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

 

[Leeea]

quoting your own post, 408gb/s

Page 36 - The anti-crypto thread

Page 36 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

That is just the memory bandwidth.

What we really want to see is the MH/s for Eth mining.

I have a feeling it is going to be 1/2 the MH/s of a 3080, but at 1/4 watts. Which would be exceptional.


-Leeea tries to catch the attention of the wolves: "hey, look over there, fresh meat!"-

 

[Hitman928]

That is just the memory bandwidth.

What we really want to see is the MH/s for Eth mining.

Someone's going to have to write some mining software for it, but I agree, at least on paper it should be very good at mining.

 

[Leeea]

Someone's going to have to write some mining software for it, but I agree, at least on paper it should be very good at mining.

-nods- exceptional. 2nd to none. This is really great news, gamers are going to be thrilled with the new m1 max.

 

[Hitman928]

SPEC usually does custom compilations of each software it benchmarks and tends to compile more for the least common denominators. For instance, in the x264/x265 tests they compile without AVX or SSE flags on x86 which is not at all how the software is actually used. This could paint a very different picture of video encoding performance as x86 kind of relies on those more modern instructions. Unless Anandtech is custom compiling the sub tests and turning these flags back on? Either way, some of the tests make it hard to really compare performance for real world apps when they're not testing the apps in the way they are actually compiled/run in the real world.

Going back and reviewing, I believe I mis-remembered the SPEC setup. I believe it is Anandtech that compiles with flags that turn off SSE for x264/x265. The actual flag is to disable ASM which makes sense as it is highly optimized for x86, but then the fallback is to not have any SSE/AVX. You'd have to check each tests compiler flags to see what is being used.

 

[insertcarehere]

Your point is valid, but from what I've seen, the 35W number in the Techspot review is with the CPU really drawing 35W. At 65W, Cezanne should score pretty close to 13,000 points in Cinebench r23 as shown in your link.

Not to belabor the point any further, but unless power was explicitly monitored (and it wasn't in the techspot review), then we have no idea what profile was actually used and therefore an accurate idea of actual power draw.

Computerbase found three power profiles for the ROG Flow X13 (AFAIK the only laptop with 5980HS), each of which exhibit clearly different behaviors, only one of which would likely really meet the criteria for 35w power draw.
https://www.computerbase.de/2021-02...chmarks_und_huerden_mit_35_42_und_80_watt_tdp

 

[simas]

I'd like to see more Rosetta benchmarks come out with the reviews. It seems like we saw a lot more of this with the M1. Obviously this wouldn't be ideal performance wise for the M1max, but at the same time it would actually tell us what kind of performance can be expected in those scenarios where native isn't supported and everyone should understand that it's not representative of architecture performance in those scenarios without accounting for the Rosetta penalty. It was actually pretty surprising with M1 how well it performed in most cases even with only a Rosetta path available.

I am with you- ideal/not ideal is irrelevant, real world is. Is this of any value to me who wants to play non-console AAA PC games? Does GamePass works on it (for game subscriptions and access to day one releases from anything Microsoft owned or contracted with)? How does it perform in say Cyberpunk which I still think has the best writing I have seen in last few years?


otherwise , it is just another Apple thingy - a real (and likely good) tool for very few people and use cases , an overpriced status symbols for vastly more wannabees and *yawn* for the rest of the PC enthusiast crowd. irrelevant, overpriced, overhyped, overmarketed and oversold gizmo.

As Apple stock owner (indirectly, through index funds), I strongly approve of anyone who wants to spend their money on it to go ahead and do it . Pay sales tax while at that , support your local government, even better.

as person spending my own money, this is irrelevant.

 

[albertmamama]

That is just the memory bandwidth.

What we really want to see is the MH/s for Eth mining.

I have a feeling it is going to be 1/2 the MH/s of a 3080, but at 1/4 watts. Which would be exceptional.


-Leeea tries to catch the attention of the wolves: "hey, look over there, fresh meat!"-

Eth mining is bandwidth-bound, so it's comparable from bandwidth alone.

 

[USER8000]

Laptop with a Ryzen 9 5900HS:

 

[BorisTheBlade82]

A 5800u at 15W limited TDP scores ~7500 points in MT for a perf/w score of 500. An M1max scores ~12400 at 34 watts for a perf/w score of 365. Clearly Zen 3 is more efficient than M1max. Additionally, for ST, an M1 chip uses just 3.8W to get the same score as M1max which uses 11W. That gives M1 a 2.9x performance efficiency advantage over M1max. They must have really screwed up the M1max design to do so bad in perf/w compared to M1. (/s)

It's easy to cherry pick data points with zero nuance/context and make whatever point you want to make.

I am not cherry picking. I am quoting Andrei Frumusanus general assessment of the competitive landscape regarding power and performance efficiency. Has it occurred to you that people like him are able to weight benchmark results not only based on some average means but also by their technical knowledge and experience? And has it occurred to you that you are the one cherry picking when he clearly states that CB is a negative outlier for the M1 Max?
The amount of ignorance you are showing tells me that you are not really interested in a grown-up discussion but instead will try to grasp at straws until the end of days.

 

[beginner99]

Apple's cores just are from totally different league than x86 rivals, and denial against it still lives strongly.

or putting it other way, x86 cpu manufacturers should be ashamed that they are beat so strongly by cpu from some lifestyle company which main target isn't selling cpus.....

Apple should be ashamed that we can't even find software to run workloads tailored for these cpus. Only half sarcastic. In data science it's very much either get an intel mac or ditch apple. The M1 isn't there yet or better said the software. And good luck if you need to do deep learning (albeit that didn't work before already)

Ditching backwards compatibility in a vertically integrated and locked platform will obviously be beneficial for performance. It's like consoles used to be. punching above their weight (here the weight is equal but the performance isn't). Neither intel or AMD can do that easily because the servers their chips need to run in often do run decades old software and need decades old instruction sets and use said instructions. Both also need to take into account the devices these things run in. their cheapest device isn't $2000. it's $300 laptops. Apples performance also directly translates to die size / silicon = cost.

 

[tamz_msc]

Going back and reviewing, I believe I mis-remembered the SPEC setup. I believe it is Anandtech that compiles with flags that turn off SSE for x264/x265. The actual flag is to disable ASM which makes sense as it is highly optimized for x86, but then the fallback is to not have any SSE/AVX. You'd have to check each tests compiler flags to see what is being used.

If you mis-remember then it is your duty to remember it properly and not make misleading claims.

The AMD 3rd Gen Ryzen Deep Dive Review: 3700X and 3900X Raising The Bar

www.anandtech.com

These are the flags used by Andrei:

-Ofast -fomit-frame-pointer
-march=x86-64
-mtune=core-avx2
-mfma -mavx -mavx2

 

[jeanlain]

Blizzard has been making games for the Mac since they started ages ago. Including World of Warcraft. So it should be a solid build.

Blizzard has been progressively abandoning the Mac.

I'd be very surprised if WoW included specific TBDR optimisations, using dedicated Metal APIs. My bet is that they simply converted their DX12 code into Metal with some automatic translation tool like everyone else does (including Feral), checked for bugs and performance issues, and recompiled the game to ARM.
All AAA games are coded for IMR GPUs, except perhaps Baldur's Gate III, which is being ported to the iPad.
In benchmark tools specifically designed to compare a wide array of devices, from mobile to desktop (e.g., GFXBench, 3DMark Wild Life), Apple GPUs perform much better.

But for the foreseeable future, Apple GPUs will underperform in AAA games, as these games will be coded with DX/Vulkan for IMR GPUs and ported to Metal as an afterthought (if they are ported at all).
We already had a large difference between performance on macOS and bootcamp on the same intel Mac. This difference will get larger, due the lack of optimisation for Apple in GPUs.

 

[jeanlain]

So in 526.blender_r and 511.povray_r, the M1 Max massacres the 5980HS, but in Cinebench R23 it . . . doesn't? Wouldn't it be nice of AT to run a current build of Blender on their M1 Max review sample instead?

The M1 has been underperforming on Cinebench, so these results are not surprising. My guess is that Cinebench favours X86 and is not as optimised for other ISAs.

 

[jeanlain]

Are you saying that M1 Pro/Max is less efficient than M1? If so, that would make sense since they only have half the efficiency cores of M1.

I think it's because a larger SoC simply uses more power, even if a single core is active.

 

[Gideon]

The M1 has been underperforming on Cinebench, so these results are not surprising. My guess is that Cinebench favours X86 and is not as optimised for other ISAs.

I'm not even quite convinced it's x86 but rather it just scales extremely well with SMT.

 

[Gideon]

While no blender there are at least some Handbrake transcoding results from Tom's Guide 14" macbook review (4K to 1080p):

4K to 1080p transcoding of a unspecified clip took:

• 4 min 48s - 14" Macbook Pro (M1 Max)
• 4 min 51s - 14" Macbook Pro (M1 Pro)
• 7 min 20s - Razer Blade 14 (Ryzen 9 5900HX)
• 7 min 46s - 13" 2020 MacBook Pro (M1)
• 8 min 10s - Dell XPS 15 (Core i7-11800H)
• 11 min 25s - Surface Laptop Studio (11 gen i5 - probably 11320H)
• 18 min 12s - Dell XPS 13 (Core i7-1185G7)

Those 11-gen XPS'es are pitted as mac competitors in look, form and quality and they are not cheap. With similar specs and the OLED screen the 15" is straight in the 16" mac pricing territory. The 13" is relatively a tiny bit cheaper, but just look at the performance.

Razer Blade 14 is pretty much the best x86 laptop currently available in the form-factor, especially for multithreaded work. It at least has a RTX 3080 that vastly outperforms the M1 Pro (and the Max often as well). The Dells usually have RTX 3050 Ti's

 

[Carfax83]

4K to 1080p transcoding of a unspecified clip took:

• 4 min 48s - 14" Macbook Pro (M1 Max)
• 4 min 51s - 14" Macbook Pro (M1 Pro)
• 7 min 20s - Razer Blade 14 (Ryzen 9 5900HX)
• 7 min 46s - 13" 2020 MacBook Pro (M1)
• 8 min 10s - Dell XPS 15 (Core i7-11800H)
• 11 min 25s - Surface Laptop Studio (11 gen i5 - probably 11320H)
• 18 min 12s - Dell XPS 13 (Core i7-1185G7)

Is that with hardware encoding enabled? I remember the original M1 was quite a bit slower than comparable x86 CPUs in software mode, which is why I am asking.

 

[Gideon]

Is that with hardware encoding enabled? I remember the original M1 was quite a bit slower than comparable x86 CPUs in software mode, which is why I am asking.

A good point, unfortunately the article is frustratingly light on details (not even the codec is mentioned).

Do you remember what reviews these were? Handbrake only added ARM support in version 1.4.0 released this July. Previously it must have ran using Rosetta 2 emulation.


EDIT:

Overall a good observation

This guy has a very detailed video comparing M1 Mini to a Intel Panther Canyon NUC (i7-1165G7) using both emulation and ARM native binaries running handbrake 10-bit x265 encode:

In his (totally unrelated and different test) it took:

• 7min 34s for the M1 Mac Mini (native)
• 7min 53s for the i7 Tiger Lake Panther Canyon NUC
• 10min 36s for the M1 Mac Mini (running under emulation)

So it is true that at least in 4K x265 encoding the difference between M1 and Core i7-11xx series is a lot closer than in the Tom's Guide transcoding results (though the XPS 13 is slim so it might also be throttling hard)

 

[igor_kavinski]

Interesting how the AT review states that Apple hasn't defined any max TDP for their SoC and it just goes as far as it can until thermal conditions prevent going any further. Someone needs to test an M1 Max in Alaska in minus temperatures. This also suggests that the Mac Pro with water cooling could be formidable.

 

[jeanlain]

I'm not even quite convinced it's x86 but rather it just scales extremely well with SMT.

I was referring to single-core scores. For instance, a tiger lake CPU scores higher than the M1 in single-thread cinebench runs, while it's the opposite in most other single-thread tasks (EDIT: 18 out of the 22 SPEC tests).

 

[igor_kavinski]

Someone's going to have to write some mining software for it, but I agree, at least on paper it should be very good at mining.

Yes, please miners, take the M1 family and scalp it to smithereens and leave our precious GPUs alone.

 

[Carfax83]

A good point, unfortunately the article is frustratingly light on details (not even the codec is mentioned).

Linustechtips, time stamped:

 

[DrMrLordX]

The one thing I don't get is all the people whining about the "lack of benchmarks", which as far as I can tell means "where are the game benchmarks".

Here's a review of an x86 mobile APU:

AMD Ryzen 9 5980HS Cezanne Review: Ryzen 5000 Mobile Tested

www.anandtech.com

Look at all of the benchmarks in that article. Now look back at their article on the M1 Pro/Max. See any differences?

 

[nickmania]

With this chip, Apple is giving the best option for filmakers, youtubers etc.... You can use one of the best video codecs, ProRes, and use your M1 system with renders and apply effects much faster than any other X86 computer, even desktops. They focus much more on this aproach instead of making a "good chip for everything". The render times in FInal Cut Pro are crazy and way ahead any X86 computer can make. So you need to use their codec, their computer and their software to obtain an increase in render times and effects. When you make the comparisions and said is not a good bench comparing render in premiere with renders in final cut, you need to understand that IS a good benchmark, if you take in consideration the user and the final result. The numbers are impressive.

Sure they are going to continue this approach in other sectors, designing the chips for specific tasks so they could sell the "whole package" a bit ahead of the x86 world, which is not going to be able to compete in the short term.

Yes, is overpriced but now they are selling something totally different than in the Intel era, and this time, it makes sense to buy one of this overpriced machines if it fit your workspace.

 

[igor_kavinski]

M1 Max GPU beats $6000 Radeon Pro W6900X in benchmark - 9to5Mac

Another boring benchmark won by M1 Max.

Review: 16-inch Apple MacBook Pro M1 Max for video editors - Part 1 by Scott Simmons - ProVideo Coalition

That's the sort of review that will have professional folks wetting their pants with sheer glee.