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

 

[Eug]

So in the, very unlikely, situation that it fails you go to AMZ and search for "ethernet USB-C adapter"?
And are back to where you would be without the power-brick scheme?

Some people just INSIST on finding flaws in a new way of doing things - we MUST maintain the ways of the ancestors no matter what...

It's always best to have it on board. People complain about randomly losing network connections with third party USB-based Ethernet adapters on Macs.

Coincidentally, elsewhere we were discussing the pros and cons of getting Gigabit Ethernet vs 10GigE on the Mac mini. One problem with getting Gigabit now and then USB 10GigE later (besides losing the USB port) is that reportedly many 10GigE adapters are flaky on Macs, and some of them actually overheat too. That said, USB Gigabit Ethernet seems to be more stable than USB 10GigE, without the overheating.

Anyhow, the real world replacement for Gigabit Ethernet on board from Apple's point of view seems to be WiFi, as evidence by Apple TV. They now have a WiFi/Ethernet model and a WiFi-only model.

Two days left.
Do we believe Gurman?
Or do we hope for the obvious way to structure and climax the whole event:

Tomorrow:
Studios with M4 Ultra announced.
Friday:
New Mac Pro with M4 Extreme announced, along with all the die shots.

No, today was the last day of Mac announcements. Apple said so itself.

 

[pj-]

I'll still use ethernet for gaming but for pretty much everything else, wifi is already far beyond good enough.

I get > 1gbps speed tests (on 2gbps service) on devices in the same room as my router, and hundreds of mbps in the rest of my apartment (urban environment with dozens of networks around and plaster w/ metal lath walls). Never have any connection instability

Edit: that being said, i am on the fence about an m4p mini, but if i do get one it'll have 10gbps for no real reason

 

[LightningZ71]

In general, USB Ethernet adapters in the 2.5+ Gbps range are cheaply made, lack proper thermal management, have flaky drivers, and typically undershoot their rated speeds by 50% or more. It really doesn't improve until you get to the formal USB/tb "docking stations."

And that's for ANY OS.

 

[Eug]

i am on the fence about an m4p mini, but if i do get one it'll have 10gbps for no real reason

I've given up on the idea of the M4 Pro, because it's a CA$600 price premium over M4. The main advantage I saw was future proofing with DisplayPort 2.1. However, I don't think I'll need DisplayPort 2.1 anytime soon. I'd be satisfied with 6K HDR 60 Hz plus HDCP 2.2, all of which the M4 mini can handle. I say "satisfied" but it's unlikely I'll buy even one of those anytime soon, since those 6K Retina monitors are hella expensive. Apple has the Pro Display XDR. Dell has one that is finicky on Macs, and Asus recently announced one but it isn't out yet.

As for Ethernet, the only reason I considered 10GigE was just because it's a cheap upgrade, but like you, I don't have any real use for it. My internet access is 1.5 Gbps, but my provider-supplied modem only has Gigabit ports anyway so it's moot. (I got 1.5 Gbps service only because it was actually cheaper on a promotion to get that instead of 500 Mbps.) And my router, switches, and NAS are all Gigabit. My in-wall cables are a mixture of CAT5e and CAT6. I have no CAT6a or CAT7.

In general, USB Ethernet adapters in the 2.5+ Gbps range are cheaply made, lack proper thermal management, have flaky drivers, and typically undershoot their rated speeds by 50% or more. It really doesn't improve until you get to the formal USB/tb "docking stations."

And that's for ANY OS.

Good to know. I was just reading about Macs using these, but it's sad to find out that they can't get it right on other OSes either.

 

[gdansk]

I had thought the notch would be transitory. Is there any indication when the MBP redesign will be, next year?

I think I can wait another year to replace my M1P MBP.

 

[mvprod123]

Apple rumor says it's considering buying Intel as chipmaker continues to struggle

Samsung is also in the mix

www.tomsguide.com

Lol

 

[gdansk]

Apple rumor says it's considering buying Intel as chipmaker continues to struggle

Samsung is also in the mix

www.tomsguide.com

Lol

MLID-sourced garbage. Probably not even worth posting.

 

[igor_kavinski]

MLID-sourced garbage. Probably not even worth posting.

The source may be garbage but the reasoning is sound. Apple, Qualcomm and Samsung are all competitors and any one of them making a successful bid for Intel and its foundries would gain a massive edge over the others. If Apple thinks there's the slightest chance that Samsung or Qualcomm could become serious market forces after acquiring exclusive access to Intel's foundries, it may be forced to open its coffers.

By the way, if any one of them deserves to get their hands on Intel fabs, it's Apple. Their close collaboration with TSMC has given them quite a bit of fab related knowledge and expertise and they would discipline the crap out of the Intel fab workforce or replace them with Taiwanese immigrants if they have to, to get the fabs back on track.

 

[oak8292]

The source may be garbage but the reasoning is sound. Apple, Qualcomm and Samsung are all competitors and any one of them making a successful bid for Intel and its foundries would gain a massive edge over the others. If Apple thinks there's the slightest chance that Samsung or Qualcomm could become serious market forces after acquiring exclusive access to Intel's foundries, it may be forced to open its coffers.

By the way, if any one of them deserves to get their hands on Intel fabs, it's Apple. Their close collaboration with TSMC has given them quite a bit of fab related knowledge and expertise and they would discipline the crap out of the Intel fab workforce or replace them with Taiwanese immigrants if they have to, to get the fabs back on track.

TSMC provides cutting edge technology to Apple and a host of fast followers. Apple is only a quarter of TSMC’s revenue. Apple would need to provide foundry services, just like Intel does to make the fabs cost effective. That isn’t where Apple is going to go.

The cost of nodes at this point in the game are bigger than any one company. Intel recognized that fact 10 years ago but did not adequately address it with foundry 1.0.

 

[The Hardcard]

I don't think we can assume that. The M4P / M4M cut line could include the extra 20 GPU cores, whatever additional SLC there may be, the additional memory controllers, and the additional encoder/display support. The extra 2 P cores the Max has? Well, about that...

There is something interesting going on with the M4 family P core cluster sizing. M4 has 4 P cores so obviously a cluster size of 4. M4M has 12 P cores, so could be three clusters of 4 or two clusters of 6. M4P has 10 P cores, is that a cluster of 4 and a cluster of 6? Two clusters of 5? Or is it three clusters of 4, and either two are binned on every M4P die or two fall on the chop line and are disabled, i.e. non functional because of part of them is missing?

That is an intriguing concept. I didn’t see it because the CPU on the M3 Max is so perfectly centered, but it doesn’t have to be. Especially if they shaped the L2 cache differently, it would be easy to put 2 cores over a cut line. It would neatly explain why the M4 Pro got so many more performance cores.

Unfortunately, it looks like we’ll have to wait for 3rd party die shots, but if an early buyer ran a core to core latency test (Geekerwan?) it would reveal a 3 cluster 6 + 4 + 4 setup. It would very unlikely Apple would layout lopsided performance clusters on a unique die.

I agree it is most likely a chop. If Gurman is thus vindicated then there won’t be any interconnect since him being right means the Ultra is a chip unto itself.

Watching the presentation Apple gave nods to its LLM and generative AI market. In the presentation video one of the screens showed LMStudio, which is a popular way to load and run LLMs. The LPDDR5x-8533 memory is going to boost those sales quite a bit. As will the Thunderbolt 5, since there is now an easy way to turn Macs into AI clusters.

I wouldn’t be surprised if a new Mac Studio would be released soon, even if the Ultra is not released, there would be multiple buyers of sets of 2 to 4. Clustered M4 Maxes are going to be a popular way to run Llama 3 405B.

Apple also made M4 Pro a viable budget machine learning platform. 273 GB/s bandwidth is very good and Thunderbolt 5 means there’s going to also be people buying multiple Mac Minis, since it’s yet another way to run 405B for less than $10,000. I’m not sure if there is a non-Apple way to run it without giving up a huge amount of speed.

 

[Eug]

That is an intriguing concept. I didn’t see it because the CPU on the M3 Max is so perfectly centered, but it doesn’t have to be. Especially if they shaped the L2 cache differently, it would be easy to put 2 cores over a cut line. It would neatly explain why the M4 Pro got so many more performance cores.

Unfortunately, it looks like we’ll have to wait for 3rd party die shots, but if an early buyer ran a core to core latency test (Geekerwan?) it would reveal a 3 cluster 6 + 4 + 4 setup. It would very unlikely Apple would layout lopsided performance clusters on a unique die.

I agree it is most likely a chop. If Gurman is thus vindicated then there won’t be any interconnect since him being right means the Ultra is a chip unto itself.

Watching the presentation Apple gave nods to its LLM and generative AI market. In the presentation video one of the screens showed LMStudio, which is a popular way to load and run LLMs. The LPDDR5x-8533 memory is going to boost those sales quite a bit. As will the Thunderbolt 5, since there is now an easy way to turn Macs into AI clusters.

I wouldn’t be surprised if a new Mac Studio would be released soon, even if the Ultra is not released, there would be multiple buyers of sets of 2 to 4. Clustered M4 Maxes are going to be a popular way to run Llama 3 405B.

Apple also made M4 Pro a viable budget machine learning platform. 273 GB/s bandwidth is very good and Thunderbolt 5 means there’s going to also be people buying multiple Mac Minis, since it’s yet another way to run 405B for less than $10,000. I’m not sure if there is a non-Apple way to run it without giving up a huge amount of speed.

I don’t think they will release the Mac Studio until the M4 Ultra is available.

BTW, where do you guys think the biggest benefits of the increased memory bandwidth will be, between M4 and M4 Pro?

 

[eek2121]

The doubling of system memory is welcome, and Apple finally now wants to sell me a Mac Mini. They must be getting desperate! 🤣

 

[mvprod123]

 

[lopri]

Normally I would not pay $100 extra for a 10 Gbe NIC in this day and age but I will make an exception for this one.

 

[Eug]

Normally I would not pay $100 extra for a 10 Gbe NIC in this day and age but I will make an exception for this one.

Just curious, why?

In my case, it turned out 10 GigE would have meant a 2 week delay so it solidified decision not to get it. I don't have ANY equipment above Gigabit anyway.



This Canadian edu order is CA$1394, which is currently US$969.69. For edu in the US, it's US$1079, or about 11% more expensive.

Regular retail in Canada is CA$1549 (US$1113.46) or US$1199 in the US, about 8% more expensive.

 

[johnsonwax]

The source may be garbage but the reasoning is sound. Apple, Qualcomm and Samsung are all competitors and any one of them making a successful bid for Intel and its foundries would gain a massive edge over the others. If Apple thinks there's the slightest chance that Samsung or Qualcomm could become serious market forces after acquiring exclusive access to Intel's foundries, it may be forced to open its coffers.

By the way, if any one of them deserves to get their hands on Intel fabs, it's Apple. Their close collaboration with TSMC has given them quite a bit of fab related knowledge and expertise and they would discipline the crap out of the Intel fab workforce or replace them with Taiwanese immigrants if they have to, to get the fabs back on track.

I don't buy this rumor. The issue over Intel fabs is going to be domestic production of ITAR components. In the past the DOD was willing to forgo leading nodes because their needs were fairly modest, but the move into AI based aircraft and the like is going to necessitate a domestic provider - why Biden and Congress have been pushing for domestic investment. Apple is never going to be a defense contractor and I think it's extremely unlikely that the US government is going to allow Apple to buy their fabs without a commitment toward contract production. My sense was always that the feds would work with a bunch of domestic companies, including Apple, to tip up a ULA-like consortium among domestic partners. It'd retain US ownership and DOD would have a contractor they can work with for cutting edge silicon.

Apple isn't leaving TSMC. When Apple can own and leverage the IP, they'll vertically integrate, but in the case of semiconductor manufacturing that isn't really viable for them, despite the experience they've gained. Not only is the IP really challenging to get right (witness Intel) but with so few players in a critical market, Apple's ability to keep it for themselves is questionable. In that situation, they're best off having a market of companies they can contract with. The question becomes whether such a market is sustainable. Samsung is still an option. Intel may be depending on who owns it (and Apple having an ownership stake in it would suffice). Even with an ownership stake, Apple could keep choosing the foundry with the most advanced IP for their A/M series, and hand off the less critical stuff (S, H, etc.) and the support silicon to this entity. Apple has no shortage of that stuff to make. It would give Apple options.

Note, Apple never bought Foxconn or any of the other assemblers. They never bought panel production, but did invest heavily in various partners. These were not industries where there were IP benefits because they commoditize too quickly. Semiconductor production is the same way, and the cost to be a cutting-edge fab is likely too high even for Apple to sustain without having customers - and Apple doesn't make components for anyone - ever. That may change, but it's been one of the most reliable rules when knocking down these kinds of rumors. Apple has never been interested in being a supplier for anyone other than consumers.

 

[StinkyPinky]

That mac mini is $500 really, I don't think Apple even checks if you have an edu email. I know I've bought from their education store with zero checks. They also have a 10% discount for gov and mil email addresses also for everyones information.

$500 for that is a steal, no doubt.

 

[Eug]

That mac mini is $500 really, I don't think Apple even checks if you have an edu email. I know I've bought from their education store with zero checks. They also have a 10% discount for gov and mil email addresses also for everyones information.

$500 for that is a steal, no doubt.

In Canada and the US, they do not check edu purchases (anymore). In some other countries they do check. However, that US$499 edu Mac mini (or US$481 edu in Canada) is with 256 GB storage for those here that don't already know. (512 GB is currently the sweet spot for me, but I'm getting 1 TB anyway. )

BTW, if someone still wants to buy retail, Apple probably will have a gift card promotion for Black Friday, and maybe Amazon and Best Buy, etc. will have Black Friday deals too.

 

[lopri]

Just curious, why?

In my case, it turned out 10 GigE would have meant a 2 week delay so it solidified decision not to get it. I don't have ANY equipment above Gigabit anyway.

View attachment 110625

This Canadian edu order is CA$1394, which is currently US$969.69. For edu in the US, it's US$1079, or about 11% more expensive.

Regular retail in Canada is CA$1549 (US$1113.46) or US$1199 in the US, about 8% more expensive.

I got sick of motherboard vendors' nickel-n-diming. But perhaps you misunderstood my previous comment. I would not mind paying for Mac Mini with a 10 GbE NIC.

 

[Eug]

I got sick of motherboard vendors' nickel-n-diming. But perhaps you misunderstood my previous comment. I would not mind paying for Mac Mini with a 10 GbE NIC.

Oh OK. Are you already able to use 10 Gbps Ethernet already in your setup? Just wondering.

Actually I mixed up your post with @pj-'s post. He said he'll get 10 Gbps Ethernet but has no real reason to do so.

i am on the fence about an m4p mini, but if i do get one it'll have 10gbps for no real reason

 

[jdubs03]

I was checking this out a bit more earnestly and it’s tempting but I’m holding off on the Mac Mini, perhaps going for the M5 version. The M4 would be a massive upgrade over my mini PC (i7-8700T), but I’d like at least Wi-Fi 7, and a base model of 512GB so I can then go up to the 1TB.
Who knows maybe after next week I’ll change my mind.

For what I use my PC for though, it does what it needs to.

I just want a MacBook Pro with FaceID, I really hope they have it for the M5 next year with their redesign. If they can fit it in the iPad Pro, it shouldn’t be that much for a stretch to put it into the top of the display.

 

[Eug]

I was checking this out a bit more earnestly and it’s tempting but I’m holding off on the Mac Mini, perhaps going for the M5 version. The M4 would be a massive upgrade over my mini PC (i7-8700T), but I’d like at least Wi-Fi 7, and a base model of 512GB so I can then go up to the 1TB.
Who knows maybe after next week I’ll change my mind.

For what I use my PC for though, it does what it needs to.

No guarantee the Mac mini will get M5. It may, or may not. Apple has a habit of skipping chip generations on its desktops. Apple's priority is the laptops. This was true for Intel, and also true on Apple Silicon:

iMac skipped M2
Mac mini skipped M3 series
Mac Studio skipped M3 series
Mac Pro skipped M3 series

P.S. I gave my son my i5-7600 27" iMac a couple of years ago. The good news with these old machines from my perspective is they have hardware h.265 HEVC decode support. M3 & M4 bring hardware AV1 decode support.

 

[Doug S]

Unfortunately, it looks like we’ll have to wait for 3rd party die shots, but if an early buyer ran a core to core latency test (Geekerwan?) it would reveal a 3 cluster 6 + 4 + 4 setup. It would very unlikely Apple would layout lopsided performance clusters on a unique die.

It should be fairly easy for someone with an M4P to establish this based on whether it has two or three SME units. If it has 3x the SME throughput of a base M4, then it has three P clusters. Conveniently, @Eug has ordered one, so if he can test his max SME throughput and compare with someone who has a base M4...

 

[Eug]

It should be fairly easy for someone with an M4P to establish this based on whether it has two or three SME units. If it has 3x the SME throughput of a base M4, then it has three P clusters. Conveniently, @Eug has ordered one, so if he can test his max SME throughput and compare with someone who has a base M4...

I ended up ordering M4, not M4 Pro.

M4 Pro would have cost 43% more with the same memory and storage. Fortunately, M4 has the 5 USB-C ports I wanted.

 

[The Hardcard]

It should be fairly easy for someone with an M4P to establish this based on whether it has two or three SME units. If it has 3x the SME throughput of a base M4, then it has three P clusters. Conveniently, @Eug has ordered one, so if he can test his max SME throughput and compare with someone who has a base M4...

That wouldn’t give the per cluster core breakdown though. I don’t doubt there are 2 performance clusters no matter what. If the M4 Pro was a unique die it would likely be two 5 core P clusters. It is not impossible for them to be designed lopsided, but that seems very unlikely.

I want a test that exposes the number of cores per cluster. Showing the 6 + 4
+ 4 is what would give strong argument for the Pro being a chop of the Max in lieu of die shots. The of course, we still anxiously await those die shots to see how they laid it out.

EDIT: I was looking at the M3 Max die shot and realized that they could rotate the clusters and cut one core from each. In fact, that could be easier. So it maybe 5 + 5 and still be chopped. We may just have to wait for die shots.