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

 

[StinkyPinky]

No way they are sending 24 core units out for review.

Very interested in the reviews. If the 32 core Max loses efficiency/performance quickly (kinda like the 6900XT) then that's not too unexpected. Apple themselves said the Pro had 2x performance and Max 1.7x so even they admit it. This seems expected behavior for basically any GPU. I suspect for the desktop Mac Pro it will be even less efficient.

 

[Eug]

No way they are sending 24 core units out for review.

Very interested in the reviews. If the 32 core Max loses efficiency/performance quickly (kinda like the 6900XT) then that's not too unexpected. Apple themselves said the Pro had 2x performance and Max 1.7x so even they admit it. This seems expected behavior for basically any GPU. I suspect for the desktop Mac Pro it will be even less efficient.

If I understand his post correctly, some guy out there is claiming that a YouTube video went up showing Metal at 72220 for M1 Max, but then it was taken down.

We'll see, but if true, that would put the M1 --> M1 Max 32-core scaling at about 3.3X to 3.4X.

The only performance difference between any of the M1 machines, is that the fanless MBA, will heat up and throttle at higher loads. Otherwise they all perform the same.

The one difference besides cooling is the 7-core GPU in the base model MacBook Air and iMac.

 

[Heartbreaker]

If I understand his post correctly, some guy out there is claiming that a YouTube video went up showing Metal at 72220 for M1 Max, but then it was taken down.

We'll see, but if true, that would put the M1 --> M1 Max 32-core scaling at about 3.3X to 3.4X.


The one difference besides cooling is the 7-core GPU in the base model MacBook Air and iMac.

It's possible that they benchmark might have scaling issues, and it's also possible that Apple has scaling issues as well.

Different benchmarks might show different levels of scaling.

 

[Eug]

It's possible that they benchmark might have scaling issues, and it's also possible that Apple has scaling issues as well.

Different benchmarks might show different levels of scaling.

GFXbench scales at 3.17X to 3.99X depending upon the test, with one outlier at 2.4X.

 

[Semel]

GFXbench is not a good benchmark to measure GPU performance for real modern tasks..

 

[Eug]

GFXbench is not a good benchmark to measure GPU performance for real modern tasks..

Didn't say it was a great test. I'm just saying that depending upon the test, there can be huge differences in reported scaling, the implication being that it's not a big surprise that Geekbench 5 Metal scales at "only" 3.3X to 3.4X or so.

 

[DrMrLordX]

Metal API was issue enough but now ARM? Developers of console\PC games just don't want to bother with this.

Just something to think about, but MS and Qualcomm are still at least nominally pushing WARM, and there is the off chance that AMD may get into the mix through their semi-custom unit. How that helps gaming on Apple is dubious, but at least we aren't talking targeting an entirely different OS and uarch, but just an entirely different OS.

 

[jeanlain]

The M1 Max compute scores reminds me of the A14 GPU being slower in the iPhone than in the iPad Air. I'm not sure if the cause has been identified, but there were speculations about the system limiting GPU power consumption on the iPhone, or something.

 

[Eug]

The M1 Max compute scores reminds me of the A14 GPU being slower in the iPhone than in the iPad Air. I'm not sure if the cause has been identified, but there were speculations about the system limiting GPU power consumption on the iPhone, or something.

Dunno about A14 but for the A12 iPhone XR I could get iPad Air 3 speeds but only if I removed the case, waited for a cool phone, and propped it up for better airflow. If I didn’t do those things, the speeds were lower.

IOW, A12 is thermally constrained in a phone.

That was CPU though. I didn’t test GPU.

 

[dc443]

Well, it's an impressive piece of hardware, and I have no doubt it will be great for video editing and such..however... gaming will still suck... unfortunately..

1) Metal API.

Keeping going Metal route instead of Vulkan was a huge mistake.

There's a thing called MoltenVK.

 

[StinkyPinky]

So apparently the new MacBooks have a "high powered mode", I wonder if that kicks in under mains power only or if that's a setting you check somewhere. Perhaps doing that gives you another nudge in performance.

 

[Eug]

So apparently the new MacBooks have a "high powered mode", I wonder if that kicks in under mains power only or if that's a setting you check somewhere. Perhaps doing that gives you another nudge in performance.

Maybe, but even Apple says it won't scale linearly. And presumably when Apple did these tests, it had turned on beast mode on the 16".

Apple unveils game-changing MacBook Pro

Apple today unveiled the completely reimagined MacBook Pro powered by the all-new M1 Pro and M1 Max — the first pro chips designed for the Mac.

www.apple.com

Featuring the 16-core GPU in M1 Pro and the 32-core GPU in M1 Max, the 14-inch MacBook Pro transforms graphics-intensive workflows with:

• • Up to 9.2x faster 4K render in Final Cut Pro with M1 Pro, and up to 13.4x faster with M1 Max.
  • Up to 5.6x faster combined vector and raster GPU performance in Affinity Photo with M1 Pro, and up to 8.5x faster with M1 Max.
  • Up to 3.6x faster effect render in Blackmagic Design DaVinci Resolve Studio with M1 Pro, and up to 5x faster with M1 Max.

Compared to previous-generation 2.3GHz quad-core Intel Core i7-based 13-inch MacBook Pro systems with Intel Iris Plus Graphics, 32GB of RAM, and 4TB SSD.

With the 16-core GPU in M1 Pro and 32-core GPU in M1 Max, the 16-inch MacBook Pro offers faster graphics performance with:

• • Up to 2.9x faster combined vector and raster GPU performance in Affinity Photo with M1 Pro, and up to 4.5x faster with M1 Max.
  • Up to 2.5x faster render in Maxon Cinema 4D with Redshift with M1 Pro, and up to 4x faster with M1 Max.
  • Up to 1.7x faster 8K render in Final Cut Pro with M1 Pro, and up to 2.9x faster with M1 Max.

Compared to previous-generation 2.4GHz 8-core Intel Core i9-based 16-inch MacBook Pro systems with Radeon Pro 5600M graphics with 8GB HBM2, 64GB of RAM, and 8TB SSD.


14" M1 Max vs M1 Pro
13.4/9.2 = 1.46X
8.5/5.6 = 1.52X
5/3.6 = 1.39X

16" M1 Max vs M1 Pro
4.5/2.9 = 1.55X
4/2.5 = 1.6X
2.9/1.7 = 1.71X

 

[The Hardcard]

So apparently the new MacBooks have a "high powered mode", I wonder if that kicks in under mains power only or if that's a setting you check somewhere. Perhaps doing that gives you another nudge in performance.

Given the silicon real estate of CPU vs. GPU, especially on the Max, this may only be for GPU. I imagine normal mode can give the CPU all the power it needs. So many permutations of what the new chips can do.

 

[Doug S]

Why would someone go to the trouble of designing an awesome GPU with first-of-its-kind 512-bit LPDDR5 and then not proudly show it off with a AAA gaming title? It's like having the most beautiful chick in the world as your wife and NOT sleeping with her! Who else in the industry has a powerful GPU specifically not marketed for gaming?

People buying Macbook Pro aren't buying them for games. Showing them off with an AAA game would be like Intel demonstrating a 28 core Xeon by running Prime95.

There are applications out there like CAD that tax a GPU at least as much as games which Macbook Pro & Mac Pro buyers do run, though. That's who the intended audience for the GPU is.

The ARM on Mac thing makes it more clear why Apple left Imagination behind and went their own way. Apple knew they needed GPUs that could scale up to full blown workstation level graphics, and they weren't going to get it from a company doing GPUs for mobile devices and embedded uses like POS systems.

Since Apple started designing their own, in addition to the multichiplet scaling stuff they've been working on no doubt they've also had more than just gaming in mind - which might be the reason for some of the uneven performance gains reported on gaming focused benchmarks, since the improvements they've made aren't all targeted at what makes AAA games faster.

 

[moinmoin]

Why would someone go to the trouble of designing an awesome GPU with first-of-its-kind 512-bit LPDDR5 and then not proudly show it off with a AAA gaming title? It's like having the most beautiful chick in the world as your wife and NOT sleeping with her! Who else in the industry has a powerful GPU specifically not marketed for gaming?

I think the metaphor you use there is exactly the kind of association Apple would want to avoid at all cost in its presentations.

That aside Apple doesn't need to push anything gaming related, because it already earns more than any other company through games, even without developing any games or pushing anything gaming specific:

Apple Doesn’t Make Videogames. But It’s the Hottest Player in Gaming.

CEO Tim Cook quietly turned the iPhone company into a superpower in the videogame business. Now he’s fighting rivals in a multiplayer universe.

www.wsj.com

 

[Eug]

Geekbench 5 single-core / multi-core scores: 1785 / 12833

I have seen no scores over 13000.



Geekbench 5 Metal score: 69212

I have seen no scores over 70000.

 

[Doug S]

OK so now there seem to be enough scores to wonder why the M1 Max scales so poorly over M1 Pro. Is it a limitation in Geekbench, a limitation in macOS or the driver, or something in the hardware itself that inhibits the same scaling observed from M1 to M1 Pro?

If it is a real limit in something Apple controls they obviously need to address it before releasing the Mac Pro, as a further 4x scaling might translate into even less of a gain.

I wouldn't be shocked if it is a driver issue. After all, this is a lot more performance than Apple has ever had available before, so some immaturity in the driver at first would not be shocking. We've seen similar from release of a new GPU from AMD or Nvidia in the more than a few times, and they have a lot more experience in high performing GPUs than Apple has.

 

[eek2121]

Just something to think about, but MS and Qualcomm are still at least nominally pushing WARM, and there is the off chance that AMD may get into the mix through their semi-custom unit. How that helps gaming on Apple is dubious, but at least we aren't talking targeting an entirely different OS and uarch, but just an entirely different OS.

Not unless they (qualcomm, you say Microsoft, haven't researched, but I doubt Microsoft is pushing anything other than DirectX) manage to drum up a 9-10 figure gaming business. AMD isn't going to touch niche APIs of any kind unless they show staying power.

There's a thing called MoltenVK.

So apparently the new MacBooks have a "high powered mode", I wonder if that kicks in under mains power only or if that's a setting you check somewhere. Perhaps doing that gives you another nudge in performance.

Are you saying they lied on their own charts? They claimed "plugged in" performance was the same thing as "battery performance". My prediction: once the real reviews roll in, most of you guys will disappear or stick around to move goal posts. Note that I'm pretty vendor neutral here, but when I see marketing calling the shots, I know what is really going on...

OK so now there seem to be enough scores to wonder why the M1 Max scales so poorly over M1 Pro. Is it a limitation in Geekbench, a limitation in macOS or the driver, or something in the hardware itself that inhibits the same scaling observed from M1 to M1 Pro?

If it is a real limit in something Apple controls they obviously need to address it before releasing the Mac Pro, as a further 4x scaling might translate into even less of a gain.

I wouldn't be shocked if it is a driver issue. After all, this is a lot more performance than Apple has ever had available before, so some immaturity in the driver at first would not be shocking. We've seen similar from release of a new GPU from AMD or Nvidia in the more than a few times, and they have a lot more experience in high performing GPUs than Apple has.

It's a limitation in "Apple products". Geekbench has had limitations in the past, but the OpenCL test has proven to scale above and beyond (it is several years old at this point and scales above even the 3090). Apple has a slow GPU, slow drivers, or both (betting on the hardware. I know from experience). A bunch of you folks are expecting Apple to perform miracles. RTX 3060 mobile performance IS a miracle in a Mac Product! It will NOT set global performance records. It will set SOME, but not many, efficiency records.99% of the records it sets are due to NVIDIA being 2-3 node generations behind Apple. If NVIDIA had GPUs on TSMC 5nm, they would be ahead. This is a fact. I'm not an NVIDIA fan. I actually only own stock in AMD (and a small amount of Apple shares), NVIDIA is very competitive when it comes to perf/watt. There is a reason when AMD, on a superior process, has trouble beating NVIDIA. NVIDIA is a giant and can throw tons of money at R&D. Apple has a good first try, but with both CPU and GPU performance, most of you aren't seeing the bigger picture:

It is easy to build an okay product on a superior node and claim a win. What happens next for all will determine who wins or loses. Thus far, Apple has several losses, including software compatibility, lack of high refresh rate displays, lack of mainstream API support, etc. Their only wins IMO, are on their own, in house, operating system in a closed ecosystem. which I don't personally consider wins, because they don't change the mainstream market. They need to step up to the plate if they want to claim real wins. Neither Windows nor Linux (GUI applications) currently run on Mac hardware unless being emulated. Get back to us when they do. If their emulated solutions can outperform cutting edge native x86 solutions, or if they figured out how to get all the rest of the software in the world to run faster than it does on an x86 CPU, I'll be all ears. As of now, MY company is ditching Macs because the fastest server CPU is x86, not ARM.

 

[DrMrLordX]

Not unless they (qualcomm, you say Microsoft, haven't researched, but I doubt Microsoft is pushing anything other than DirectX) manage to drum up a 9-10 figure gaming business. AMD isn't going to touch niche APIs of any kind unless they show staying power.

WARM = Windows on ARM. You'd best believe DirectX is a part of that. Vulkan is a legit API target on that platform as well.

As for AMD, they've already been approached by their semi-custom partners about producing ARM products for various niches. That may include possible gaming products since their console APU division is wedded to their semi-custom unit, and has been since the PS4/Xbox One days. When it comes to semi-custom, they'll build whatever it is their customer wants so long as the bills get paid.

Furthermore, as @moinmoin pointed out, Apple's iOS app store makes Apple a ton of money off gaming.

Gaming on ARM is a thing. It's just not necessarily the traditional AAA gaming that we're used to around here. But at the same time, some of those games on ARM in the mobile arena are becoming increasingly complex.

 

[jeanlain]

It's a limitation in "Apple products". Geekbench has had limitations in the past, but the OpenCL test has proven to scale above and beyond (it is several years old at this point and scales above even the 3090). Apple has a slow GPU, slow drivers, or both (betting on the hardware.

Slow drivers? Doubtful. Apple writes Metal drivers for its own GPUs.
Slow GPU? Why wouldn't the M1 Max be ~2x faster than the M1 Pro? It has double the core and memory bandwidth.
I'd say it's either the low power mode messing with the test or an issue with geekbench. Keep it mind that the Apple silicon version of this tool has only been tested on M1, and it may be derived from the iOS version. Primate labs never got the opportunity to check that it scales well with many GPU cores.

 

[Semel]

There's a thing called MoltenVK.

It doesn't matter coz it lacks features available to DX12 and Vulkan that are used for modern game development.

 

[Eug]

Slow drivers? Doubtful. Apple writes Metal drivers for its own GPUs.
Slow GPU? Why wouldn't the M1 Max be ~2x faster than the M1 Pro? It has double the core and memory bandwidth.
I'd say it's either the low power mode messing with the test or an issue with geekbench. Keep it mind that the Apple silicon version of this tool has only been tested on M1, and it may be derived from the iOS version. Primate labs never got the opportunity to check that it scales well with many GPU cores.

1. I doubt Apple would have run the tests in low power mode.

2. In one of my previous posts I went through all of the GPU related performance benchmarks that Apple published in its press release, and calculated the scaling between M1 Pro 16 and M1 Max 32. Not a single GPU test published in Apple's press release - from Apple's own lab - scaled at near 2X. The range was 1.39X to 1.71X.

So, if anything, in this case Geekbench 5's Metal test is providing a reasonably accurate picture of M1 Pro 16 to M1 Max 32 scaling: ~69000 (Max) / ~42000 (Pro) = ~1.6X.

 

[The Hardcard]

1. I doubt Apple would have run the tests in low power mode.

2. In one of my previous posts I went through all of the GPU related performance benchmarks that Apple published in its press release, and calculated the scaling between M1 Pro 16 and M1 Max 32. Not a single GPU test published in Apple's press release - from Apple's own lab - scaled at near 2X. The range was 1.39X to 1.71X.

So, if anything, in this case Geekbench 5's Metal test is providing a reasonably accurate picture of M1 Pro 16 to M1 Max 32 scaling: ~69000 (Max) / ~42000 (Pro) = ~1.6X.

it is becoming clearer that the M1 Max is not set up to run wide open in laptops. The scores are the result of Apple-set wattage limits. We probably won’t see the full power of the M1 Max until it is in a desktop.

The 16 inch has he “high power mode“ which has to be manually activated in the system preferences. Even then, there’s no indication that high power mode in the 16 inch will represent the absolute maximum that the chip is capable of.

EDIT: At least the GPU. Since the M1 Max is really a GPU with an integrated CPU, that side probably won’t give more performance in ”high power “ or even desktops for that matter.

 

[Eug]

M1 Max PugetBench Premiere Pro



Ryzen 9 5950X 16-core with GeForce RTX 3080



It seems in terms of just the GPU score, the desktop 3080 does much better, but in terms of the overall score they are similar.

Part of the reason is due to what I was talking about earlier, which was the addition of hardcore hardware acceleration in M1 Pro/Max. The robust hardware ProRes acceleration in M1 Pro/Max is going to go a very long way in making this model a darling of video editors.

 

[lobz]

M1 Max PugetBench Premiere Pro

View attachment 51752

Ryzen 9 5950X 16-core with GeForce RTX 3080

View attachment 51753

It seems in terms of just the GPU score, the desktop 3080 does much better, but in terms of the overall score they are similar.

Part of the reason is due to what I was talking about earlier, which was the addition of hardcore hardware acceleration in M1 Pro/Max. The robust hardware ProRes acceleration in M1 Pro/Max is going to go a very long way in making this model a darling of video editors.

Because of how extremely well their driver is tuned to use and utilize the fixed function hardware during playback tests. Old console times.