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]

Geekbench Search - Geekbench

Apple m3 geekbench single core

Apple M3 Max 16-core gets >21000 in Geekbench 6 multi-core.

M3 Max Geekbench 6

That is a touch below M2 Ultra and Core i9-13900K.

 

[poke01]

Apple M3 Max 16-core gets >21000 in Geekbench 6 multi-core.

M3 Max Geekbench 6

That is a touch below M2 Ultra and Core i9-13900K.

Also is closely equal to 14900k

 

[Doug S]

Curious, I would have thought the Pro was the sweet spot for many. Well, Apple knows what's selling and what's not - no way for me to second guess them.

I also saw an interesting comment by repoman27 in the Anandtech article about the M3 where he mentioned that Apple was double stacking the memory controllers in the Pro & Max dies previously (you can see from die photos) but for M3 only do that in the Max.

Cutting back on the width may have been necessary to avoid that double stacking which would make the packaging steps that combine the SoC with the LPDDR modules easier/cheaper. That also explains why the Pro is no longer a "chop". That packaging change helps Apple cut the cost of the Pro. Like you I don't know what's selling and what isn't but Apple does and they clearly decided they wanted to make that rather large price move in the Pro so they probably tasked the engineers with finding ways to reduce the cost to help make that a reality.

I remember when the first rumors came out about the 36 GB configuration being tested on an M3 Macbook Pro. I took that to mean that Apple was bumping the memory width by 50% as there was no way they could keep a 256 bit width and get a 36 GB config. I pointed that out immediately when it was mentioned here and I'm likely the reason people started expecting wider memory configs with M3. I didn't consider the possibility of reducing memory width, so I'll have to take my mea culpas on that. I suppose if I had considered that though I would have suggested "well if they go to LPDDR5X-8500 then they can do that while maintaining the same memory bandwidth". I guess we'll have to wait for M4 for that.

What I am curious to see is whether that reduction in memory bandwidth actually matters in practice. i.e. can anyone find a single benchmark that runs faster on M2 Pro than it does on M3 Pro?

 

[repoman27]

I remember when the first rumors came out about the 36 GB configuration being tested on an M3 Macbook Pro. I took that to mean that Apple was bumping the memory width by 50% as there was no way they could keep a 256 bit width and get a 36 GB config. I pointed that out immediately when it was mentioned here and I'm likely the reason people started expecting wider memory configs with M3. I didn't consider the possibility of reducing memory width, so I'll have to take my mea culpas on that. I suppose if I had considered that though I would have suggested "well if they go to LPDDR5X-8500 then they can do that while maintaining the same memory bandwidth". I guess we'll have to wait for M4 for that.

Yeah, I read that back when you posted about it and quickly realized it was most likely a 192-bit bus, but totally expected LPDDRX-8533 to bring it right back to 204.8 GB/s bandwidth.

A lot of folks have referred to that 36GB config as being an oddball because it is divisible by 3 (even Chips and Cheese did so), but what's more notable is that it's divisible by 9. In other words, 3 is a factor twice. 12 and 24 Gbit LPDDR5 dies have been around for a while, and Apple was already using them. But to get another 3 in there, you either have to do 6-high stacks in the packages, or go to a multiple of three for your memory channels. The latter was more likely, and a 192-bit interface with 3 x64 packages was really the only option that was in the cards.

 

[GC2:CS]

I also saw an interesting comment by repoman27 in the Anandtech article about the M3 where he mentioned that Apple was double stacking the memory controllers in the Pro & Max dies previously (you can see from die photos) but for M3 only do that in the Max.

Cutting back on the width may have been necessary to avoid that double stacking which would make the packaging steps that combine the SoC with the LPDDR modules easier/cheaper. That also explains why the Pro is no longer a "chop". That packaging change helps Apple cut the cost of the Pro. Like you I don't know what's selling and what isn't but Apple does and they clearly decided they wanted to make that rather large price move in the Pro so they probably tasked the engineers with finding ways to reduce the cost to help make that a reality.

I remember when the first rumors came out about the 36 GB configuration being tested on an M3 Macbook Pro. I took that to mean that Apple was bumping the memory width by 50% as there was no way they could keep a 256 bit width and get a 36 GB config. I pointed that out immediately when it was mentioned here and I'm likely the reason people started expecting wider memory configs with M3. I didn't consider the possibility of reducing memory width, so I'll have to take my mea culpas on that. I suppose if I had considered that though I would have suggested "well if they go to LPDDR5X-8500 then they can do that while maintaining the same memory bandwidth". I guess we'll have to wait for M4 for that.

What I am curious to see is whether that reduction in memory bandwidth actually matters in practice. i.e. can anyone find a single benchmark that runs faster on M2 Pro than it does on M3 Pro?

What does it even mean to double stack a memory controler ? Like having two layers of conections one on the edge of the chip and another one a bit behind ?

 

[FlameTail]

Yeah, I read that back when you posted about it and quickly realized it was most likely a 192-bit bus, but totally expected LPDDRX-8533 to bring it right back to 204.8 GB/s bandwidth.

A lot of folks have referred to that 36GB config as being an oddball because it is divisible by 3 (even Chips and Cheese did so), but what's more notable is that it's divisible by 9. In other words, 3 is a factor twice. 12 and 24 Gbit LPDDR5 dies have been around for a while, and Apple was already using them. But to get another 3 in there, you either have to do 6-high stacks in the packages, or go to a multiple of three for your memory channels. The latter was more likely, and a 192-bit interface with 3 x64 packages was really the only option that was in the cards.

For what it's worth, 18 GB LPDDR5 dies do exist. You can see Samsung's website where they have listed a 144 Gb density.

A few smartphones have this unusual 18 GB capacity.

 

[MadRat]

So M3 is literally M technology x3?

 

[Glo.]

M3 MacBook Pro: 8 GB RAM, 2 thunderbolt 4 ports, 1 HDMI port, 1 external display support from Thunderbolt, and 1 from HDMI, everything else is the same as...
M2 Pro MacBook Pro, refurbished, 16 GB of RAM, 2 more CPU cores, 6 more GPU cores, 1 more Thunderbolt 4 port, support for much more external displays, and 4K/240 Hz from HDMI port.

Both computers cost the same 1599$.

What is the point of it?

 

[repoman27]

What does it even mean to double stack a memory controler ? Like having two layers of conections one on the edge of the chip and another one a bit behind ?

That's exactly what I meant, but it might be more like they folded the memory controller layout in half so it's twice as deep but only uses half the amount of shoreline. It probably makes fan-out more challenging, but it's the only way they can realistically fit a 512-bit bus on that die, seeing as the bottom edge is totally dedicated to the d2d interconnect.

visual aid:

For what it's worth, 18 GB LPDDR5 dies do exist. You can see Samsung's website where they have listed a 144 Gb density.

A few smartphones have this unusual 18 GB capacity.

AFAIK, LPDDR5 dies are available in 8, 12, 16, 24, and 32 Gbit densities. 18 GB packages are totally a thing, but I believe they are all made using 6x 24 Gbit dies. As I said, I thought it unlikely that Apple would go to 6-high stacks.

M3 MacBook Pro: 8 GB RAM, 2 thunderbolt 4 ports, 1 HDMI port, 1 external display support from Thunderbolt, and 1 from HDMI, everything else is the same as...
M2 Pro MacBook Pro, refurbished, 16 GB of RAM, 2 more CPU cores, 6 more GPU cores, 1 more Thunderbolt 4 port, support for much more external displays, and 4K/240 Hz from HDMI port.

Both computers cost the same 1599$.

What is the point of it?

I bet that inventory of refurb M2 Pro MBPs dries up real quick.

 

[FlameTail]

AFAIK, LPDDR5 dies are available in 8, 12, 16, 24, and 32 Gbit densities. 18 GB packages are totally a thing, but I believe they are all made using 6x 24 Gbit dies. As I said, I thought it unlikely that Apple would go to 6-high stacks.

LPDDR5 | DRAM | Samsung Semiconductor Global

Meet LPDDR5 powering next-generation applications with performance and efficiency by 6,400 Mbps of pin speed, massive transfer at 51.2Gb/s, and 20% power saving.

semiconductor.samsung.com

 

[igor_kavinski]

M3 MacBook Pro: 8 GB RAM, 2 thunderbolt 4 ports, 1 HDMI port, 1 external display support from Thunderbolt, and 1 from HDMI, everything else is the same as...
M2 Pro MacBook Pro, refurbished, 16 GB of RAM, 2 more CPU cores, 6 more GPU cores, 1 more Thunderbolt 4 port, support for much more external displays, and 4K/240 Hz from HDMI port.

Both computers cost the same 1599$.

What is the point of it?

Raytracing

 

[Mopetar]

What an odd comparison. The M3 is designed to go in ultrabooks and iPads. That it also powers regular laptops and iMacs doesn't mean it's relevant to compare it with desktop gaming CPUs.

The Zen core is also used in those same markets, so I don't think it's that much of a stretch. Generally the v-cache CPUs use less power on top of offering better gaming performance so from a certain standpoint, it would make more sense for AMD to use that technology in a premium ultra book product.

Really the only reason not to is that the v-cache parts don't get much extra performance outside of gaming, so there's not much reason to use one when for anything outside of a really high-end gaming laptop, the performance bottleneck will be the GPU.

Raytracing

RT doesn't have much of a presence in gaming as is and of the titles that do use it extensively, you need a 4090 to have any kind of reasonable performance.

Macs already aren't seen as a gaming platform and the paltry amount of RT the M3 brings to the table isn't going to matter for most consumers.

Anyone using professional software that adds support for it is probably holding out for an M3 Max or perhaps even an Ultra.

 

[ashFTW]

Geekbench Search - Geekbench

SC GB6 score of 3227 reported for M3 Max

 

[richardskrad]

I don't understand why people are saying Apple is doomed. The M3 family of chips on the laptop side are literally the most advanced chips on earth. No other chip maker can compete on performance per watt. Qualcomm's new M-series competitors are promising but they'll come out halfway through 2024. At that point, Apple is preparing to release the M4.

 

[gdansk]

A17/M3 isn't their best work. It is a skip generation despite being on a new node. I know that offends consumers on principle but people are comparing M1/M2 to M3 and it isn't compelling unless you spend $5k on a laptop.

 

[FlameTail]

Geekbench Search - Geekbench

SC GB6 score of 3227 reported for M3 Max

Snapdragon X Elite slain already.

 

[igor_kavinski]

Snapdragon X Elite slain already.

But was the M3 Max score achieved at room temp?

 

[amrnuke]

But was the M3 Max score achieved at room temp?

Highest M2 Max SC scores in the GB database are generally the 2800s
The earliest numbers were largely in the 2700s, though
M2 Max official listing is 2803
Similar trend for other CPUs
Also might be irrelevant whether it was at room temp, especially in context of comparing it to unverified marketing slides from Qualcomm on a product that's a year from release

 

[igor_kavinski]

Apple M3 and the State of CPUs

Apple’s M3 launch stood out for its focus on how people actually use computers, the lack of AI acronyms and the fact that they launched 3 chips all at once. It is only getting harder to compe…

digitstodollars.com

Perhaps the most significant facet of the M3 launch was the fact that they are launching three chips at once. Our initial assumption was that these were the same chip, ‘binned’ in production, but that was wrong. These are three different chips, with the differences manifest in their very different sizes of each and visible to the naked eye in the die photos above. This is somewhat staggering, as we have to assume that tape out costs alone for the three has to be close to $1 billion. Very few companies can afford this large an undertaking.

$1 billion for Apple is like, under the carpet/bed/in the kitchen cabinet money.

 

[darkswordsman17]

I don't understand why people are saying Apple is doomed. The M3 family of chips on the laptop side are literally the most advanced chips on earth. No other chip maker can compete on performance per watt. Qualcomm's new M-series competitors are promising but they'll come out halfway through 2024. At that point, Apple is preparing to release the M4.

Its just how people are. Apple will be fine even if they lose performance crown, and most likely we'll end up back where things are with Android or Windows, where it'll depend if there's native Apple versions of the software, etc, where due to Apple's tighter software/hardware integration they'll have the edge in some aspects (like battery life), even if they don't have absolute performance across the board.

While I'm glad for the competition drivings things (hoping we get this Snapdragon chip in premium thin 'n light $1000 and below laptops; also hope it forces Apple to up core counts and other; hope it spurs AMD and Intel to make inroads and/or leverage other advantages to make devices that these others won't like thin and light laptops that have external GPUs), I'm honestly not looking forward to it for 3 reasons. The first being that when there's real competition this place (and nerds in general) become insufferable fanboy a-holes who make mountains over minor issues (what you're talking about is a prime example). The 2nd being that most of this will be overshadowed by them forcing AI into everything. And 3rd is that computing is more than simple performance or any one metric, and computing is going through big changes, so fussing over pretty small differences in benchmark scores is largely pointless.

A17/M3 isn't their best work. It is a skip generation despite being on a new node. I know that offends consumers on principle but people are comparing M1/M2 to M3 and it isn't compelling unless you spend $5k on a laptop.

M2 was a skip gen more than M3 is. It'll be as compelling as M2 was once it is in all of the products those chips were in. But most M1 users likely can skip until M5 gen before they'll really find enough compelling to warrant an upgrade as most are not likely even taxing the M1 performance. Maybe by then we'll start seeing native Apple versions of big name games being made.

Snapdragon X Elite slain already.

Eh, that's a silly way to look at things. There's a multitude of factors that will likely play a MUCH larger role than single benchmark scores. Even if things were reversed, I highly doubt Apple users would be ditching them for this chip.

 

[repoman27]

LPDDR5 | DRAM | Samsung Semiconductor Global

Meet LPDDR5 powering next-generation applications with performance and efficiency by 6,400 Mbps of pin speed, massive transfer at 51.2Gb/s, and 20% power saving.

semiconductor.samsung.com

View attachment 88242

As I said, 18 GB packages are a thing (144 Gb = 18 GB). They are made using six 24 Gbit dies. 6 x 24 = 144.

AFAIK, 18 Gbit dies are not a thing.

 

[jeanlain]

But was the M3 Max score achieved at room temp?

Apple SoCs can't be overclocked so I don't think operating temperature is much relevant to a single-threaded short test.

 

[igor_kavinski]

Apple SoCs can't be overclocked so I don't think operating temperature is much relevant to a single-threaded short test.

What about turbo boost? Do Apple SoCs boost as far as their thermal limits allow?

 

[gdansk]

M2 was a skip gen more than M3 is

Yes but not really M3 Pro is even more of a skip than M2 Pro. At least M2 Pro was a straightforward improvement over its predecessor.

 

[jeanlain]

What about turbo boost? Do Apple SoCs boost as far as their thermal limits allow?

Apple SoCs don't have such feature, AFAIK.