batterty testing.

AMD is no where near Apple's efficiency in laptops.

4000 series is quite old.

6000 series is much better. Where Apple shines is battery runtime at load, otherwise AMD is getting close.

Lenovo ThinkPad Z13 laptop review: AMD's premium ThinkPad with long battery life - NotebookCheck.net Reviews

 

[richardskrad]

The M2 MacBook Air is fan-less. Think about that for a second. AMD and Intel chips might have similar peformace but the laptop will melt without a fan. As as I am concerned, Intel and AMD make gasoline cars and Apple makes EV's. They are that far ahead.

4000 series is quite old.

6000 series is much better. Where Apple shines is battery runtime at load, otherwise AMD is getting close.

Lenovo ThinkPad Z13 laptop review: AMD's premium ThinkPad with long battery life - NotebookCheck.net Reviews

View attachment 68867

 

[igor_kavinski]

View attachment 68867

Load battery life of i5-1250P is really good! Seems it is "leaking" battery in the other tests.

 

[mmaenpaa]

The M2 MacBook Air is fan-less. Think about that for a second. AMD and Intel chips might have similar peformace but the laptop will melt without a fan. As as I am concerned, Intel and AMD make gasoline cars and Apple makes EV's. They are that far ahead.

That may well be. I was only referring to battery runtime. For example if we compare Z13 to Macbook Pro 13 2022 M2 (which is not fanless), Apple manages H.264 playback 19.8 minutes / Wh (1153/58.2), AMD 19 minutes / Wh (979/51.5). On WiFi test Apple shows what they can do when they control every aspect of hardware, manufacturing & software (19.6 min /Wh vs. 16.7 min /Wh). Still not bad job from AMD as they are on inferior node compared to Apple and probably will stay that way for years to come

 

[Mopetar]

It's not strictly fair to just compare AMD to Apple as OEMs make decisions about other components. Apple as a company probably pushes harder for battery life than just about anyone else. There's a market for battery life in the x86 notebook world, but performance is seen as just as important and consumers often care about these other factors more.

 

[Markfw]

It's not strictly fair to just compare AMD to Apple as OEMs make decisions about other components. Apple as a company probably pushes harder for battery life than just about anyone else. There's a market for battery life in the x86 notebook world, but performance is seen as just as important and consumers often care about these other factors more.

For a laptop, I use it on AC most of the time. The times I DON'T are on airplanes and other times when AC is not available. So 9-12 hours on battery is plenty for me. Then performance comes next. The 6850H laptop I just got seems to be the best of both worlds. Plus I only spent a little over $1000 for it.

 

[Doug S]

Don't mislead by TSMC PR talk, at least refer to Anandtech article to get more insight information

TSMC Readies Five 3nm Process Technologies, Adds FinFlex For Design Flexibility

www.anandtech.com

N4P only improve 6% in performance but nothing else. The main improvement already done with N4 which improve upon N5 with 1.06X logic density.

Here is interesting part the article did not say: How TSMC do it? The answer is: by adding one more EUV layer total 15 layers. And you may wonder why TSMC keep mention about reducing mask layers: that is actually compared to past DUV version which with multi-patterning can go up to 50 layers. I don't have pricing info on DUV with insane layers but I believe N7+ with 4 layers of EUV mask mark new beginning of lithography process.

So no, the price of advanced process will be keep on going up by adding more EUV mask layers. It is simple as that. That's why Jensen said Moore Law is dead. And that is why AMD is using chiplet technique to prolong it.

Only when TSMC begin new N2 process with GAA we might see big changes in EUV mask layers...

Given that Anandtech is parroting other TSMC information I'm not sure why you are asking me to ignore some information while listening to other information - especially when that other information does not include ANYTHING about cost.

You are wrong if you think you can count EUV layers and believe that equals cost, simple as that. There are over 300 process steps required to fab a wafer, there is room to create efficiencies by reducing the number of steps or the cost of those steps regardless of what happens with the number of EUV exposures.

Not sure why you discount TSMC's own claims and substitute your own, but I guess Dunning-Kruger is strong in some people.

 

[Mopetar]

For a laptop, I use it on AC most of the time. The times I DON'T are on airplanes and other times when AC is not available. So 9-12 hours on battery is plenty for me. Then performance comes next. The 6850H laptop I just got seems to be the best of both worlds. Plus I only spent a little over $1000 for it.

I'm on AC most of the time as well, but Apple's newest notebooks with their own chips are a cut above.

On the previous MBP I had with an Intel chip running a Zoom call while demonstrating some software that would tax the CPU would suck the battery dry pretty quickly, especially if I had to increase the screen brightness. In some cases I couldn't expect too much more than 2 hours of that kind of workload.

With the new M1 MBP I have that only seems to tickle it now and I could probably go at least 6 hours before needing to plug in. Although I never had problems with the previous one where it would run out of juice or require me to plug in, it's far less stressful knowing that you could actually go an entire day away from an outlet.

It's great when you're traveling and can just work on the plane without having to worry about being able to find a place to charge at the next store between flights because it'll last without. There are probably plenty of PC notebooks that are built for that use case as well, but at that point you're paying just as much as you would for a Mac.

 

[Eug]

Mark Gurman claims to have the specs of one Mac Pro SKU undergoing testing.

Gurman: Apple Silicon Mac Pro Testing Ramping Up, New MacBook Pro and Mac Mini to Launch Within Months

Apple is planning to debut several new Macs over the next several months, including new MacBook Pro, Mac mini, and Mac Pro models, according to...

www.macrumors.com

24 CPU cores (16 performance + 8 efficiency)
76 GPU cores
192 GB RAM
Ventura 13.3

He also says it will support up to 48 CPU cores and 152 GPU cores, and up to 256 GB memory, to be released in 2023. Not sure why the memory would be limited to 256 GB RAM if the 24-core model above can support 192 GB RAM. I wonder what the configuration is, and if perhaps 384 GB RAM could be supported.

M2 Max will have 12 CPU cores (8+4) and 38 GPU cores, up to 64 GB RAM.

It should be noted that M1 Max already supports 64 GB RAM, and M1 Ultra supports 128 GB RAM.

Even if it is 384 GB RAM maximum for the 2023 Mac Pro, that's still a far cry from the 1.5 TB maximum in the current Intel Mac Pro. I wonder how many people actually buy these things with more than 384 GB RAM though.

 

[Doug S]

Mark Gurman claims to have the specs of one Mac Pro SKU undergoing testing.

Gurman: Apple Silicon Mac Pro Testing Ramping Up, New MacBook Pro and Mac Mini to Launch Within Months

Apple is planning to debut several new Macs over the next several months, including new MacBook Pro, Mac mini, and Mac Pro models, according to...

www.macrumors.com

24 CPU cores (16 performance + 8 efficiency)
76 GPU cores
192 GB RAM
Ventura 13.3

He also says it will support up to 48 CPU cores and 152 GPU cores, and up to 256 GB memory, to be released in 2023. Not sure why the memory would be limited to 256 GB RAM if the 24-core model above can support 192 GB RAM. I wonder what the configuration is, and if perhaps 384 GB RAM could be supported.

M2 Max will have 12 CPU cores (8+4) and 38 GPU cores, up to 64 GB RAM.

It should be noted that M1 Max already supports 64 GB RAM, and M1 Ultra supports 128 GB RAM.

Even if it is 384 GB RAM maximum for the 2023 Mac Pro, that's still a far cry from the 1.5 TB maximum in the current Intel Mac Pro. I wonder how many people actually buy these things with more than 384 GB RAM though.

How much memory is offered and how much memory is possible are two different things. With the current LPDDR5 chips used to make the modules in the M1 Ultra Studio, if doubled to an "Mx Extreme" would allow for up to 2 TB of memory if they maxed out the size of the stacks. Talking about a 192 GB M2 Ultra there may be an indication they are using newer LPDDR5X Micron chips that are 50% more dense, which would raise the potential ceiling to 3 TB.

The problem with offering larger memory configurations with on package LPDDR is that it has to be built that way at TSMC's packaging facility! That's a big complication over even soldering chips to the motherboard, because at least that is happening inside your own (well Foxconn's, but same thing for Apple) manufacturing facility.

If you max out at 384 GB that covers the needs (and budget) of the overwhelming majority of customers. Let's say 5% of your customers want more than that, the problem is they aren't going to all want the same amount - and the cost difference between "what I want" vs "what SKUs Apple is willing to deal with" might be pretty severe - i.e. "I need 1 TB but Apple only sells 768 GB and 1.5 TB". Paying for 512 GB of LPDDR5X you don't need would put a dent in the 'ol budget! There's enough flexibility Apple should be able to build a 1 TB model but how much variation would they want to deal with, given that every SKU would have less and less demand?

I wouldn't be shocked to see them only list the standard 384 GB of DRAM, but have sort of under the table special order thing where you can buy larger configurations than the standard 384 GB, but only if you are ordering a certain number, are willing to put down a sizeable deposit on that order in advance, and are willing to wait a few months for delivery.

 

[Eug]

I wouldn't be shocked to see them only list the standard 384 GB of DRAM, but have sort of under the table special order thing where you can buy larger configurations than the standard 384 GB, but only if you are ordering a certain number, are willing to put down a sizeable deposit on that order in advance, and are willing to wait a few months for delivery.

Well, one point I was making is that he said that he knows a 192 GB 24-core machine is being tested, but also that 256 is the likely max configuration even for 48-core. If 24-core can do 192 GB, I might have guessed 48-core could do 384 GB. But then again, maybe the on-board memory configuration wouldn't have made that cost effective. Or maybe it's just wrong.

It's also interesting to see that listed GPU configs are multiples of 19, instead of the previous 20. I guess those all could be multiples of 20 binned to 19.

I don't think Apple has ever shipped a machine with more RAM than it listed as supported.

 

[Doug S]

Well, one point I was making is that he said that he knows a 192 GB 24-core machine is being tested, but also that 256 is the likely max configuration even for 48-core. If 24-core can do 192 GB, I might have guessed 48-core could do 384 GB. But then again, maybe the on-board memory configuration wouldn't have made that cost effective. Or maybe it's just wrong.

It's also interesting to see that listed GPU configs are multiples of 19, instead of the previous 20. I guess those all could be multiples of 20 binned to 19.

I don't think Apple has ever shipped a machine with more RAM than it listed as supported.

I'm very suspicious of the "low end" Mac Pro having 3/4 the DRAM of the high end model, unless there is a higher higher end out there with 384 GB or more. It doesn't make sense since it would reduce the cost/price gap between the base config of the two models.

I find it strange the extent Apple seems to bin on GPU cores. Even iPhones have done so. Obviously it isn't defect related, since defects don't occur preferentially in a GPU. So I'm guessing this is binning for power, and knocking out the highest power core. Maybe there is something about the design of the GPU or transistor types they use in those blocks that leads to great variability in power draw?

 

[jpiniero]

I find it strange the extent Apple seems to bin on GPU cores. Even iPhones have done so. Obviously it isn't defect related, since defects don't occur preferentially in a GPU.

Of course they do. GPU designs are possibly more resilient than say CPUs. But it's still an issue.

 

[Doug S]

Of course they do. GPU designs are possibly more resilient than say CPUs. But it's still an issue.

What do you mean "of course they do"? How would defects be more likely to occur in a GPU core than any other block of the same size?

 

[Eug]

Take this with a grain of salt, but I post it because he had pre-release info of Mac Studio before it was launched.

Basically the claim is that the Mac Pro sample board has PCIe slots, and will actually recognize an AMD GPU, but the GPU doesn't work. Furthermore, no external RAM is supported. All is unified RAM, up to 192 GB in the 24-core model.

40 core ASI with PCI-E slot is around the corner!!

I want to confirm Gurman info: Latest Prototype Mac Pro is now based on 24 Core M2 and 192GB Prototype board is allocate in 7,1 Case with More than 1 PCI-E Slot, GPU is now show it's name in system preference. but not working properly. Still no ram slot. PS. That latest Mac Pro include beefy...

forums.macrumors.com

I want to confirm Gurman info:

Latest Prototype Mac Pro is now based on 24 Core M2 and 192GB <<Unified Memory.>>
Prototype board is allocate in 7,1 Case with More than 1 PCI-E Slot, <<my pay tell me that 6 pci-e++>>
GPU is now show it's name in system preference. but not working properly.
Still no ram slot.

PS. That latest Mac Pro include beefy SoC thermal cooler.

40 core ASI with PCI-E slot is around the corner!!

1. sorry it is my typo, prototype machine is contain 6 pcie slot. 2. it recognize (shown as RX6900XT) but not working.

forums.macrumors.com

1. sorry it is my typo, prototype machine is contain 6 pcie slot.
2. it recognize (shown as RX6900XT) but not working.

 

[Eug]

Commercial Times claims M2 Pro/Max is on 3 nm, even though some here say the timing is off and that N3 volume would be on N3E.

Apple Has Reportedly Given TSMC a Large Order to Mass Produce Its M2 Pro, M2 Max and Future SoCs on the Latest 3nm Process

A new report states that TSMC has taken a large order from Apple to mass produce M2 Pro, M2 Max and future chipsets on its 3nm node

wccftech.com

蘋果M2打核心戰 台積傳獨吞3奈米大單

雖然個人電腦（PC）市場需求低迷不振，但蘋果仍將在未來幾個月推出多款新機，而蘋果亦加速第二代Apple Silicon的M2處理器研發及量產，據外電及業界消息，蘋果M2處理器將大打核心戰，可能推出搭載48核心的最高階M2 Extreme搶攻高階PC及工作站應用，晶圓代工龍頭台積電可望獨吞3奈米晶圓代...

ctee.com.tw

It also claims that this time around M2 Pro will have fewer CPU cores than M2 Max.

 

[poke01]

It also claims that this time around M2 Pro will have fewer CPU cores than M2 Max.

The base M1 Pro in the 14" also has fewer CPU cores then M1 Max. So it's not a huge suprise.

 

[Eug]

The base M1 Pro in the 14" also has fewer CPU cores then M1 Max. So it's not a huge suprise.

Those are binned though, and the binned variants also have fewer GPU cores.

For M2 Pro, the chip spec has its full complement of GPU cores. It sounds like a non-binned part.