M4 series core count speculation

[prime17569]

Now that the initial M3 series chips are out, we can begin to speculate about the core counts for the M4 series chips

One of the headline changes for the M3 series chip architecture is the switch to six-core clusters for the Pro and Max P-cores. The Pro has one of these clusters and the Max has two. Other multiples of three also pop up: the M3 Pro and lower-tier M3 Max use triple channel memory, and the M3 Pro has six E-cores.

Given this, I think that the M4 series will have the following core counts. Notably, I think the M4 Max will use the six-core E-core cluster layout that was debuted in the M3 Pro as a space-efficient way to boost multicore performance and potentially efficiency.

Base M4: 4 P-cores, 4 E-cores, 10 GPU cores
M4 Pro: 6 P-cores, 6 E-cores, 20 GPU cores (up from 18 in M3 Pro and 19 in M2 Pro)
M4 Max: 12 P-cores, 6 E-cores (up from 4 in M3 Max), 40 GPU cores

Additionally, I think that the base M4 will begin to use 6GB RAM chips for a base spec of 12 GB RAM, especially considering that the iPhone 15 Pro already has 8 GB of RAM.

Do these sound realistic?

 

[altaic]

If you look at the die shots of the "6-core P-clusters" in the M3 Pro/Max, you can see that they're actually designed as 8-core clusters, but with other stuff taking up a 2-core sub-cluster spot. And, in the case of the M3, other stuff takes up two such sub-clusters, resulting in 4 P-cores.

So, the M4 Pro/Max could just as easily have multiples of 6 cores as the M3 Pro/Max, multiples of 8 cores, multiples of 4 cores, or even a combination. Heck, even the M3 Ultra (or M3 Extreme or whatever) could have multiples of 8 cores, 4 cores, etc. Hope that helps

 

[prime17569]

If you look at the die shots of the "6-core P-clusters" in the M3 Pro/Max, you can see that they're actually designed as 8-core clusters, but with other stuff taking up a 2-core sub-cluster spot. And, in the case of the M3, other stuff takes up two such sub-clusters, resulting in 4 P-cores.

So, the M4 Pro/Max could just as easily have multiples of 6 cores as the M3 Pro/Max, multiples of 8 cores, multiples of 4 cores, or even a combination. Heck, even the M3 Ultra (or M3 Extreme or whatever) could have multiples of 8 cores, 4 cores, etc. Hope that helps

According to the die shot analyses of the chips, the other stuff is the AMX blocks. On the M3, that additional two-core space is used by the Neural Engine.

I'm not sure that they would go to the trouble of designing an all-new six-core cluster layout to only use it for a year. The M1 and M2 Pro/Max have two clusters of four cores in an identical layout. The only difference is that the M2 Pro/Max added two additional E-cores. That is why I am speculating that they will keep the six-core P-core clusters and the six-core E-core cluster will make its way to the M4 Max.

 

[Boil]

Since the thread limit for macOS is 64, and hoping for a quad chip configuration (Mn Extreme) at some point, I would think the Mn Max would stick to the current 12P/4E (16 total) CPU core counts; dual chip gives one 32 cores & quad chip gives one 64 cores...

The area that needs "more of" is the GPU cores, but adding more GPU cores will make for a larger SoC; maybe Apple can extend the core count to 48 or 56, or maybe they go all out and make the largest SoC they can and jam 64 GPU cores onto a Mn Max die...?!?

Another method to get more GPU cores into the mix would be an asymmetrical configuration, pairing a "standard" CPU/GPU/etc. SoC with a "GPU specific" SoC...?

Mn Max

• 16-core CPU (12P/4E)
• 48-core GPU
• 16-core Neural Engine

Mn Compute

• 96-core GPU
• 16-core Neural Engine

Mn Ultra Compute SoC (asymmetrical configuration)

• 16-core CPU (12P/4E)
• 144-core GPU
• 32-core Neural Engine

Mn Extreme Compute SoC (asymmetrical configuration)

• 32-core CPU (24P/8E)
• 288-core GPU
• 64-core Neural Engine

Get LPDDR5X RAM into the mix, with 64GB chips and using an in-line ECC scheme, one could see up to 960GB of LPDDR5X RAM in a quad chip configuration (standard or asymmetrical)...!

 

[MF878]

I wonder if the M3 Pro moving from 8+4 to 6+6 was a response to the high cost/low yield of the N3B process.

If N3E resolves these issues on the M4, then it’s possible the M4 Pro could revert back to 8+4 with clusters of 4 P-cores across the board (1 on M4, 2 on M4 Pro, 3 on M4 Max).

I don’t see Apple upping E-cores on the Max, because then you’d end up with 12 E-cores on the Ultra, which aren’t that useful on a desktop-only chip.

If there is any room to increase core counts on M4, then I suspect that will go to GPU cores as Apple is absolutely crushing it on CPU already. M4 Pro moving up to 20 makes sense (the drop from 19 to 18 again feels like a cost/yield move), I’m not sure if N3E will be enough of an efficiency/yield improvement to jump to 6-core GPU clusters? If so we could end up with 12 on the M4, 24 on the Pro and 48 on the Max.

 

[Beau10]

If there is any room to increase core counts on M4, then I suspect that will go to GPU cores as Apple is absolutely crushing it on CPU already. M4 Pro moving up to 20 makes sense (the drop from 19 to 18 again feels like a cost/yield move), I’m not sure if N3E will be enough of an efficiency/yield improvement to jump to 6-core GPU clusters? If so we could end up with 12 on the M4, 24 on the Pro and 48 on the Max.

I think Apple is always going to try to improve compute but it may just make the cores more efficient/faster in the meantime... but if they can't get at least 15% faster here between generations, we may see them add another 2 cores.

Agreed that they likely will swap 2 eff cores back as that move seems to only be here to reduce transistor counts in the face of a shaky yield process.

Agreed that GPU is where they need to focus with the push from the ML/gaming front, likely trying to pace things up to maybe 40% improvement generation over generation to catch up with the best of NVidia in a few generations.

 

[picpicmac]

Since I view Apple as primarily a maker of mobile computing devices, I suspect Apple's next suite of processors will once again move towards more efficiency.

The desktop Macintosh products are but a tiny fraction of the overall Apple corporate revenue, so I am not expecting a push towards replacing Nvidia at the high end.

As such, shrinking the physical size and power consumption are what matter.

Notice how the M3 Pro has fewer transistors than the M2 Pro.

I wonder if that will continue with the next generation.

 

[Beau10]

Notice how the M3 Pro has fewer transistors than the M2 Pro.

That's only because of the yield issue which has been well documented. It will reverse with the next gen.

 

[prime17569]

I’m not sure if N3E will be enough of an efficiency/yield improvement to jump to 6-core GPU clusters? If so we could end up with 12 on the M4, 24 on the Pro and 48 on the Max.

6-core GPU clusters could be a possibility. The A17 Pro has six GPU cores, meaning that the M4 might have 12.

This also makes sense considering past precedent:
A14: 4 GPU cores -> M1: 8 GPU cores
A15: 5 GPU cores -> M2: 10 GPU cores
A16: 5 GPU cores -> M3*: 10 GPU cores
A17: 6 GPU cores -> M4: 12 GPU cores (presumably)

*Continuing the trend that started with the first Mac Studio, all M3 Macs have a 15 in their model identifier (e.g. Mac15,x), which matches the iPhone15,x notation used for the iPhone 14 Pro and regular iPhone 15 models that have the A16.

This, and the fact that the M3 series chips didn't get the updated 35 TOPS neural engine from A17 Pro, suggests that the M3 series chips are based on the core designs that were originally intended to go into the A16 before TSMC's N3B node was delayed and A16 had to be backported to 4nm.

 

[Basic75]

A16: 5 GPU cores -> M3*: 10 GPU cores
A17: 6 GPU cores -> M4: 12 GPU cores (presumably)

But the M3 GPU is a larger version of the A17 GPU.

 

[theorist9]

Since the thread limit for macOS is 64, and hoping for a quad chip configuration (Mn Extreme) at some point, I would think the Mn Max would stick to the current 12P/4E (16 total) CPU core counts; dual chip gives one 32 cores & quad chip gives one 64 cores...

What's the source of the 64-thread limit, and how fundamental is it to the OS—would allowing more than 64 simultaneous threads require a minor rewrite, a significant rewrite, or a complete overhaul to the OS?

Knowing that would tell us whether the current thread limit constitutes a significant barrier to Apple offering an Extreme with >64 cores.

[To forestall potential confusion from others: You may have read that MacOS can support thousands of threads, and that's true; but that's not thousands of simultaenously-operating threads, which is what Boil and I are discussing; the latter is limited to 64.]

Edit: According to https://dortania.github.io/OpenCore-Install-Guide/macos-limits.html#cpu-support , it seems one can go beyond the 64-thread limit....

 

[galad]

It's not really an issue for Apple. If they want to ship a CPU with more than 64bit, they will just rework that part of the kernel and increase the limit.
It's an issue on Hackintosh because you can't ship a custom kernel.

 

[senttoschool]

That's only because of the yield issue which has been well documented. It will reverse with the next gen.

I would hope so. But I don't think so. I think Apple wants to separate the Pro line from the Max line more. They want a bigger gap so people have to pay more to jump to the next level of performance.

 

[theorist9]

It's not really an issue for Apple. If they want to ship a CPU with more than 64bit, they will just rework that part of the kernel and increase the limit.
It's an issue on Hackintosh because you can't ship a custom kernel.

We're discussing the max number of threads (and thus the max number of supported cores) for MacOS, not the number of bits. The number of bits is entirely different, and determines the size of addressable memory.

A bit of history (no pun intended!): MacOS began its transition to a 64-bit kernel with Panther, and completed it with Snow Leopard:

Mac OS X 10.6 Snow Leopard: the Ars Technica review

No new features.

arstechnica.com

 

[Basic75]

What's the source of the 64-thread limit, and how fundamental is it to the OS—would allowing more than 64 simultaneous threads require a minor rewrite, a significant rewrite, or a complete overhaul to the OS?

I can't be bothered to dig through the xnu source, but perhaps somebody else wants to step up? They probably have some CPU bit mask in a uint64 and it's annoying to extend to more, especially when that creates overhead for devices like phones and watches that are still very far from the 64 core limit.

 

[falainber]

It's probably unrealistic to expect significant core count increases be it CPU or GPU simply because TSMC process density is not going to increase in a meaningful way any time soon. Apple was able to do it for M3 because they switched from N5 to N3 process. They are going to stay on N3 for a while.

 

[theorist9]

It's probably unrealistic to expect significant core count increases be it CPU or GPU simply because TSMC process density is not going to increase in a meaningful way any time soon. Apple was able to do it for M3 because they switched from N5 to N3 process. They are going to stay on N3 for a while.

They could make the die somewhat larger if they wanted. The M3 Max is well below the reticle limit. It's just a matter of economics.

 

[Beau10]

I would hope so. But I don't think so. I think Apple wants to separate the Pro line from the Max line more. They want a bigger gap so people have to pay more to jump to the next level of performance.

Yeah, I've changed my stance on that since, ultimately bought an M3 Max thinking the next pro will probably be fairly underpowered still.

 

[senttoschool]

Yeah, I've changed my stance on that since, ultimately bought an M3 Max thinking the next pro will probably be fairly underpowered still.

The biggest telling is that the base M3 increased transistors from M2 by 25% (20b --> 25b). But the M3 Pro decreased from M2 Pro (40b --> 37b). That tells you that it's not a yield issue. It's business segmentation. If there is a yield issue, Apple wouldn't have increased the base M by a whopping 25%.

 

[Superman041]

I've see the below posted on Reddit and wanted to know your thoughts on it?

Edit - I have seen you discussing what I think it something similar above. (Apologies I am not technical and trying to understand it all to make a purchasing decision on whether to wait for the M4. Would be grateful if someone could explain it to me like a dummy)

The M3 is built using TSCM N3B process which has a very low production yield. Only 55% of the chips are good, 45% of production is a failure. This was the only option because the N3E process will not go into production before the end of the year.

Will Apple move to N3E as soon as it becomes available? If so, they could do that during 2024. It would be logical to name the silicon on N3E as “M4” because it should use slightly less power and/or have better performance.

 

[Zorori]

Agreed that GPU is where they need to focus with the push from the ML/gaming front, likely trying to pace things up to maybe 40% improvement generation over generation to catch up with the best of NVidia in a few generations.

And sadly for many, this is a complete waste

If Apple keep focusing on GPU improvements they’ll see users gained during the M1 jump leave as Intel/AMD catch up on what’s important to their workflows

 

[quarkysg]

If Apple keep focusing on GPU improvements they’ll see users gained during the M1 jump leave as Intel/AMD catch up on what’s important to their workflows

All Apple need to do is increase CPU clock speed and Intel will not be able to keep up, if power draw be damned.

 

[JinxVi]

If Apple keep focusing on GPU improvements they’ll see users gained during the M1 jump leave as Intel/AMD catch up on what’s important to their workflows.

I don’t even know, if Apple gained any users with the transition? There’s a constant growth of the platform, because of the iPhone halo effect. And then there’s a wave of Intel Mac users, who preponed their upgrade cycle because of Apple Silicon.

The excellent battery life and power efficiency of Apple Silicon might have created new interest in the platform. But we can’t know how large the effect was until many years later. Advancing the GPU capabilities of the entire ecosystem from iPhone to iPad can only help the Mac.

 

[aj_niner]

A question...

Do you guys think that the base M4 3nm chip will outperform a

- 2019 Mac Pro Xeon 14nm
- 2020 iMac 27" Core i9 14nm

 

[hovscorpion12]

A question...

Do you guys think that the base M4 3nm chip will outperform a

- 2019 Mac Pro Xeon 14nm
- 2020 iMac 27" Core i9 14nm

Not the base Chips. The base M-Series will most likely keep the 8-Core CPU/10-Core GPU design for pricing reasons.

It's more the Pro/Max/Ultra chips that will embarrass the Intel Mac Pro/iMac models.