So why exactly did m1 pro/max only have 2 efficiency cores?

[JPack]

But that doesn’t matter what Apple is doing in terms of E core performance, they don’t even talk about it publicly. Our resident chip expert @cmaier explained how far ahead Apple’s E cores are from both a performance AND extreme low power consumption prospective a while back. I’ll see if I can dig that up after work.

Edit: Apologies, I mixed up users, it was actually @lemans (another hugely knowledgeable member in chip design).

Apple "benchmarking themselves" meaning Apple looks at their own needs with macOS.

If M1 E-cores weren't powerful enough based on Apple modeling with typical usage, and the end result would be firing up the main cores anyway, then it wouldn't make sense to add more cores.

With A15, the E-core performance went up significantly, to a point where users could run on Low Power mode consistently. Once that was the case, it was worth investing die area for additional E-cores.

 

[NT1440]

Apple "benchmarking themselves" meaning Apple looks at their own needs with macOS.

If M1 E-cores weren't powerful enough based on Apple modeling with typical usage, and the end result would be firing up the main cores anyway, then it wouldn't make sense to add more cores.

With A15, the E-core performance went up significantly, to a point where users could run on Low Power mode consistently. Once that was the case, it was worth investing die area for additional E-cores.

Ah, thanks for the clarification, that makes sense.

 

[dmccloud]

Apple "benchmarking themselves" meaning Apple looks at their own needs with macOS.

If M1 E-cores weren't powerful enough based on Apple modeling with typical usage, and the end result would be firing up the main cores anyway, then it wouldn't make sense to add more cores.

With A15, the E-core performance went up significantly, to a point where users could run on Low Power mode consistently. Once that was the case, it was worth investing die area for additional E-cores.

This makes sense, largely because Apple has a track record of making changes with their A-series already based on observable behaviors (i.e., usage) and how the SoC is utilized in those tasks. This is a much different approach to what Intel is doing, or even what Qualcomm does with Snapdragon or Samsung with Exynos. Apple also has the advantage of designing both the hardware and software, which nobody else has (outside of Microsoft's partnership with Qualcomm).

 

[nbjustforfun]

So, title. Why did the original M1 Pro/Max only have 2 efficiency cores? Was it because apple thought they didn't need more than 2? Or was it a die space issue? Or was it simply a design decision?

Also, why did they add 2 more with M2 Pro/Max? Was it just to boost the cpu performance a bit more? Was it because they knew m2 was going to be less efficient so they added 2 more efficiency cores?

Thoughts? Why couldn't they just have 4 efficiency cores on M1 Pro/Max to start with?

Apple essentially scammed buyers KNOWING the performance cores would remain idle 98% of the time. This means the computer will be slower on day to day tasks. So swapping out performance cores for more efficiency cores on the M2 and now the M3 they can claim faster day to day performance. Here is my M2 Max studio CPU history- those performance cores are dead to me. Waste of money! My M1 16" pro looks even worse- 2 efficiency and 8 performance. Day to day tasks grind to a halt cause they get all jammed up on only 2 cores all day.

 

[pshufd]

I have an M1 Pro and what I discovered is that my workload mostly runs in the 2 e Cores with a little use from a p core. So 4 e Cores would have been beneficial for my use case. There are times when I use several of the P cores (when running a Windows VM) but I don't do that very often. The M3 Pro MacBook Pro would probably be ideal for me in extending battery life but the M1 more than meets my needs right now.

There are all kinds of customers with different needs. Some do really need a ton of power from their systems; some want monitor support, ports, RAM or storage. So I got the best thing that Apple made that I thought would meet my needs.

I do think that the way they did the E/P cores on M3 better differentiates the Pro vs Max chips.

 

[Zest28]

Because Apple didn't make this machine for answering e-mails and browsing the internet. So they thought they could increase their profit margin by excluding 2 efficiency cores.

Also Apple introduced this "fault" as an easy win for a future MacBook Pro. It's the same reason why the 16" M1 MBP has a big ass notch. You know there is no reason for the notch to be so big when it has no Face ID. But Apple has to save something for later.

 

[leman]

Apple essentially scammed buyers KNOWING the performance cores would remain idle 98% of the time. [...] Here is my M2 Max studio CPU history- those performance cores are dead to me. Waste of money!

If you don't do anything that needs P-cores, there is no reason to burn extra energy activating them. If you literally don't do anything intensive with your computer, why did you buy an expensive performance-oriented model instead of base M2 mini?

So swapping out performance cores for more efficiency cores on the M2 and now the M3 they can claim faster day to day performance.

They are not swapping out P-cores for E-cores (except on the M3 Pro, where E-cores compensate the loss of sustained performance of removing 2 P-cores). And the E-cores are getting faster every year.

This means the computer will be slower on day to day tasks. [...] Day to day tasks grind to a halt cause they get all jammed up on only 2 cores all day.

These are nonsensical statements. If your P-cores don't get activated it simply means there is no advantage to activating them.

 

[MacCheetah3]

Apple essentially scammed buyers KNOWING the performance cores would remain idle 98% of the time. This means the computer will be slower on day to day tasks. So swapping out performance cores for more efficiency cores on the M2 and now the M3 they can claim faster day to day performance. Here is my M2 Max studio CPU history- those performance cores are dead to me. Waste of money! My M1 16" pro looks even worse- 2 efficiency and 8 performance. Day to day tasks grind to a halt cause they get all jammed up on only 2 cores all day. View attachment 2338204

How paranoid/cynic/conspiracy theories err claims are so easily squashed.

How does macOS decide when to use M1's performance vs. efficiency cores?

I haven't found the answer by googling. I'm wondering because I can't make sense of what I see on my M1 MacBook Air when giving it a heavy load of BOINC jobs when on battery vs. plugged in. Screen...

apple.stackexchange.com

By the way, this also applies to:

I have an M1 Pro and what I discovered is that my workload mostly runs in the 2 e Cores with a little use from a p core. So 4 e Cores would have been beneficial for my use case. There are times when I use several of the P cores (when running a Windows VM) but I don't do that very often.

 

[mr_roboto]

Because Apple didn't make this machine for answering e-mails and browsing the internet. So they thought they could increase their profit margin by excluding 2 efficiency cores.

Also Apple introduced this "fault" as an easy win for a future MacBook Pro. It's the same reason why the 16" M1 MBP has a big ass notch. You know there is no reason for the notch to be so big when it has no Face ID. But Apple has to save something for later.

I love it when people pretend to know what they're talking about while spinning conspiracy theories.

You can fit approximately four M1 E cores in the footprint of a single M1 P core. Since E cores provide more than 1/4 the performance of P cores, E cores actually provide better performance per unit area.

If Apple wanted to penny pinch with no regard for what users of Pro/Max chips actually need, they could have skewed the chip the opposite way. For example, they could have done 2P+24E in less area than the 8P+2E configuration we actually got. But Apple knew that only a tiny fraction of power users run workloads that make a sea of E cores look good, so they took the more expensive (to them) route and designed the chip around P cores with a small number of E cores in an ancillary support role.

It does mean that under many low-load scenarios M1 Pro and Max are less energy efficient than plain M1, since Pro/Max may need to run things on P cores more often. But Pro and Max were designed to go into bigger Macs that have more battery capacity and cooling, so that's not such a bad tradeoff.

And no, the display notch obviously isn't some sandbagged thing so they can deliver artificial improvements in the future. It actually costs Apple more money to notch the display than to not notch it, and the sensor module in there does need some room to exist. Yes there are windows laptops out there with ultra tiny camera sensors, but they suffer from reduced image quality.

They may yet introduce face ID, but it'll take a redesigned sensor module - they can't just borrow the face ID sensors from iPhones, those need an even bigger space in a dimension the MBP display module doesn't have. iPhones are much thicker than the MBP display and their sensors get to take advantage of that in a big way.

 

[Zest28]

I love it when people pretend to know what they're talking about while spinning conspiracy theories.

You can fit approximately four M1 E cores in the footprint of a single M1 P core. Since E cores provide more than 1/4 the performance of P cores, E cores actually provide better performance per unit area.

If Apple wanted to penny pinch with no regard for what users of Pro/Max chips actually need, they could have skewed the chip the opposite way. For example, they could have done 2P+24E in less area than the 8P+2E configuration we actually got. But Apple knew that only a tiny fraction of power users run workloads that make a sea of E cores look good, so they took the more expensive (to them) route and designed the chip around P cores with a small number of E cores in an ancillary support role.

It does mean that under many low-load scenarios M1 Pro and Max are less energy efficient than plain M1, since Pro/Max may need to run things on P cores more often. But Pro and Max were designed to go into bigger Macs that have more battery capacity and cooling, so that's not such a bad tradeoff.

And no, the display notch obviously isn't some sandbagged thing so they can deliver artificial improvements in the future. It actually costs Apple more money to notch the display than to not notch it, and the sensor module in there does need some room to exist. Yes there are windows laptops out there with ultra tiny camera sensors, but they suffer from reduced image quality.

They may yet introduce face ID, but it'll take a redesigned sensor module - they can't just borrow the face ID sensors from iPhones, those need an even bigger space in a dimension the MBP display module doesn't have. iPhones are much thicker than the MBP display and their sensors get to take advantage of that in a big way.

Please, the M2 Max has 4E cores so it could have easily been included in the 16" M1 Max MBP no problem at all. The only reason why it is not included is because of $$$$.

The fans don't even spin up on my 16" M1 Max MacBook Pro most of the time so this machine has waaaaay more thermal capabilities than the M1 Max can provide.

 

[smirking]

Apple essentially scammed buyers KNOWING the performance cores would remain idle 98% of the time.

So now that you've cracked the code, your next machine is going to be a Macbook Air, right?

Right?

 

[pshufd]

So now that you've cracked the code, your next machine is going to be a Macbook Air, right?

Right?

No.

I need 32 GB of RAM.

I bought an M1 mini in 2021 to test and size my workload and I correctly bought a 32 GB M1 Pro MacBook Pro because the M1 mini could only go to 16 GB of RAM. But the CPU was enough for my needs. And I wanted a 16 inch display to run at native resolution. 4k would have been better but Apple doesn't seem to do standard native resolutions anymore.

And then I bought a Studio for my desktop because I wanted to run with 4 external monitors. If you want to run with 4 external monitors, then you need to go with the Max. I tested DisplayLink on the mini and it's subpar.

If you need one aspect of the Pro/Max/Ultra systems, then you get them all. It's not really that bad though because the leap is so high compared to the Intel machines.

 

[esou]

I'm pretty sure Apple put only 2 e-cores because 4 cores wouldn't fit nicely.

A 4 e-core cluster would have been slightly larger, and would have pushed other components out of the way. That would have required a large reorganization of the layout.

From an efficiency, a performance, or even a density point of view it doesn't really make a lot of sense to remove e-cores. The E-cores are more efficient and denser than P-cores. By putting more of them Apple could achieve better performance while using less space and less energy. A 4P+32E would be faster, denser and more efficient than a 8P+2E.

Now, on the other side of the equation, I think Apple has favored heavy P-core configurations because they are having problems scaling the core count. Apple has always designed tiny, ultra-power efficient SoCs. Their SoC design focuses on efficiency first and foremost, leaving scalability and flexibility behind. This is pretty evident when you see how Apple continues to use core clusters with shared L2 cache.

To pack more cores, Apple would either

1. Need to make large clusters, where the shared cache would be under extreme contention. This seems to be where Apple is heading, with the 6 core clusters we have seen on M3.
2. Add many clusters, which would crush the space efficiency if the cores cannot access the other cluster cache or generate tons of noise across clusters if they can.
3. Redesign their cores with private L2 caches and a new memory subsystem with slow and scalable L3 cache (where everyone else went).

 

[aj_niner]

Product segmentation for typical use case of the next 4 years.

 

[leman]

I'm pretty sure Apple put only 2 e-cores because 4 cores wouldn't fit nicely.

A 4 e-core cluster would have been slightly larger, and would have pushed other components out of the way. That would have required a large reorganization of the layout.

From an efficiency, a performance, or even a density point of view it doesn't really make a lot of sense to remove e-cores. The E-cores are more efficient and denser than P-cores. By putting more of them Apple could achieve better performance while using less space and less energy.

It's also possible that they thought their high-performance designs did not need more than 2 E-cores, or maybe there were additional constraints in how the internal chip fabric works. We simply don't know. One can speculate about it of course, but without more information the speculation is probably not going to yield any useful revelations.

A 4P+32E would be faster, denser and more efficient than a 8P+2E.

Faster for some very narrowly defined parallel workloads, surely. Faster on a wide variety workloads Apple and their users care about? Unlikely. Don't forget the added routing complexity and thread migration overhead. Apple Silicon inter-cluster latency is already extremely high (150-200ns), with that many E-core clusters it would become unbearable.

 

[pshufd]

Faster for some very narrowly defined parallel workloads, surely. Faster on a wide variety workloads Apple and their users care about? Unlikely. Don't forget the added routing complexity and thread migration overhead. Apple Silicon inter-cluster latency is already extremely high (150-200ns), with that many E-core clusters it would become unbearable.

I'd love to try out my workload on an M3 Pro. My guess is that the P cores would be used very little. Basically startup of my professional workload and when running Windows. Scaling is always an issue as you have diminishing returns outside of specific workloads. I have to say that the AS MacBook Pros (Pro and Max) are very nice machines, even if over-spec'd for many users.

 

[dgdosen]

It's interesting to look back on this 'old' thread and see thoughts from early 2023.

With respect to CPU cores - the High Yield youtube video - "M3, M3 Pro & M3 Max — Chip Analysis" described that on all the M chips - the CPU cores take up < 3% of die space - and that percentage has decreased over time (although transistor count increased). GPU die space has grown to 35% of die space.

Doesn't that imply that Apple could ship a CPU powerhouse by moving that die allocation between CPU/GPU more towards CPUs?

 

[leman]

It's interesting to look back on this 'old' thread and see thoughts from early 2023.

With respect to CPU cores - the High Yield youtube video - "M3, M3 Pro & M3 Max — Chip Analysis" described that on all the M chips - the CPU cores take up < 3% of die space - and that percentage has decreased over time (although transistor count increased). GPU die space has grown to 35% of die space.

Maybe 3% per individual core, the CPU clusters take much more than that. GPU on the Max is around half the die I think.

Doesn't that imply that Apple could ship a CPU powerhouse by moving that die allocation between CPU/GPU more towards CPUs?

Well, we knew that from day one. If Apple wanted to go after the server market they would have murdered everyone already.

 

[esou]

Faster for some very narrowly defined parallel workloads, surely. Faster on a wide variety workloads Apple and their users care about? Unlikely. Don't forget the added routing complexity and thread migration overhead. Apple Silicon inter-cluster latency is already extremely high (150-200ns), with that many E-core clusters it would become unbearable.

I would argue that most workloads that scales to 8 cores, can scale to more cores.

And about the intra-cluster latency, yeah, that's part of the problems of scaling. And I did mention the scaling problems that come from both packing cores in clusters or number of clusters.

 

[Sydde]

Does anyone know if AS E-cores are functionally equivalent to the P-cores? Because, I have heard that *Lake E-cores are missing some features compared to the P-cores.

 

[leman]

Does anyone know if AS E-cores are functionally equivalent to the P-cores? Because, I have heard that *Lake E-cores are missing some features compared to the P-cores.

So far it seems they are functionally equivalent, at least according to sources I’ve seen. E-clusters even have their own AMX unit.

 

[name99]

So, title. Why did the original M1 Pro/Max only have 2 efficiency cores? Was it because apple thought they didn't need more than 2? Or was it a die space issue? Or was it simply a design decision?

Also, why did they add 2 more with M2 Pro/Max? Was it just to boost the cpu performance a bit more? Was it because they knew m2 was going to be less efficient so they added 2 more efficiency cores?

Thoughts? Why couldn't they just have 4 efficiency cores on M1 Pro/Max to start with?

I suspect it's as simple as they weren't quite sure how much code would land up running on E-cores, in other words how useful they would be, especially for desktop machines. Maybe the limited amount of code the OS and apps would mark as "background, low priority" would be so small that two cores could easily handle it?

The pattern since then has been a combination of Apple learning just how much code can run well on E-cores and developers likewise learning the same thing (usually lagging by a year or so).
Simultaneously Apple keeps rearchitecting the OS into more, smaller pieces that can run well on an E-core (certainly anything OS-related that either involves waiting, for network, IO, or whatever; or anything unpredictable, so that all the caching and speculation smarts in a P-core are of limited utility).
Simultaneously with that the E-core just keeps getting better in performance while retaining that low energy.

So I'd see it more as "we tried 2 cores, and realized there was a lot more we could do with E cores. So then we tried 4 cores, and they were still frequently overloaded. So now with M3 Pro we're trying 6 E-cores!"
Who knows, maybe next year, based on all this experience, we'll see something like M gets 6 E-cores, and Pro gets 8 E-cores? Obviously there's no point in adding extra E-cores if they are rarely used. On the other hand, why NOT add them if they often all can be used, since they are small?
It's not an exact science

It feels like Intel is doing more or less the same thing, creating hybrids of every possible P to E ratio and seeing how each sell, because they also don't really have a good idea as to which is more useful for which classes of users.
Apple has a rather different business model from Intel, so they can't experiment with creating 17 different variants of the M1 (or M3) each with slightly different P, E, and GPU counts; instead they have to do their experiments as they do, with things like "two E cores for the M1 Pro, then four E-cores for the M2 Pro, then six vs four E-cores for the M3 Pro vs Max" The 6 E-cores make the M3 Pro feel like a sidegrade rather than a downgrade relative to the M2 Pro, while the 6 E vs 4 E on the Max vs Pro allows Apple to see whether 6 E-cores has actual real value at the high end or the sweet spot really is just 4 E-cores everywhere all the time.

 

[dgdosen]

Does anyone know if AS E-cores are functionally equivalent to the P-cores? Because, I have heard that *Lake E-cores are missing some features compared to the P-cores.

I found this analysis of P/E cores pretty interesting:

Evaluating M3 Pro CPU cores: 1 General performance

Using assembly language test loops to understand the differences between M1 Pro and M3 Pro CPU cores casts new light on their differences.

eclecticlight.co

Makes it look like those E cores are pretty versatile.

 

[Basic75]

It feels like Intel is doing more or less the same thing, creating hybrids of every possible P to E ratio and seeing how each sell, because they also don't really have a good idea as to which is more useful for which classes of users.

I don't think Apple's approach to P and E cores is comparable to Intel's. Apple is trying to reduce energy consumption, Intel is trying to maximise multi-threaded performance. That's why Intel's E cores are more about area efficiency while Apple's are more about energy efficiency. Also Intel has more pressure on pricing, so they are more likely to put 4 E cores in the space of 1 P core because together they are about twice as fast but only use about the same area. Large chips are expensive to make. From that viewpoint Intel's strategy makes a lot of sense to me. It also explains why Intel's E cores are faster than Apple's, and why the difference between P and E cores is smaller for Intel.

 

[pshufd]

I don't think Apple's approach to P and E cores is comparable to Intel's. Apple is trying to reduce energy consumption, Intel is trying to maximise multi-threaded performance. That's why Intel's E cores are more about area efficiency while Apple's are more about energy efficiency. Also Intel has more pressure on pricing, so they are more likely to put 4 E cores in the space of 1 P core because together they are about twice as fast but only use about the same area. Large chips are expensive to make. From that viewpoint Intel's strategy makes a lot of sense to me. It also explains why Intel's E cores are faster than Apple's, and why the difference between P and E cores is smaller for Intel.

I really do wish that Intel would put more effort into efficiency.

I keep reading about their newer generations of CPUs and I'd have to upgrade my PSU to use the newer stuff.

It seems like power is no object in their quest for performance.

I'm looking at replacing the GPU on my Windows desktop and I'd like to replace the motherboard and CPU as well; Intel just is not compelling at this time. I'd also like to do a build where the motherboard is good for 4-5 generations. I don't know if that will ever be the case going forwards.