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

 

[DrMrLordX]

Thanks for the heads up on the DX level of that test. Apparently, it doesn't run on RTX or AMD DX12 hardware, which is weird. Probably GFX Bench's compatibility issue?

Not sure, I can run Aztec Ruins and Aztec Ruins High on my Radeon VII in all three APIs. Or at least I used to be able to with an older build. Looks like all the DX12 and Vulkan results on newer dGPUs have been purged from Kishonti's website.

I was going to reinstall since my old install can't communicate with Kishonti's servers for some unknown reason, but Windows is freaking out over the lack of publisher data, and virustotal is showing 1 AV engine (Acronis) detecting malware in the installer. Specifically, Trojan/Generic.ASMalwS.3463160 . Not sure if that's a false positive but I don't want to touch it right now.

Anyway, Kishonti's GFXBench 5 will run on newer hardware in DX11 mode, and they have scores listed in their database.

 

[nxre]

So the 32 core GPU (if that is the one being benchmarked) seems to be in RTX 3060 performance class. Crazy.
Especially considering its low power consumption should allow it to run faster for longer than most laptop discrete GPUs.

 

[uzzi38]

Compared to a 6600M

GFXBench is a waste of time of a benchmark. In fact it's so much of a waste of time that the M1 can lose to the A14.

 

[igor_kavinski]

Apple's GPU is giving the same score while working on 3.6 times more pixels, using max 60W compared to 3050's max 75W!

 

[Eug]

So the 32 core GPU (if that is the one being benchmarked) seems to be in RTX 3060 performance class. Crazy.
Especially considering its low power consumption should allow it to run faster for longer than most laptop discrete GPUs.

The king of all iGPUs?

BTW, FWIW:

MacBookPro18,2 - Geekbench

Benchmark results for a MacBookPro18,2 with an Apple M1 Max processor.

browser.geekbench.com

 

[igor_kavinski]

View attachment 51650

Beating the 2060 Max-Q with 5W less TDP.

 

[igor_kavinski]

Apple is in a truly unique position to launch a Nintendo Switch killer with 20 hours battery life.

 

[Mopetar]

Everything about statement is incorrect. The M1 contained 8 CPU cores. 4 big and 4 small. The M1 Pro, for example, ALSO only contains 8 cores. Do I need to further elaborate?

The M1 Pro/M1 Pro Max must be evaluated on their own terms thanks to significantly different power targets and core types.

First of all, that's a horrible take because you're mixing big and little cores together into a single number. This might be forgivable if we were on a knitting forum and you honestly didn't know better, but we aren't and I hope you do. Worse you've compared the binned M1 Pro against the M1 instead of the full die which even by your own tortured logic would have 10 cores.

We've already seen enough benchmark scores posted to think that the extrapolation is in the ballpark and actually does match up to Apple's figures, or is at least close enough that you can see how they wound up on a marketing slide without being stretched too far. This is also a tech forum where posters speculate on performance based on rumors or in lieu of actually performance metrics.

 

[thunng8]

First you are falling into the trap of measuring Watt/Perf , not Perf/watt ..

Both of these are venerable to being missleading, dependent on where you set the reference point, but more so with the former.

In this case you're choosing a performance point (~1500 1T CB23 ) that thanks to a significant IPC deficit (approx 35-37% higher ST IPC for M1) is only achievable for Zen3 at much higher frequencies - Frequencies it can achieve -and beyond, but at the expense of power consumption.

IF AMD decided not to clock Zen 3 beyond say , 3.2ghz, you would simply NOT be able to make this comparison. You'd instead find , rather conveniently, that at these frequencies Zen 3 vs M1 ST Power consumption is essentially the same (by numbers quoted in your thread - i.e 3.8W for M1 @ 1C-1T 3.2Ghz) , This means the M1's advantage in perf/watt is literally it's IPC .. i.e about 35-37% . A far cry from 400-500%.

I could make similar claims for Zen 3 vs Zen 2 by making a Zen3 CPU match the ST score of a 3950x, running at maximum 1C turbo - The IPC difference means the former will only have to run at , say 3.9Ghz, to match a 4.6-4.7Ghz 3950x With orders of magnitude less power as a result (around ~8w vs ~18w) With that it would be all to easy to say Zen 3 is over 200% more efficient.. Which we all know is not the case.


Secondly, this doesn't account for Throughput.. The M1 core is HEAVILY geared towards maximum 1T IPC, and low frequencies. Once you put two threads down a Zen3 core with SMT , the efficiency delta drops even more. since core power does not increase in line with perf uplift.

Some examples below: Different SKU's sustaining different frequencies..


View attachment 51635


View attachment 51637


Anyway. this will all become more than apparent if the M1X/pro scores as predicted - i.e around 12000 @ 30w - but with the help of 2x efficiency cores. Vs Cezzane @ 10500 or so @ 35w 8 core.

So no - there's no magic happening with the M1. It's just a very focused, high IPC, low clock design.

M1 can play the low frequency angle as well. I’ve tested my MacBook Air running cinebench and when it is thermally throttled the power usage goes down to 7w and the cinebench scores goes down to approx 6200-6400. Dropping the frequency down by 20% halves the power usage.

I believe dropping down further doesn’t improve performance per watt.

 

[Shivansps]

Apple is in a truly unique position to launch a Nintendo Switch killer with 20 hours battery life.

Im petty sure that the TDP of the PRO and MAX are way beyond what can be used on these devices. In fact the Max, based on those Apple numbers, should be faster than the PS5.

But yeah, that what i was thinking, gaming is not Apple priority, in fact they dont care AT ALL, but they now have some very interesting SoC that allows them to enter that market, if they wish.

 

[beginner99]

While the performance and performance/watt is impressive, it's also clearly they artificially cripple it by shipping it right of the bat with too little RAM. My 3 year old company laptop has 32 gb RAM and yes I do need it. And data is only going to grow meaning more memory needed.

Yes 32 gb is enough for most use-cases but you aren't buying such a machine for an average "office" use case. I kind of don't see the balance in performance especially also from the GPU which shares the RAM pool. And then you need to buy a overpowered iGPU to get 64 GB. This is already somewhat an issue in PC laptops to a degree, want ton of CPU and RAM you also get a bigger display and usually a gpu and other features you don't care about. But needing to shell out a lot more just to get 64GB RAM? Really? It's clear this will be the reason one needs to replace this machine. And anyone making actually use of that 32.core GPU, I fear could even be limited by 64 Gb RAM, AI...8k video,...

 

[lobz]

That is simply not true. At best about 25% of their advantage is caused by the process. Just take a look at what TSMC officially claims and then take a look at Apple's superiority again.

If you're gonna use such simplistic terms as 'superior', please also insert 2 claims (1 TSMC and 1 Apple) into your post. I'm sure you know where they're found, you seem to know their implications exceptionally well.
I'm not trying to mock you, I'm honestly trying to find out, exactly where am I supposed to see a 500% (!!!) power efficiency advantage.

I'm absolutely willing to compare the exact claims and / or independent findings of these 2 offerings. Of course, my emphasis on the word 'exact' is no coincidence. I'm sure we can all agree that 5 European men can in a football stadium can carry 30 stacked A4 papers on a plate from one post to the other one while maintaining the integrity of all 30 individual papers, but 5 American men can not do the same thing - if it's raining on that day in the American stadium..... I'm sure you can understand my concerns about your unbelievably bold 500% claim 🙂

 

[Ajay]

So please - to all those that think that 5nm vs. 7nm was the reason for their superiority: It is not.
The process advantage is almost negligible when looking at the numbers.

Just keep in mind that the M1 Max would not have designed and manufactured on 7N instead of 5N.
TSMC's charts show a density increase of 1.8 (max, much much less for the large amount of SRAM on the M1 line up) for 5N. The M1 Max, in particular, would have been unreasonably large on 7N.

To what degree 5N helped the power usage on these new M1 SoCs is something that only Apple and TSMC know. It's a complex function of design within the limits of the PDK and RTL characteristics of individual functional blocks. Apple faced the classic tradeoff on 5N FF as to the degree they took advantage of the potential performance gains vs power savings, and we don't know what that tradeoff was.

Lastly, there needs to be more extensive testing to determine actual performance across a broader range of standardized tests (like SPEC) and performance on software typically used by the MacBook Pro's users. Unfortunately, IIRC, because of Apple's closed system, we cannot run power vs frequency tests (well, some enterprising EE could, if they have deep pockets and can afford to potentially brick a MacBook Pro. This is easily done with AMD and Intel CPUs.

So, don't be too hasty in making such a broad proclamation; for the sake of your own credibility.

 

[lobz]

You have to give Apple credit for bringing meaningful performance increases to the Macbook product line. Intel had basically failed to do that for the past 7 years.

I'm not sure it's about giving credit per se. It's more about bringing doom upon yourself because you see a single rival as forever defeated, and then trying to sell that one single thing to your shareholders for a whole decade and doing EVERYTHING in order to make them unaware of the (obviously inevitable) changes that occur in every market.

Now put all that in perspective with the tone of what Pat Gelsinger is using in almost every interview he gives...

Hector Ruiz must be spinning in his grave, yelling "WHY THE HECK HAVE I BEEN FIRED?!"

 

[Eug]

While the performance and performance/watt is impressive, it's also clearly they artificially cripple it by shipping it right of the bat with too little RAM. My 3 year old company laptop has 32 gb RAM and yes I do need it. And data is only going to grow meaning more memory needed.

Yes 32 gb is enough for most use-cases but you aren't buying such a machine for an average "office" use case. I kind of don't see the balance in performance especially also from the GPU which shares the RAM pool. And then you need to buy a overpowered iGPU to get 64 GB. This is already somewhat an issue in PC laptops to a degree, want ton of CPU and RAM you also get a bigger display and usually a gpu and other features you don't care about. But needing to shell out a lot more just to get 64GB RAM? Really? It's clear this will be the reason one needs to replace this machine. And anyone making actually use of that 32.core GPU, I fear could even be limited by 64 Gb RAM, AI...8k video,...

I'd say your use case is not the norm for these machines, and the projected use case for others buying this machine is also not always quite accurate.

The majority of the creatives are fine with 32 GB RAM, with only a subset really requiring 64 GB. The creatives I know are going for 32 GB RAM. And no, most of them don't need the M1 Max either. Those who do would include most of the subset that need 64 GB RAM, but the need for 128 GB in their laptops is very rare. The 0.01% who really do need 128 GB or 256 GB are going to have the option of the Mac Pro in 2022. Now, do I agree that having cheaper 64 GB options would be better? Yes! Do I think it will hurt their sales? No, not to any significant extent.

And for most office use cases, even 16 GB is enough, and I also think it's perfectly reasonable for someone to buy a 16" M1 Pro 16 GB just to work in Excel, since it's gonna have a really nice and big screen with lots of ports. They just don't need to buy the M1 Max.

 

[StinkyPinky]

I'd say your use case is not the norm for these machines, and the projected use case for others buying this machine is also not always quite accurate.

The majority of the creatives are fine with 32 GB RAM, with only a subset really requiring 64 GB. The creatives I know are going for 32 GB RAM. And no, most of them don't need the M1 Max either. Those who do would be the subset that need 64 GB RAM, but the need for 128 GB in their laptops is very rare. The 0.01% who really do need 128 GB or 256 GB are going to have the option of the Mac Pro in 2022. Now, do I agree that having cheaper 64 GB options would be better? Yes! Do I think it will hurt their sales? No, not to any significant extent.

And for most office use cases, even 16 GB is enough, and I also think it's perfectly reasonable for someone to buy a 16" M1 Pro 16 GB just to work in Excel, since it's gonna have a really nice and big screen with lots of ports. They just don't need to buy the M1 Max.

All this is true, but why do people buy nice muscle cars when a Nissan Versa will do? Why do people buy nice leather jackets when a a cheap hoodie will do? Why do people eat fine foods when a salad will do? People are enthusiasts about lots of things, and they buy this stuff often out of enjoyment rather than need.

 

[Eug]

All this is true, but why do people buy nice muscle cars when a Nissan Versa will do? Why do people buy nice leather jackets when a a cheap hoodie will do? Why do people eat fine foods when a salad will do? People are enthusiasts about lots of things, and they buy this stuff often out of enjoyment rather than need.

I agree with that. Like I said before, people are strongly considering going for 10-core M1 Pros with 32 GB RAM just to run Office and Chrome despite others telling them it's a waste of money.

However, what I was saying is that the suggestion that the lack of 64 GB options on the low end is a major issue isn't really justified in the greater scheme of things. If they really want 64 GB, they can just suck it up and pay the extra $200 to get M1 Max as well.

It certainly isn't going to hurt Apple sales revenue.

In my case I want more ports than the current M1 offers. I would love to get an M2 with more ports in 2022, but I suspect that will be crippled like the M1 in terms of number of ports. Does that suck? Yes. But in the end I'm thinking I'll (hopefully) only have to pay $200 more to jump up to M1 Pro 8-core (6+2) to get those extra ports (and a new case design). And in retrospect it's not actually as bad it seems in my case since I'd have to pay $100 anyway to get a Mac mini-shaped hub to go with that port-starved M2 Mac mini, with more potential for weird compatibility issues. I just hope it's not $300 more to get those damn ports.

BTW, regarding that car analogy: I don't drive muscle cars. I currently drive a Prius Plug-in. The performance isn't great, but I don't really care much. What I do care more about though is the lack of Apple CarPlay (old model) and the poor sound insulation, as well as the lack of AWD. The heater also sucks. For my next car I'd be happy to buy another Prius (non-Plug-in) if it fixed all those things. It turns out the new Prius has AWD and Apple CarPlay but likely still poor sound insulation and I'm not sure about the heater. So, for my next car I haven't decided yet but I may end up jumping up a tier and getting an AWD Lexus even though I don't really care that much about the engine performance. But the improved handling and pickup of the Lexus would still be a nice bonus even though they aren't priorities for me.

What I'm saying here is I'd buy a luxury AWD Prius, but they don't exist so I can't. However, I realize that market would be quite small, and small enough that it doesn't really make sense for Toyota to sell a model to cater to that small niche crowd, since they know luxury buyers tend to want more performance too. Similarly, while it'd be nice to have a 64 GB option for M1 Pro buyers, in the end it likely doesn't make much business sense for Apple to offer that (even ignoring technical considerations for the moment), esp. considering a significant portion of those wanting 64 GB would want more GPU performance as well.

 

[LightningZ71]

So, I've been looking for someone to say this, and I haven't seen it yet...

Will it mine? With 400GB/sec of memory bandwidth, and with it just sipping power, these devices have the makings of VERY efficient coin miners. I wonder if someone is going to try to get mining going on these things beyond the first few proof-of-concept efforts that were made at the end of 2020 with absolutely no optimization for the architecture. It would seem to me that it should be able to do about 8 processes of XMR on the performance cores and however much on the GPU side.

 

[linkgoron]

View attachment 51648
View attachment 51649

Apple's GPU is giving the same score while working on 3.6 times more pixels, using max 60W compared to 3050's max 75W!

Well, they're using metal vs OpenGL, on a way better node (8nm Samsung vs 5nm TSMC) with at least 4 times the number of transistors. I also wouldn't compare "max" TDP without actual real world numbers ("max" does not mean that it's actually using that amount of power). Not saying Apple's numbers are not impressive, but it's not exactly a correct comparison.

 

[igor_kavinski]

So, I've been looking for someone to say this, and I haven't seen it yet...

Will it mine? With 400GB/sec of memory bandwidth, and with it just sipping power, these devices have the makings of VERY efficient coin miners. I wonder if someone is going to try to get mining going on these things beyond the first few proof-of-concept efforts that were made at the end of 2020 with absolutely no optimization for the architecture. It would seem to me that it should be able to do about 8 processes of XMR on the performance cores and however much on the GPU side.

Miners are more interested in creating mining farms with dozens of GPU's. Doing that with MBP's is cost prohibitive, unless a single MBP starts replacing an entire mining rig.

 

[uzzi38]

Miners are more interested in creating mining farms with dozens of GPU's. Doing that with MBP's is cost prohibitive, unless a single MBP starts replacing an entire mining rig.

Perhaps you missed the images of miners using gaming laptops?

 

[Eug]

The MacBook Pro review embargo supposedly lifts on Monday, October 25. However, I suspect they'll still be the non-technical sometimes-fluff pieces as usual.

The good news though is that the delivery dates for some purchased units begin Tuesday, October 26, so we should have solid benchmarking of M1 Pro (8-core and 10-core) and M1 Max with various applications by early next week.

I'm especially curious about fan noise, but it looks like the cooling design has been significantly improved in the current models, sort of like how much better the iMac Pro's cooling was compared to the iMac.

 

[LightningZ71]

Perhaps you missed the images of miners using gaming laptops?

Exactly, miners don't care what they use to mine with... Witness the laptop farms...

If the M1 Pro/Max can mine with the best of them, and sip power while doing it, they WILL be bought for mining... Especially since the binned chips retain full memory bandwidth, making them more attractive from a price/performance standpoint. Loosing CPU cores doesn't hurt as mining XMR is always just gravy. It's the Eth that pays, and that memory bandwidth with sufficient memory to hold the tables will certainly pay the bills.

 

[igor_kavinski]

Exactly, miners don't care what they use to mine with... Witness the laptop farms...

Well, I would welcome that if it means that the market will be flooded with affordable used M1 Max laptops once the crypto bubble bursts.

 

[moinmoin]

So, I've been looking for someone to say this, and I haven't seen it yet...

Will it mine? With 400GB/sec of memory bandwidth, and with it just sipping power, these devices have the makings of VERY efficient coin miners. I wonder if someone is going to try to get mining going on these things beyond the first few proof-of-concept efforts that were made at the end of 2020 with absolutely no optimization for the architecture. It would seem to me that it should be able to do about 8 processes of XMR on the performance cores and however much on the GPU side.

Shush!