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

 

[nxre]

AT got review units for M1, so maybe they got them too for M1 Max/Pro which could mean we get a review as early as tomorrow depending on when embargo lifts.

 

[Pix12]

We need AT analysis and SPEC benchmark to confirm the performance.

Don't know why nobody (or almost) but AT makes a serious analysis of performance with proper benchmarks.

Yeah, absolutely agree.

 

[Doug S]

Interestingly, A15 in the iPhone 13 Pro & Pro Max now support ProRes video recording too. Not sure how much that helps iPhone users though, since it takes up 6 GB per minute. However, I guess that means A15/A16 based M2 will also support ProRes acceleration, something M1 doesn't have.

A lot of amateur/indy filmmakers shoot with iPhones. 6GB per minute isn't an issue as that's a good day's shooting for a 1 TB phone but having video shot directly in the format they use for editing is a big win for them.

It isn't a feature for the average consumer, or even the prosumer, but there is a niche market that was over the moon with joy at this announcement.

 

[Doug S]

I think Apple is fine with not having the absolute performance crown. Macs never sold based on that, and Apple is investing a lot of die space on accelerators that make the Mac the best machine, desktop or mobile, for certain workloads, instead of a one-size-fits all approach.
Still, the results are extremely competitive for a year-old architecture with no clock increases versus Intel biggest architectural change in a decade. I was actually expecting an even bigger gap, but I guess these are engineering samples that will still get tweaked and ADL will score higher when it hits consumers.

Exactly. For years people have been slagging on Mac because they're sold at premium prices but don't use the fastest possible Intel chip available at the time (though that's limited by TDP class or available SKUs in the socket type they need) and then don't update them for a couple years so they fall further behind PC laptops.

Apple shipping something that's faster than anything Intel has, even if a benchmark shows it getting beat by Intel a week later (though still not coming close to Apple's power efficiency) is a big change.

I've long maintained that Apple would wait to go ARM on the Mac until they could convincing beat Intel x86 performance for the same class of CPU being replaced in a given Mac model. What was necessary for the Mac ARM transition to be successful was to show people that they not only wouldn't be giving up performance by Apple switching from Intel, but there would be a meaningful increase in both performance and efficiency.

By that standard they've scored a home run with everything so far, with only the very high end Mac Pro class of machine left to do. Being able to beat Intel's fastest 'K' class CPUs that Apple never used in their products due to their higher power draw / cooling requirements would be sweet but was not required for a successful transition.

 

[DrMrLordX]

We need AT analysis and SPEC benchmark to confirm the performance.

Don't know why nobody (or almost) but AT makes a serious analysis of performance with proper benchmarks.

This only matters when you can run the same software as the competition, which is one of the reasons why you see so many GB5 and (to a lesser extent) SPEC scores. At least now we can use CinebenchR23 too, and Blender.

 

[biostud]

Doesn't Apple simply need to be good at everything, and then excel (destroy competition) in content creation to be a success? Nobody is going to to buy this primarily as a gaming rig, but obviously it will be nice if it can handle games as well.

 

[Doug S]

Doesn't Apple simply need to be good at everything, and then excel (destroy competition) in content creation to be a success? Nobody is going to to buy this primarily as a gaming rig, but obviously it will be nice if it can handle games as well.

Apple's primary goal is to take care of the existing Mac customer base. A lot of that is content creation, that's why they spend so much time talking about that and not talking about frame rate in the AAA game of the moment or about how quickly Excel calculations complete.

Once Apple feels that the existing Mac customer base is satisfied with the transition then they can see if there are opportunities to expand the Mac customer base, either by capturing parts of their existing markets they don't currently have (i.e. people doing content creation work on PCs rather than Macs) or by gaining customers in markets not traditionally viewed as strengths on the Mac platform. I have to think games would be well down that list, since the problem there isn't about convincing gamers to buy a Mac - first they would have to convince game developers it is worth porting to Metal on a "build it and they will come" hope.

I think the Mac has plenty of potential to grow its market share while continuing to completely ignore the gaming market, so I don't see Apple changing its policy of not caring about gaming on the Mac anytime soon. If Apple did care about convincing game developers to port, increasing the Mac's market share is more likely to succeed than anything else they could do, so ignoring the gaming market in favor of other markets might in a roundabout way be the best way to make that happen!

 

[igor_kavinski]

They also have Apple Arcade subscription service but it is more geared towards casual gaming.

 

[Gideon]

BTW there is a very nice website to watch for the state of M1 Mac gaming:

M1 compatible games master list - Gaming on M1 Apple silicon Macs and MacBooks, bugs, fixes, compatiblity and troubleshooting guides - AppleGamingWiki

www.applegamingwiki.com

Overall the state is pretty bad. Hugely fragmented and almost no new AAA games actually run - unless they have metal implementations (in which case they actually do work very well).

What does work quite well are:

• Native titles and those originally ported to x86 macs (in 64 bit versions only!)
• Many indie-titles in Steam work well in both rosetta or native.
• Older windows games (DX-11 and earlier) either in CrossOver (Wine) or Parallels (VM) for 32 bit games Parallels is the best pick

Just to prove the point. Here is a list of latest games used in Techpowerup GPU reviews:

Game	Method of running	Runs well	Runs at all	Unsupported API?
Assassins Creed Valhalla	-	0	0	DX12
Battlefield V	-	0	0
Borderlands 3	rosetta	1	1
Civilization IV	rosetta	1	1
Control	parallels	0	1
Days gone	-	0	0
Cyberpunk 2077	-	0	0	DX12
Deathloop	-	0	0	DX12
Death Stranding	-	0	0	DX12
Detroit Became Human	-	0	0	Vulkan
Divinity Original Sin II	rosetta	1	1
DOOM Eternal	-	0	0	Vulkan
F1 2021	-	0	0	DX12
Far Cry 5	-	0	0
Far Cry 6	-	0	0	DX12
Gears 5	-	0	0	DX12
Hitman 3	-	0	0	DX12
Metro Exodus	rosetta	1	1
Red Dead Redemption 2	-	0	0	DX12
Resident Evil Village	-	0	0	DX12
Sekiro: Shadows Die Twice	Wine	1	1
Shadow of Tomb Rider	rosetta	1	1
Star Wars Squadrons	-	0	0
Wither 3	-	0	0
Watch Dogs Legion	-	0	0
Total (out of 25)		6	7

• So only 6 out of 25 run at all
• 5 run actually well enough - just install via steam. Interesting is that none are actually ARM native! (BG3 is pretty much the only high-profile ARM native game)
• 1 only runs when using wine, but runs well (Sekiro)
• 1 runs barely in a a VM (Control)

Luckily older games (DX9-DX11) are usually better supported.

The most obvious problems are: lack of DX12 support in Parallels. If they somehow manage to enable this, things would probably go from bad to passable (though that alone is no guarantee 4 DX11 capable titles in the list don't run).

And of course the self-inflicted wound of not supporting Vulkan. The latter can be bypassed by devs using MoltenVK or other such tools, but that is rarely a plug-and-play solution in practice and requires plenty of dev resources.


TL;DR - Gaming on ARM Macs:

Legacy/indie titles should work (let's say at least half of the time). Latest AAA games fail to run 4 times out of 5. If the title only supports DX12, it's a guaranteed fail.

 

[moinmoin]

Regarding games I don't think Apple will make any more effort than pushing universal apps as the solution of covering all their platforms:

 

[Gideon]

Regarding games I don't think Apple will make any more effort than pushing universal apps as the solution of covering all their platforms:

Probably. I know Apple doesn't care about gaming and their users don't game at all, other than their favorites from other platforms that just happen to run in macOS as well.

I just really wish they did something Vulkan in some limited manner. Pretty much as they do with Rosetta 2 and x86 support. A separate executable you have to download that comes with an EULA that essentially says that Apple has no commitments, etc ...

That would allow them to:

• Only support M1 Macs - essentially only one generation of GPUs and no headaces for iOS
• Instantly allow to be 9/10 as good in gaming as Linux (as Proton could then essentially be ported to MacOS).
• All of this would actually put a dent in Windows exclusivity to gaming and motivate devs to support Vulkan vs DX12

Metal and Vulkan are similar enough that such a product shouldn't be that big of a deal (for Arm Macs only). I'm sure most of it could even be implemented as a side-project by Apple's gaming-enthusiast devs, if they only gave a green light.

 

[igor_kavinski]

Metal and Vulkan are similar enough that such a product shouldn't be that big of a deal (for Arm Macs only). I'm sure most of it could even be implemented as a side-project by Apple's gaming-enthusiast devs, if they only gave a green light.

Apple doesn't even have to do anything.

Portability benchmark of Dota2 on MacOS - gfx-rs nuts and bolts

It's up to the developers to target MoltenVK. Epic Games' Tim Sweeney would have likely extolled the virtues of the M1's GPU with his Unreal Engine but Apple decided to alienate him through litigation in the Epic Games v. Apple case (I know Epic sued them but Apple failed to see the bigger picture and refused to settle).

 

[Shivansps]

Because Apple is Apple, they dont support Vulkan because no one would use Metal anymore and supporting Vulkan also means supporting development that could be used on diferent platforms.

Still, MoltenVK can be used but thats petty much a fancy wrapper, so perf loss is expected, to support DX12 means implementing VKD3D on top of MoltenVK, thats two wrappers to get the game running.
Steam can get this done by porting Proton.

 

[igor_kavinski]

Steam can get this done by porting Proton.

I hope they get around to doing that sooner than later. There's no love lost between Valve and Microsoft.

 

[Gideon]

Aaaand reviews!

Apple's M1 Pro, M1 Max SoCs Investigated: New Performance and Efficiency Heights

www.anandtech.com

 

[bigggggggg]

Started reading the review, first things to notice: oh my god, this thing is stupidly fast

edit: 11% faster than a desktop 5900X for floating point workloads. It is just incredible, i have no words.
Maybe i should regret my comments about the 12900hk being faster than this

 

[Gideon]

And there were people already claiming that the power-efficiency is nothing to write home about, well how about that?



In Cinebench 23 M1 has 5% worse ST score and 3-4% MT score than a 11980HK. For ST workload the package draws 3x less and the whole machine 4x less power. For MT workload the difference is 2.5x.

 

[tamz_msc]

And there were people already claiming that the power-efficiency is nothing to write home about, well how about that?

View attachment 51890

In Cinebench 23 M1 has 5% worse ST score and 3-4% MT score than a 11980HK. For ST workload the package draws 3x less and the whole machine 4x less power. For MT workload the difference is 2.5x.

Cinebench isn't a very relevant benchmark. Look instead at gcc and bwaves in SPEC. The former is a typical integer workload and the latter is a typical floating point workload. The perf/W advantage is 4x and 6x respectively in those two workloads compared to Tiger Lake + 10nm SF.

Part of the lead in bwaves is no doubt due to the ~5x effective memory bandwidth compared to what's available on x86 competition, but people who say node advantage in favour of Apple should finally shut up.

There's no company providing server-class memory bandwidth in a laptop.

 

[Eug]

Well, this may explain the discrepant clock speeds reported by Geekbench:

The CPU cores clock up to 3228MHz peak, however vary in frequency depending on how many cores are active within a cluster, clocking down to 3132 at 2, and 3036 MHz at 3 and 4 cores active. I say “per cluster”, because the 8 performance cores in the M1 Pro and M1 Max are indeed consisting of two 4-core clusters, both with their own 12MB L2 caches, and each being able to clock their CPUs independently from each other, so it’s actually possible to have four active cores in one cluster at 3036MHz and one active core in the other cluster running at 3.23GHz.

--

GPU runs at 1.3 GHz. This is not being reported correctly by Geekbench.

--

CPUs cannot max out memory bandwidth on their own. CPU memory scaling maxes out at 243 GB/s on M1 Max. M1 Pro can max out the 204 GB/s though.

Andrei can get the GPU to use up to 90 GB/s.

This allows extra memory bandwidth overhead for the rest of the SoC.

--

M1 Max idle power 0.2 Watts! <--- This is my favourite performance stat of the M1 Max of the entire review. Holy frickin' wow! Wall power of the entire 16" MacBook Pro at idle with screen turned down to minimum was 7.2 Watts.

Aztec High Offscreen + 511.povray 92 Watts. (In comparison, i9-11980HK uses 220 Watts.)

--

@Doug S was right about the CPU scaling at 1.6X theoretically.

Because the new M1 Pro and Max have 2 less E-cores, just assuming linear scaling, the theoretical peak of the M1 Pro/Max should be +62% over the M1. Of course, the new chips should behave better than linear, due to the better memory subsystem.

However, memory makes a huge difference so in some workloads the scaling is much higher due to the higher memory bandwidth, with some scores scaling at >2X.

The 2020 Mac Mini Unleashed: Putting Apple Silicon M1 To The Test

www.anandtech.com

--

GFXBench Aztec High Offscreen for M1 Max scales at 4X M1, and M1 Pro scales at 2X M1. Perfect scaling. Same goes for 3D Mark Wild Life Extreme.

Gaming not great though, not surprisingly.

--

Andrei agrees with my comments about the the ProRes/ProRes RAW acceleration:

To further improve content creation, the new media engine is a key feature of the chip. Particularly video editors working with ProRes or ProRes RAW, will see a many-fold improvement in their workflow as the new chips can handle the formats like a breeze – this along is likely going to have many users of that professional background quickly adopt the new MacBook Pro’s.

--

Overall, this solidifies my resolve to buy the lowest M1 Pro SoC available in the Mac mini next year, or else the M2, as long as it has enough ports. The M1 is already twice as fast as my current primary machine, a Core i5-7600, and I'm actually buying this for a secondary machine.

An 8-core (6+2) M1 Pro machine would be way more performance than I need, and M2 machine at say 10% faster than an M1 machine would also be more performance than I need. But I would prefer more ports than in the current M1 Mac mini.

 

[Qwertilot]

All mildly amazing, yes.

One thing very gently intriguing me - shouldn't be there some GPU accelerated compute tasks where the massive memory pool and the shared memory between CPU & GPU (+ other bits) mean it gets even sillier? cf Linus Torvald talking about APU's having some distinct advantages, years ago now.

Maybe that is mostly seen in power savings.

 

[Eug]

sixcolors: 14-inch MacBook Pro review: A Mac Pro in your backpack



The coder geeks here will like that compile time. It's twice as fast as a 3.2 GHz 8-core Xeon W-2140B iMac Pro.

---

BTW, Andrei responded to someone about why Geekbench isn't scaling GPU scores so well on M1 Max. This is second hand information, but apparently Andrei thinks the Geekbench tests are too short, and the M1 Max GPU is not ramping up quick enough on such a short test.

 

[insertcarehere]

Looks like any benchmark & workflow that is even somewhat bandwidth-sensitive will absolutely clean house on the M1 Pro/Max. Which in practical terms, should bode well for the sort of professional users which need to manipulate/edit/playback tons of high-resolution media at a time.

This crazy level of engineering work (512 bit LPDDR5 on an SoC!) almost justifies the crazy money Apple is charging for the Pros, almost.

 

[Eug]

M1 Pro review:

He says it takes about 10 minutes of full CPU load (as in Cinebench R23) to get the fans to become audible.

 

[Hitman928]

And there were people already claiming that the power-efficiency is nothing to write home about, well how about that?

View attachment 51890

In Cinebench 23 M1 has 5% worse ST score and 3-4% MT score than a 11980HK. For ST workload the package draws 3x less and the whole machine 4x less power. For MT workload the difference is 2.5x.

I don't think anyone was claiming the power efficiency wouldn't be really good, it's just that a particular poster was claiming 500% better than the most efficienct x86 competition which is bollocks and these new tests show as much. Also, everything looks super power efficient compared to Intel's chips .

 

[Cali3350]

The Anandtech article notes that the max CPU clock of the performance cores in the M1 PRO is dictated by the number of active cores in a 4 core cluster.

Apple sells a M1 Pro binned chip with 8 cores (6 performance + 2 efficiency). Do you believe they force the binning such that only 1 performance core can be disabled from each cluster or no? It seems it could matter (however slightly) as you could have 2 cores at 3132 mhz on a cluster with both other cores disabled and 4 cores at 3036 on the other, vs a 1 + 1 disable where all cores would run at 3036.
Very slight difference, but it would be a difference in a chip Apple is selling as one SKU, which I am not familiar with happening before?