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

Nothingness · Feb 19, 2024

FlameTail said:
View attachment 94139
Look at the huge rise in power consumption with A17P.

N3B yields might be a reason, but it's not the only one.

The strategy of relying on clock speed increases to increase performance is not a sustainable one.

For SPECint, A16 is 1.8 SPECint/W, A17 is 1.62.
For SPECfp, A16 is 2.47 SPECint/W, A17 is 2.44.

The decrease in efficiency is not what I'd qualify as "huge".

Doug S · Feb 19, 2024

mikegg said:
The vast majority of members here were sure that Apple Silicon was on an 18 month cadence.

I know, and I have always disagreed. The reason everyone here believes the 18 month thing is because of a Mark Gurman tweet a few years ago, but I'm guessing he got hold of an internal roadmap showing the dates for M2 and M3 and figured out it came out to around 18 months and assumed it would be like that forever.

We'll know with M4. If it comes out this fall then I was right all along and Apple's cadence was not a deliberate choice but just how it worked out as far as resource allocation for getting things off the ground combined with TSMC messing up their process cadence with N3.

If we don't see a sign of M4 in 2024 I'll admit I was wrong and while I'll continue to question the sanity of an 18 month cadence I will admit that Apple has deliberately made that insane choice.

Doug S · Feb 19, 2024

Nothingness said:
For SPECint, A16 is 1.8 SPECint/W, A17 is 1.62.
For SPECfp, A16 is 2.47 SPECint/W, A17 is 2.44.

The decrease in efficiency is not what I'd qualify as "huge".

He wasn't making a claim about efficiency he was making a claim about power consumption, a claim which is not wrong.

mikegg · Feb 19, 2024

SteinFG said:
M1 macbook - Nov, 2020
M2 macbook - Jul, 2022 (20 months)
M3 macbook - Nov, 2023 (16 months)
It's not a lot of samlple points, but the average is 18 😛
If the M4 macbook comes out before March 2025, we can revise that theory, but so far it's right.

M3 Pro/Max came out 10 months after M2 Pro/Max.

okoroezenwa · Feb 19, 2024

SteinFG said:
M1 macbook - Nov, 2020
M2 macbook - Jul, 2022 (20 months)
M3 macbook - Nov, 2023 (16 months)
It's not a lot of samlple points, but the average is 18 😛
If the M4 macbook comes out before March 2025, we can revise that theory, but so far it's right.

IMO, it's only "right" so far if you take the M chips alone. But it doesn't take into account the other chips.

M1 Pro/Max - Oct 2021
M2 Pro/Max - Jan 2023 (15 months; supposedly was supposed to be an Oct 2022 thing, but whatever)
M3 Pro/Max - Oct 2023 (9 months)
This averages out to 1 year.

M1 Ultra - Mar 2022
M2 Ultra - June 2023
15 months here, so less than the supposed 18 month cadence. We'll see if M3 Ultra makes it in June as rumoured (which would make it 12 months there and the average much closer to 1 year than 18 months).

Also if you look at the time it takes for each gen to complete it seems to trend toward one year:

M1: Nov 2020 - Mar 2022
M2: Jun 2022 - June 2023
M3: Oct 2023 - ? (rumoured Jun 2024)

I just can't imagine Apple deciding to twiddle their thumbs in the meantime instead of releasing something (especially when redesigns aren't expected).

I think @Doug S is correct about them intending to have a 1-year cadence here but COVID messed things up. It also lines up with their upgrade cycles for Macs (at least laptops and iMacs) in the Intel times.

FlameTail · Feb 19, 2024

Ok, when did this appear in the thread...

FlameTail · Feb 19, 2024

Doug S said:
He wasn't making a claim about efficiency he was making a claim about power consumption, a claim which is not wrong.

Yeah, technically that's right. But I was alluding to the power efficiency too.

In an era where Moore's Law is dead (pun unintended), each node brings smaller and smaller gains. In this landscape, to increase performance by bumping the frequency is an uphill battle, and not a sustainable one.

In contrast to Intel/AMD, Apple's architecture was designed to run wide and slow. As Andrei Frumusanu noted in his review of the M1, it is one of the factors that contributed to their massive power efficiency lead.

SteinFG · Feb 19, 2024

FlameTail said:
View attachment 94139
Look at the huge rise in power consumption with A17P.

N3B yields might be a reason, but it's not the only one.

The strategy of relying on clock speed increases to increase performance is not a sustainable one.

5GHz A20 soon? (throttles after 1sec)

FlameTail · Feb 19, 2024

It will be very interesting to see what the former-Apple engineers now working at Qualcomm will do with Oryon.

Will they beat the Apple Silicon team? Or will they run into the same roadblocks the Apple Silicon team has?

I think their is a very real competition between these two. Apple hasn't publicly acknowledged it (of course), but the other has.

Nothingness · Feb 19, 2024

Doug S said:
He wasn't making a claim about efficiency he was making a claim about power consumption, a claim which is not wrong.

I know what he claimed. I was only pointing out efficiency did not go down the toilet.

FlameTail · Feb 19, 2024

FlameTail · Feb 19, 2024

The PPC gains dropped off a cliff after the A13 Bionic in 2019- coincidentally the same year the Exodus happened and Nuvia was founded.

FlameTail · Feb 19, 2024

M1 vs M3
3 years later, and not even 10% PPC gain....

FlameTail · Feb 19, 2024

Meanwhile, the PPC gains from A10 -> A13 look spectacular.

Those must have been the good times👀

soresu · Feb 19, 2024

FlameTail said:
The PPC gains dropped off a cliff after the A13 Bionic in 2019- coincidentally the same year the Exodus happened and Nuvia was founded.

µArch designs are years in the making as we see from the time gap between Nuvia being founded, acquired and the actual first silicon of Oryon.

The exodus was more likely simply a matter of engineers being sick of being in the same place as we have seen before multiple times with Jim Keller.

Complacency breeds stagnation.

Nothingness · Feb 19, 2024

FlameTail said:
M1 vs M3
3 years later, and not even 10% PPC gain....

As a rule of thumb, for the same uarch and memory, the higher the frequency of the core is, the lower performance per clock will be (except for tests that don't depend on memory).
Also as you increase frequency, you might have to add extra cycle to access caches.

FlameTail said:
The PPC gains dropped off a cliff after the A13 Bionic in 2019- coincidentally the same year the Exodus happened and Nuvia was founded.

As you advance, you get less and less low hanging fruits so progress is slower.

SteinFG · Feb 19, 2024

FlameTail said:
The PPC gains dropped off a cliff after the A13 Bionic in 2019- coincidentally the same year the Exodus happened and Nuvia was founded.

Maybe it's deliberate. All smartphone tech has plateaued for the last 3 or so years, and Apple sees that most people are happy with the power efficiency that A15 offers, plus people want an 800 dollar phone. With increasing transistor costs and higher inflation, Apple knows that the iphone with more features will inevitably cost more. So why even bother then, if everyone will keep buying an 800 dollar option. In short: apple found peak 800 dollar phone and knows that's what most want.

FlameTail · Feb 19, 2024

SteinFG said:
Maybe it's deliberate. All smartphone tech has plateaued for the last 3 or so years, and Apple sees that most people are happy with the power efficiency that A15 offers, plus people want an 800 dollar phone. With increasing transistor costs and higher inflation, Apple knows that the iphone with more features will inevitably cost more. So why even bother then, if everyone will keep buying an 800 dollar option. In short: apple found peak 800 dollar phone and knows that's what most want.

Ah, but this isn't only about iPhones!

Apple uses these same core designs in their M series chips for Macs.

FlameTail · Feb 19, 2024

Nothingness said:
As a rule of thumb, for the same uarch and memory, the higher the frequency of the core is, the lower performance per clock will be (except for tests that don't depend on memory).
Also as you increase frequency, you might have to add extra cycle to access caches.

Exactly. Another reason why performance increase via clock speed increase is unsustainable.

Doug S · Feb 19, 2024

FlameTail said:
The PPC gains dropped off a cliff after the A13 Bionic in 2019- coincidentally the same year the Exodus happened and Nuvia was founded.

When Moses and the gang (to continue your capitalized "Exodus" theme) left in 2019 the A14 design would have months from tapeout and all the major planning for A15 would have been done. A16 would probably have on the drawing board though not fully formed. The first cores not touched in any way by Williams et al are found in the current A17P/M3. You could probably make the argument they didn't have all that much to do with A16, but the lackluster gains from A14 and A15 can't be blamed on the departures.

I think they didn't run out of ideas, they ran out of time, so to speak. People keep forgetting it was a massive effort to bring Apple's iPhone SoC into the Mac. Obviously it had been worked on for years and there were FrankenMacs running in secret labs so software guys could do their thing, but actually taking it to released products shipping to customers is a whole different story. Not only that, they were going from one "big brother SoC" every other year or so (e.g. A10X, A12X) to three big brother SoCs every 12 to 18 months (and the Ultra on top of it) which required a lot of engineering and management effort from the people at the top.

If you're allocating that effort and you don't have enough resources you can't do everything you want, you have to make some choices and the extra effort MUST go toward making Apple Silicon successful. There will be no second chances, so if that means smaller gains in your "big core" during that time because you lack engineering bandwidth so be it. Even if Apple planned ahead and beefed up their team via hiring it will take the new additions significant time to be able to function on the level of team members who have been around a few years and already worked at least one SoC the entire life cycle from clean sheet to mass production.

If my "12 months was always the plan" theory is true, the simultaneous release of all three M3 chips immediately following A17P could be a sign that they've finally "caught up" as it were to a steady state. Where they have the proper resources to handle the steady state demands for the Apple Silicon designs and perhaps will finally have enough time to devote to more "major" architectural moves. But if so that would mean not seeing the fruits of that until A19 or A20. Plus we have to keep in mind that making 20-25% yearly gains as they were in the A10->A13 years (and even larger before that) becomes more and more difficult the higher your starting point. And process doesn't help as much as it used to either. Like I always say, diminishing returns is a thing.

moinmoin · Feb 19, 2024

Doug S said:
When Moses and the gang (to continue your capitalized "Exodus" theme) left in 2019 the A14 design would have months from tapeout and all the major planning for A15 would have been done. A16 would probably have on the drawing board though not fully formed. The first cores not touched in any way by Williams et al are found in the current A17P/M3. You could probably make the argument they didn't have all that much to do with A16, but the lackluster gains from A14 and A15 can't be blamed on the departures.

The reports regarding Gerard Williams III back in 2019 were oddly specific in that he was said to be only responsible for A7 to A12X.

LightningZ71 · Feb 19, 2024

Getting back to NAND based AI tech, does this not seem like a best case scenario for 3d X-point/Optane memory? Nand density, near RAM endurance, and already developed tech. A couple of 32GB/64 GB chips would give plenty of storage for a REASONABLY sized model. It's also easily updated if needed.

DrMrLordX · Feb 20, 2024

LightningZ71 said:
Getting back to NAND based AI tech, does this not seem like a best case scenario for 3d X-point/Optane memory? Nand density, near RAM endurance, and already developed tech. A couple of 32GB/64 GB chips would give plenty of storage for a REASONABLY sized model. It's also easily updated if needed.

How much power does it require vs. NAND for long term storage?

Gerard Williams · Feb 21, 2024

moinmoin said:
The reports regarding Gerard Williams III back in 2019 were oddly specific in that he was said to be only responsible for A7 to A12X.

Was responsible for A7 to A14. And Apple M1.

FlameTail · Feb 21, 2024

It's him!🔥

Discussion Apple Silicon SoC thread

Lifer

Platinum Member

Platinum Member

Platinum Member

Golden Member

Member

Platinum Member

Platinum Member

Senior member

Platinum Member

Platinum Member

Platinum Member

Platinum Member

Platinum Member

Platinum Member

Platinum Member

Platinum Member

Senior member

Platinum Member

Platinum Member

Platinum Member

Diamond Member

Golden Member

Lifer

Junior Member

Platinum Member