Discussion Apple Silicon SoC thread

Eug · Nov 10, 2020

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

mvprod123 · Oct 29, 2024

FlameTail said:
So do M4 and M4 Pro have identical display support capabilities?

igor_kavinski · Oct 29, 2024

gdansk said:
WiFi 6E and the power button being on the bottom are a few odd choices.
Guess you can't have everything!

jdubs03 said:
I saw there is a 12-core CPU for the base M4 Pro. So assuming that’s 8P+4E.

Disappointing that it’s only WiFI 6E but 10 gigabit Ethernet I suppose is the trade off there.

The Hardcard said:
Disappointing, somewhat. Weirdly to me, the M4 Pro core layout makes both the chop, fusion, and separate die possibilities all feasible.

'Tis the year of disappointment!

Where's the jubilation, folks? Apple seems to be providing more performance per dollar than AMD/Intel/Qualcomm, if you are married to Apple store.

poke01 · Oct 29, 2024

igor_kavinski said:
'Tis the year of disappointment!

Where's the jubilation, folks? Apple seems to be providing more performance per dollar than AMD/Intel/Qualcomm, if you are married to Apple store.

I’m happy, it’s very good value over the previous mini and M4 Pro is great. I didn’t expect 10P+4E at all.

poke01 · Oct 29, 2024

I ordered the base model for $485US ($741AUD) excluding tax using EDU discount. It will be a server NUC. The base being 16GB and the full M4 is what me pull the trigger.

igor_kavinski · Oct 29, 2024

poke01 said:
I ordered the base model for $485US ($741AUD) excluding tax using EDU discount.

Helluva deal!

Got a discount for me too?

perry mason · Oct 29, 2024

igor_kavinski said:
'Tis the year of disappointment!

Where's the jubilation, folks? Apple seems to be providing more performance per dollar than AMD/Intel/Qualcomm, if you are married to Apple store.

Ah, the "married to Apple store" trope. You are aware that there are hundred if not thousands of available OS X programs and tools, even open source, that one can download outside of any "App Store." Not that the App Store, which is far and away the most prolific app store among its competitors, should really be called any sort of walled garden anymore.

The Hardcard · Oct 29, 2024

mvprod123 said:
I don't believe in the 20+8 configuration for the M4 Max. Despite the power efficiency of Apple's silicon, such a chip will have too high a TDP for a laptop chassis, especially for a 14-inch model. And the die size will be too large. I'm already curious to see the die shots of the M4 Pro/Max dies. Apple is silent about the number of transistors. Apparently, this information with die shots will be announced tomorrow.

Die size wouldn’t be an issue, that core count would only be for a Max as 2 fused Pro dies. The 14-inch not being able to run the Max full bore I think is a given no matter what. The 16-inch easily handles the Max.This wouldn’t be a problem.

I also want to see die shots. In any way it will be yet another product adjustment for Apple. M1 -> M2 was the only straightforward update. M3 series completely reworked the layouts and obviously M4 is another full rework.

The Hardcard · Oct 29, 2024

igor_kavinski said:
'Tis the year of disappointment!

Where's the jubilation, folks? Apple seems to be providing more performance per dollar than AMD/Intel/Qualcomm, if you are married to Apple store.

To be sure, my disappointment concerns the multi-day announcements. The M4 is great and the M4 Pro is phenomenal. Apple opened up a huge gap on all M4 Pro competitors. It has far more grunt than top end laptop chips outside of the lower volt desktop-chip-in-a-laptop variants. And M4 Max looks likely to handle all those HX’s.

I am confident I’ll be happy tomorrow. Just got to make it through today. Seriously looking forward to 550 GB/s memory in a laptop. I think AMD was too conservative with planning Strix Halo. You’d have to just hate Macs and/or Apple to pick it over the M4 Max because it ain’t going to be that cheap.

igor_kavinski · Oct 29, 2024

The Hardcard said:
I think AMD was too conservative with planning Strix Halo.

It's an experiment I think. The sales volume will tell them if they should take it more seriously.

CouncilorIrissa · Oct 29, 2024

10 P-cores on the Pros is obscene. I wonder how those who bought the M3 Pros are now feeling...

Doug S · Oct 29, 2024

mvprod123 said:
I don't believe in the 20+8 configuration for the M4 Max. Despite the power efficiency of Apple's silicon, such a chip will have too high a TDP for a laptop chassis, especially for a 14-inch model. And the die size will be too large. I'm already curious to see the die shots of the M4 Pro/Max dies. Apple is silent about the number of transistors. Apparently, this information with die shots will be announced tomorrow.

Yep, I was gonna say the same thing. What's the point of putting that many CPU cores in a laptop when they'd quickly throttle if you tried to use them all? A Max built from two Pros is a non starter.

Well...unless they were both binned Pros with some CPU, GPU, SLC, TB, MC cut out. If they bin on enough blocks to account for an appreciable portion of the total die size they could fuse two dies and get something that's ~ 50-75% faster than Pro and recover as many as half the defective Pro dies. I highly doubt that's what happens but it is interesting to consider. I'll bet it is either a return to the chop or a separate die.

name99 · Oct 29, 2024

Eug said:
That's probably not the reason. I think the reason is because the iMac is only 11.5 mm thick, and that includes the screen and shell.

The RJ45 connector's depth may be the problem. You'll note that they don't even have the headphone jack on the back because a headphone plug is too long. So, they stuck it on the side.

Also, I can't speak for other people, but I have my in-wall ethernet jack close to my wall power socket.
It's actually nicer, in terms of cable layout, to just run a short cable across the wall, and not have that extra ethernet cable also running to the desk.

Doug S · Oct 29, 2024

The Hardcard said:
Gurman explicitly claimed that the next Ultra would get its own die unrelated to the the Max. That would make sense since Apple plans to build huge numbers of Ultras for Private Cloud Compute datacenters. There are many IP blocks that aren’t needed for cloud machines, like media encoders. They are rumored to sell consumer Ultras, but they could stick a secondary chip to support consumer functions rather than have hundreds of thousands of datacenter Ultras with large chunks of unused, expensive transistors in each.

Also, I’ll reiterate the rumors that Apple is working on an Ultra with several technologies that won’t be available until late next year. I still think there is probability that there won’t be an M4 Ultra. If these new technologies are important to their datacenter Ultra plans, the M5 cores will be ready before the next Ultra can launch.

But if Apple builds enough of them to make it worth doing a separate die for the very low volume Ultra, they are building enough of them to do a separate die for the private cloud datacenter. Then they cut out the stuff they don't need - but add in much more of what they do need. Maybe stuff that the Ultra doesn't support like DDR5 controllers or something.

It just doesn't make sense to design the Ultra as a separate die because they want to shoehorn it into private cloud servers, rather than design exactly what they do need for that role. If they do, somewhere GW III will be saying "I told you so" lol

Meteor Late · Oct 29, 2024

The Hardcard said:
I think AMD was too conservative with planning Strix Halo. You’d have to just hate Macs and/or Apple to pick it over the M4 Max because it ain’t going to be that cheap.

And it's a whole different product, in the sense that Strix Halo won't be a low power consumption chip in low load situations due to the use of chiplets, idle power will be bad.

name99 · Oct 29, 2024

gdansk said:
16GB, full 10 core M4 for $600. No one can beat that price to performance.

Throw in just how many of the you can mount in a rack.
And how MLX can distribute training across Macs. (See Alex Cheema and exolabs.)

The only real question is how long it takes till everyone smaller than FAANG-sized realizes they're foolish to be paying nV prices...

pj- · Oct 29, 2024

Anyone here using apple silicon for machine learning? I tried it a while ago with an m1 mbp and later an m3 mbp but immediately encountered bugs when trying to use the GPU for training (loss divergence) with very simple example models. I wasn't able to troubleshoot much other than finding a couple of apple developer forum posts with similar issues.

I tried googling around just now on the subject and discussions overall seem rare, which is a problem in itself, but it looks like the typical sort of apple problems still exist. Half baked tools/libraries, black boxes, bugs.

M4 mac mini hardware looks great on paper for ML learning/experimenting, but doesn't seem like the software is there to support it

johnsonwax · Oct 29, 2024

igor_kavinski said:
'Tis the year of disappointment!

Where's the jubilation, folks? Apple seems to be providing more performance per dollar than AMD/Intel/Qualcomm, if you are married to Apple store.

This is the best value for money by a long shot if you are doing non-AI data science using open source tools. Like, ridiculously so. No store access needed.

Pro tip - you can extend brew to also manage your App Store apps if you do want to use the store.

johnsonwax · Oct 29, 2024

name99 said:
Also, I can't speak for other people, but I have my in-wall ethernet jack close to my wall power socket.
It's actually nicer, in terms of cable layout, to just run a short cable across the wall, and not have that extra ethernet cable also running to the desk.

I'm mixed on the idea. I do like it for cable management - much more handy, but it then introduces a proprietary component to the machine for the shared power/data.

It seems quite obvious that ethernets days are numbered in the consumer space.

poke01 · Oct 29, 2024

pj- said:
Anyone here using apple silicon for machine learning? I tried it a while ago with an m1 mbp and later an m3 mbp but immediately encountered bugs when trying to use the GPU for training (loss divergence) with very simple example models. I wasn't able to troubleshoot much other than finding a couple of apple developer forum posts with similar issues.

I tried googling around just now on the subject and discussions overall seem rare, which is a problem in itself, but it looks like the typical sort of apple problems still exist. Half baked tools/libraries, black boxes, bugs.

M4 mac mini hardware looks great on paper for ML learning/experimenting, but doesn't seem like the software is there to support it

have you used MLX?

MLX — MLX 0.19.1 documentation

I found these regarding loss

More losses! · Issue #76 · ml-explore/mlx

Hi, I thought it could be nice if there were a few more losses ready to go in mlx. To start with I was thinking: L1_loss (L1 loss) mse_loss (mean squared error) Maybe also: nll_loss (negative log l...

github.com

Loss Functions — MLX 0.19.1 documentation

johnsonwax · Oct 29, 2024

pj- said:
Anyone here using apple silicon for machine learning? I tried it a while ago with an m1 mbp and later an m3 mbp but immediately encountered bugs when trying to use the GPU for training (loss divergence) with very simple example models. I wasn't able to troubleshoot much other than finding a couple of apple developer forum posts with similar issues.

I tried googling around just now on the subject and discussions overall seem rare, which is a problem in itself, but it looks like the typical sort of apple problems still exist. Half baked tools/libraries, black boxes, bugs.

M4 mac mini hardware looks great on paper for ML learning/experimenting, but doesn't seem like the software is there to support it

I don't personally, but there is a whole community of people doing it. This isn't a terrible starting point: https://www.reddit.com/r/LocalLLaMA/

There are a number of developers maintaining Apple Silicon support now. This may be another good starting point.

Eug · Oct 29, 2024

johnsonwax said:
I'm mixed on the idea. I do like it for cable management - much more handy, but it then introduces a proprietary component to the machine for the shared power/data.

It seems quite obvious that ethernets days are numbered in the consumer space.

One problem with the Ethernet-on-the-AC-adapter approach for the iMac is that this unit is quite difficult to get after the fact. It's not sold at the Apple Store, and even if you can find it on the used market, it's very expensive.

Mopetar · Oct 29, 2024

Doug S said:
Can't be getting much yield increase by binning off two efficiency cores

The binning is almost certainly based on parametric yields. Most hardware is fully functional, but some cores can't hit the desired clock speeds at target voltages.

Get enough chips through testing and you can pick some number to satisfy 60% (or whatever percentage you want based on past sales data) of chips qualifying for a particular bin. That determines the clock speeds for the shipping product.

Even with one defective core, a chip may still manage to make the lower bin. Anything worse than that just gets thrown away if it can't be used or gets saved for some other product that requires even less computational power and has a low enough volume to utilize the worst ~2% of silicon they get back.

mvprod123 said:
I don't believe in the 20+8 configuration for the M4 Max. Despite the power efficiency of Apple's silicon, such a chip will have too high a TDP for a laptop chassis

16 cores seems like a possibility. The last 100 MHz is always the most expensive so Apple may dial down the max all-core frequency by a little bit to get more cores into a design.

It could also be a physical 20 cores with some disabled to just 16 for the MBP. That lets them have a lower bin for any chips with defective hardware and to select the 16 most efficient cores to stay within TDP constraints. I'm sure they could also increase the TDP with a better cooling system in their existing chassis.

My MBP (M1 Max) only really gets warm when the GPU is running full load in a game, so I'd be more concerned about that than the CPU cores. Even when I've got those fully loaded, the fan rarely even kicks on unless some other part of the SoC is also being used heavily as well.

pj- · Oct 29, 2024

poke01 said:
have you used MLX?

MLX — MLX 0.19.1 documentation

I found these regarding loss

More losses! · Issue #76 · ml-explore/mlx

Hi, I thought it could be nice if there were a few more losses ready to go in mlx. To start with I was thinking: L1_loss (L1 loss) mse_loss (mean squared error) Maybe also: nll_loss (negative log l...

github.com

Loss Functions — MLX 0.19.1 documentation

I saw MLX when I was looking around today, but the sub-1.0 version and general lack of discussion around it has me leaning a bit toward putting it in the "half-baked" category, or at least "in the oven" to be more charitable. Does seem to be under quite active development, though, which is cool. Will play around with it a bit

johnsonwax said:
I don't personally, but there is a whole community of people doing it. This isn't a terrible starting point: https://www.reddit.com/r/LocalLLaMA/

There are a number of developers maintaining Apple Silicon support now. This may be another good starting point.

I've looked at running LLMs locally before but personally I am looking at more of the model creation side than inference on a pretrained model. Although it seems like it's becoming practical to run an entirely local home automation assistant like Alexa, using Home Assistant, cheap hardware, and an LLM

Meteor Late · Oct 29, 2024

mvprod123 said:
I don't believe in the 20+8 configuration for the M4 Max. Despite the power efficiency of Apple's silicon, such a chip will have too high a TDP for a laptop chassis, especially for a 14-inch model. And the die size will be too large. I'm already curious to see the die shots of the M4 Pro/Max dies. Apple is silent about the number of transistors. Apparently, this information with die shots will be announced tomorrow.

It's not really a TDP issue if Apple really wanted to do that, you just use a power limit and lower frequencies, just like AMD and Intel do with their Laptop parts, or does someone believe a 5800U or a 7840U can run at 4GHz in Cinebench MT if capped at 15W?

Eug · Oct 29, 2024

Luke Miani summarizes what he thinks about the internal arrangement of the Mac mini:

Apple only tells us the air gets sucked in and gets blown around and then blown out.

Based on that Apple animation, he believes the top board is the power supply and the bottom board is the motherboard. A heatsink is not obvious from this side view, but the curved structure on the bottom curving up at the right appears to be a heat pipe, overlying the vents.

In the animation, the air is drawn in as cool blue, but becomes warm orange after passing by the power supply and motherboard and eventually the heat pipe, and then gets blown out.

Discussion Apple Silicon SoC thread

Lifer

Member

Lifer

Platinum Member

Platinum Member

Lifer

Junior Member

Senior member

Senior member

Lifer

Senior member

Platinum Member

Senior member

Platinum Member

Member

Senior member

Senior member

Member

Member

Platinum Member

Member

Lifer

Diamond Member

Senior member

Member

Lifer