M3 Chip Generation - Discussion Megathread

[leman]

Work done for an hour... People, please stop treating the A17 Pro as a desktop chip to drive home academic points. A17 Pro performance (or lack of) should be discussed in light of the typical work load of an average iPhone user. That being said, the N3 node does not yet seem to be the salvation many hoped for prior to the A17 Pro release.

But what if A17 Pro is a desktop chip? I hardly think that this is just an academic discussion, Mac business is not some theoretical enterprise. The hypothesis that A17 cores are developed to take better advantage of the available thermal range on a desktop would explain the performance and power consumption data we observe with these new processors, so I find this to be a compelling line of though. Not to mention that there is additional evidence for this, like the increase in size of internal OoO structures (Apple has been shrinking them in A15/A16 to improve efficiency). At least the "desktop hypothesis" appears to be explaining what we see a bit better than the hypothesis that either Apple or N3B is failing — there are many things they could have done if their target was simply to develop a more efficient smartphone CPU.

For the average iPhone user, A17 Pro is roughly comparable to A16, which is already the fastest smartphone CPU on the market by a wide margin. So I don't see any concessions being made here.

 

[Confused-User]

That's exactly what I was referring to. We know it's just Apple buying these chips so if they were going solely into A17s I think there's a good chance we would have heard that by now, especially if there was a change of plans from N3B to N3E (which would have been a while ago). Combine that with the fact that M1 and M2 are both built on the same process as the preceding A-series chip and that volume production of N3E didn't begin until a few months before Apple's upcoming event (originally-planned for this month) event and it seems safe to assume that M3 will be N3B-based.

Oh, I see what you mean.

I thought this myself until recently. But now it seems quite up in the air. The most common expectation now (and I hope it's wrong! But I have doubts) is that the M3 is NOT coming this year. This is one key factor. The other is the widespread expectation that the A17 cores are already being reimplemented on N3E for the regular A17 (or A18?) due in next year's iPhone 16. That one I do believe, because it makes a ton of sense.

So, if the M3 is not coming until 2024, and the A17 cores are being reimplemented in N3E anyway, it's entirely plausible that they'll come out first in the M3.

Even if that's wrong, it's not likely that a half million M3s for the AVP is going to make any difference at all to TSMC a couple months into 2024. At that point, iphone demand is way down from peaks, supply is all caught up, and there's plenty of capacity.

There is no indication at all that TSMC is having trouble making enough chips on N3, given the actual demand, and lots of indication that they have spare capacity. There's indication that it's pricier than they'd like, but that's not due to supply constraints.

 

[PgR7]

Ahh ... then it is not the same amount of work done in the 1 hour period. Are you sure you understand the concept?

Are you then not also juggling to diss Apple using the A17 Pro as a weapon?

It is the same work, thats why they take different time. Everyone looking at those numbers of performance and consumption with a minimun knowledge in hardware and overclock knows that. Im not juggling im just doing basic maths, mmmm let me guess, a processor 10% faster takes 50% more power consumption...mmm I wonder what CPU is more efficient...Im saying at máximum clocks the A17 is less efficient than the A16, in some loads you dont get that

 

[quarkysg]

It is the same work, thats why they take different time. Everyone looking at those numbers of performance and consumption with a minimun knowledge in hardware and overclock knows that. Im not juggling im just doing basic maths, mmmm let me guess, a processor 10% faster takes 50% more power consumption...mmm I wonder what CPU is more efficient...Im saying at máximum clocks the A17 is less efficient than the A16, in some loads you dont get that

Actually, after reading all your replies, you presented your thoughts without any substantial figures to back up your assertion.

Take for example, your 1 hour of work between A17 Pro and A16, with the w-h computed, and declared that A17 Pro sucks in terms of efficiency. But when asked how much work is done, there's nothing to quatify the work done. How then is efficiency derived then?

Two person working non-stop for 1 hour does not make both equally efficient, assuming they both take the same amount of energy. We also need to take into account the amount of work completed by each person to derive efficiency.

If one person needed 20% more food but completed 50% more work in the same hour, who is more efficient? You seem to think/suggest that the one eating less food is more efficient?

 

[PgR7]

Let's assume that you're correct about this (and I think you are in many cases, though whole-system power is more complex than A17 power, and it's not always going to be that way).

You still seem to be missing completely the point I was making (and that @name99 was making, too). This isn't a bad thing, it's a good thing. Making the system more responsive, "snappier", is generally a win. Pushing it too far, of course, you get overheated chips and poor battery life. So what makes this the right choice?

That answer is subjective, of course. You can argue that for your particular use case and preferences, Apple's choice is suboptimal. You can't however say that they built a bad product.

I suspect that what actually happened was that Apple looked at the whole system and said "we want battery life to be as good as last year, and we'd like better performance, what can do within those constraints?". And the 15Pro is what we got. I am quite happy with their choice, and so are lots of other people. You wanted more battery and are disappointed? Fair, but not a reason to say their product has a problem. They just built something not optimal for you.

Mmm no, the way It actually is "We Apple prefer more consumption in more proportion than the proportion of the performance gains". This phone will be a phone where the battery life will depend a lot about your usage.

 

[PgR7]

Actually, after reading all your replies, you presented your thoughts without any substantial figures to back up your assertion.

Take for example, your 1 hour of work between A17 Pro and A16, with the w-h computed, and declared that A17 Pro sucks in terms of efficiency. But when asked how much work is done, there's nothing to quatify the work done. How then is efficiency derived then?

Two person working non-stop for 1 hour does not make both equally efficient, assuming they both take the same amount of energy. We also need to take into account the amount of work completed by each person to derive efficiency.

If one person needed 20% more food but completed 50% more work in the same hour, who is more efficient? You seem to think/suggest that the one eating less food is more efficient?

?? With the data of the image you get the performance of both, that is the amount of work per unit of time, and you get the power consumption of both CPUs. What It is that you dont understand

"If one person needed 20% more food but completed 50% more work in the same hour, who is more efficient? You seem to think/suggest that the one eating less food is more efficient?"
That person is more efficient like the A16 compared to the A17

 

[Confused-User]

Mmm no, the way It actually is "We Apple prefer more consumption in more proportion than the proportion of the performance gains". This phone will be a phone where the battery life will depend a lot about your usage.

I'm sorry, I don't understand your words. But I think you're saying that Apple gave up efficiency in exchange for performance. To some extent, depending on workload, yes, they did.

I'm not clear on why you keep returning to this. That's obvious from the numbers we've seen. What's not obvious is why this matters. In the M3, this will be a good thing: the ability to trade off some efficiency for raw performance is greatly desired by nearly everyone. In the iPhone, where the A17 actually is, the only impact is on battery. If you're one of the rare people for whom the battery is not sufficient (after the promised bug fixes coming soon form Apple), then, for you Apple made a poor tradeoff. For most people... it's a good tradeoff even for phones.

Why is this hard to understand?

 

[Confused-User]

Actually, after reading all your replies, you presented your thoughts without any substantial figures to back up your assertion.

I don't think that's fair. They were using the figures posted by someone else here. And probably some others posted earlier.

Take for example, your 1 hour of work between A17 Pro and A16, with the w-h computed, and declared that A17 Pro sucks in terms of efficiency. But when asked how much work is done, there's nothing to quatify the work done. How then is efficiency derived then?

There is clearly a quantification, based on that chart, and their conclusion is correct: each running at their maximum clocks, the A17 isn't as efficient as the A16. This should surprise nobody. But it's also not very interesting, except to the extent that it showed us some early indicators about Apple's future intentions for this design.

After that, they lose me. There seems to be some confusion about whether this is supposed to mean that the A17 is a bad design, N3 is a bad process, or Apple made bad choices. I don't think any of those things are true.

 

[PgR7]

I'm sorry, I don't understand your words. But I think you're saying that Apple gave up efficiency in exchange for performance. To some extent, depending on workload, yes, they did.

I'm not clear on why you keep returning to this. That's obvious from the numbers we've seen. What's not obvious is why this matters. In the M3, this will be a good thing: the ability to trade off some efficiency for raw performance is greatly desired by nearly everyone. In the iPhone, where the A17 actually is, the only impact is on battery. If you're one of the rare people for whom the battery is not sufficient (after the promised bug fixes coming soon form Apple), then, for you Apple made a poor tradeoff. For most people... it's a good tradeoff even for phones.

Why is this hard to understand?

Finally you admit It, for me and lots of people is not enough, lots of people comment that 1 year later his 14 PM dont do 1 day on 1 charge, so imagine if they make It last to 3-5 years. And no, why I have to do every few years a battery replacement, what if they dont do battery replacements anymore. Efficiency is always good in every device, in phones good battery life, more sustained performance, in desktops less fan noise, less heat in the room, less electrical bill.

 

[quarkysg]

I don't think that's fair. They were using the figures posted by someone else here. And probably some others posted earlier.


There is clearly a quantification, based on that chart, and their conclusion is correct: each running at their maximum clocks, the A17 isn't as efficient as the A16. This should surprise nobody. But it's also not very interesting, except to the extent that it showed us some early indicators about Apple's future intentions for this design.

After that, they lose me. There seems to be some confusion about whether this is supposed to mean that the A17 is a bad design, N3 is a bad process, or Apple made bad choices. I don't think any of those things are true.

Actually the diagram shows the wattage consumed while doing the Spec benchmark. I don’t see any timing. Efficiency cannot be inferred from that chart.

Many folks says Nvidia GPUs are more efficient as it complete more work at the same time period although consuming obscene amount of energy.

So my point is there’s just not enough data to say one way or another.

@leman’s work in another thread suggest A17 Pro is more efficient at the same clock, because his benchmark made the A16 and A17 Pro do the same work at the same frequency, IPC increase not-withstanding.

In any case the poster I’m responsible by seems to have an axe to grind. So I think I’ll just leave it at that.

 

[Confused-User]

Actually the diagram shows the wattage consumed while doing the Spec benchmark. I don’t see any timing. Efficiency cannot be inferred from that chart.

As far as I can see, run time and power are both on that chart.

Many folks says Nvidia GPUs are more efficient as it complete more work at the same time period although consuming obscene amount of energy.

You can't fix stupid. So what?

So my point is there’s just not enough data to say one way or another.

@leman’s work in another thread suggest A17 Pro is more efficient at the same clock, because his benchmark made the A16 and A17 Pro do the same work at the same frequency, IPC increase not-withstanding.

Yes, as I have said here and there more than once for the benefit of the clueless. (And the IPC increase, modest as it is, contributes to that.) But leman's chart supports this poster's contention. The problem is that his point is not interesting or meaningful to nearly anyone.

In any case the poster I’m responsible by seems to have an axe to grind. So I think I’ll just leave it at that.

Yeah, I'm done with him too.

 

[leman]

It is indeed the case that operating at maximal load over prolonged period of time the A17 Pro CPU is less efficient than A16. Running my stress test for example, A17 Pro about 10% less efficient in single core (in multi core their efficiency is practically identical, but both are much less efficient than A14). On more complex workloads the difference will likely shrink thanks to increased IPC on A17.

But I would caution agains extrapolating from these kind of stress tests to any real world usage. That's a very different cup of tea. I mean, by @PgR7's logic, both A15 and A16 based model should have considerably lower battery life than the iPhone 12, but that's not what we are observing in practice. And sure, they have larger batteries (as does the 15 series btw), but their display is also brighter. And in my tests I see 25-30% difference in MC efficiency for A14 and A16 for example.

 

[MRMSFC]

Lmao, I guess I Will have to leave this forum, I expected hardware entusiasts, not Apple fanboys, I have never seen something like that but then just look at the name of the website what do you expect lol. Apple never do anything bad, you must be doing something wrong

I don’t think it gets more enthusiastic than plotting graphs comparing performance between two phones, based one ones own research. A la leman’s thread.

And sure there’s a level of fanboyism here, who else would join a forum for Apple enthusiasts except Apple enthusiasts and people who like to troll them.

For example:

It is indeed the case that operating at maximal load over prolonged period of time the A17 Pro CPU is less efficient than A16. Running my stress test for example, A17 Pro about 10% less efficient in single core (in multi core their efficiency is practically identical, but both are much less efficient than A14). On more complex workloads the difference will likely shrink thanks to increased IPC on A17.

But I would caution agains extrapolating from these kind of stress tests to any real world usage. That's a very different cup of tea. I mean, by @PgR7's logic, both A15 and A16 based model should have considerably lower battery life than the iPhone 12, but that's not what we are observing in practice. And sure, they have larger batteries (as does the 15 series btw), but their display is also brighter. And in my tests I see 25-30% difference in MC efficiency for A14 and A16 for example.

Fairly points out that there’s more changes to the iPhone than just the processor, and that we can’t necessarily isolate the increased battery draw to decreased efficiency.

 

[MrGunny94]

I'm still wondering whether the M3 Air will bring OLED and base 12GB of RAM.

That might convince me to get a 2nd Apple Silicon machine.. Would leave the 14" Pro at home when travelling...

 

[name99]

Am I missing something here? Assuming that their published data is correct, it's not 21%, it's 38%. (I think you subtracted 17 twice?) That does not result in a favorable energy-delay product.

Of course, I'm not trusting their data without backup. And if their data is correct, it's likely still a good tradeoff given the bursty nature of most smartphone use.

(a) power = energy / time. therefore energy = power*time
(b) 1/(1+epsilon)~1-epsilon and similarly for (1+eps1)(1+eps2)~1+eps1+eps2
We agree?

OK, then power goes up by 38%. But time has gone down by 17%. So energy has gone up by?
(1+.38)=(1+.17)*(1+x). x=21%
Therefore energy has gone up by 21%.

This is the last I will post on the issue.
If people don't know enough physics 101 to understand the argument, go read a book!

 

[name99]

But is the amount of work done in that 1 hours the same for the A17 Pro and A16? Do you know that?

In principle we do in that these numbers are for SPEC which is a constant amount of work.
The presentation is garbage but essentially we CAN conclude:
- the amounts of work are equal
- it's not completely accurate to say that a 17% higher SPEC score means the block of work (ie the set of benchmarks) completed in 17% shorter time, but it's "good enough" if the two microarchitectures are similar and the scores are similar
- likewise reporting a "power" for the run is crazy, but assuming they're not just making **** up, then presumably what they ACTUALLY means is "total energy to execute the run"/"duration of the run" which is kinda sorta an average power level, and can at least allow us to back out the total energy used for the benchmark run.

 

[Chuckeee]

This is the last I will post on the issue.
If people don't know enough physics 101 to understand the argument, go read a book!

In that case:

A man walks into a hardware store and speaks to the cashier...
"Have you any two watt bulbs?"
"For what?"
"That’ll do, I'll take two."
"Two what?"
"I thought you didn’t have any."
"Any what?"
"Yes please!"


I was gonna tell another joke about Electricity, but I forgot Watt it was.

 

[Confused-User]

(a) power = energy / time. therefore energy = power*time
(b) 1/(1+epsilon)~1-epsilon and similarly for (1+eps1)(1+eps2)~1+eps1+eps2
We agree?

OK, then power goes up by 38%. But time has gone down by 17%. So energy has gone up by?
(1+.38)=(1+.17)*(1+x). x=21%
Therefore energy has gone up by 21%.

This is the last I will post on the issue.
If people don't know enough physics 101 to understand the argument, go read a book!

:-(

This will teach me not to post (long) after bedtime. Sorry, that was obvious.

 

[smalm]

But what if A17 Pro is a desktop chip? I hardly think that this is just an academic discussion, Mac business is not some theoretical enterprise.

Whatever Apple may do be assured their $200B iPhone bussiness always comes first.
The A17 is built for the iPhone and it is what is is because Apple want that in the iPhone.

(1+.38)=(1+.17)*(1+x). x=21%

Maynard! Please!

 

[APCX]

Whatever Apple may do be assured their $200B iPhone bussiness always comes first.
The A17 is built for the iPhone and it is what is is because Apple want that in the iPhone.


Maynard! Please!

If they could serve their very lucrative Mac business while also serving their even more lucrative phone business, why wouldn’t they? Those two don’t seem to be in opposition.

 

[quarkysg]

In principle we do in that these numbers are for SPEC which is a constant amount of work.
The presentation is garbage but essentially we CAN conclude:
- the amounts of work are equal
- it's not completely accurate to say that a 17% higher SPEC score means the block of work (ie the set of benchmarks) completed in 17% shorter time, but it's "good enough" if the two microarchitectures are similar and the scores are similar
- likewise reporting a "power" for the run is crazy, but assuming they're not just making **** up, then presumably what they ACTUALLY means is "total energy to execute the run"/"duration of the run" which is kinda sorta an average power level, and can at least allow us to back out the total energy used for the benchmark run.

Well, my response to the poster throwing out the 1 hour figure (post #1,042) was because those figures are arbitrarily posted (IMHO anyway) to assert a point. Poster claim that at full tilt, A16 vs A17Pro, doing "some" work, the A16 needed 1.17 hours compared to 1 hour for the A17 Pro. Using that to compute energy used for the "work" which was never defined.

So far, the only hard figured we have at the moment are from @leman in which it shows the same work done for both A16 and A17 Pro based on clock freq. and power consumed. That graph shows that A17 Pro is indeed more efficient than A16. It is just that A16's max freq is limited, while the A17 Pro has been designed to use more if given. So declaring A17 Pro suck due to that is a little ingenious IMHO. Instead of having a civil debate poster gets hot and bothered and started calling names. Kind of in-mature and rude. On the other hand, the work from @leman is looking at a very narrow subset of the CPU cores, so it may not be representative of actual, but it gives another data point of reference.

As for the SPEC results, there are a few variables. Altho the work done is the same, the results are different, as the A17 Pro scored higher. And I do not seem to remember seeing the time it took both A16 and A17 Pro to complete said benchmark that arrived at their respective scores. So using that to then declare one is more efficient over another is a little ingenious IMHO.

My understanding of SPEC is that it is used to benchmark performance of a system, not efficiency, as it was not designed to determine a system's efficiency.

IMHO, whether the A17 Pro is more or less efficient over the A16 is unclear.

But it amuses me to no end that some random poster in an Internet forum thinks they are a lot smarter than Apple's SoC architects and designers tho.

Happy to be corrected tho.

 

[leman]

So far, the only hard figured we have at the moment are from @leman in which it shows the same work done for both A16 and A17 Pro based on clock freq. and power consumed. That graph shows that A17 Pro is indeed more efficient than A16. It is just that A16's max freq is limited, while the A17 Pro has been designed to use more if given. So declaring A17 Pro suck due to that is a little ingenious IMHO. Instead of having a civil debate poster gets hot and bothered and started calling names. Kind of in-mature and rude. On the other hand, the work from @leman is looking at a very narrow subset of the CPU cores, so it may not be representative of actual, but it gives another data point of reference.

Careful here! In my tests, A17 uses more energy per workload unit than any other A-series chip. It is indeed uses less power at the same frequency than A16, but its average operating frequency is higher. But as mentioned, my tests are designed to stress the CPU and study the maximal power consumption per core, not to evaluate how long the CPU needs to run useful work!

For practical energy efficiency discussions, I think it makes more sense to look at more circumstantial evidence like battery life. At the end of the day, it's the practical usability that matters. In the new review by notebookcheck (who are usually quite good at it) they show that iPhone 15 pro max outperforms any previous iPhone in pretty much every category except battery life under continuous high load. Which exactly confirms the notion that the system will consume more power when pushed hard, but in actual practical use it ends up being more efficient.

 

[mr_roboto]

Two big issues at play there. First and foremost is the limited production of 3nm silicon in comparison to the 5nm silicon used in every current product NOT named iPhone 15 Pro/Pro Max. Apple is not going to cannibalize an already limited supply of 3nm parts in order to bump the specs for Vision Pro, especially when it's likely to sell considerably fewer units given the ~$3000 price point. The second issue is the fact that it's not the M2 which handles all the input from the cameras, sensors, microphones, etc. That is the job of the R1 chip:

Focusing on M2 vs. M3 actually overlooks what is doing the bulk of the work inside Vision Pro, and I'd bet the farm that Apple isn't about to replace R1 just a few months before VP actually launches.

I wouldn't discount the M2's role in Vision Pro quite as much as that. The Vision Pro platform is essentially a Mac plus a bunch of cameras and sensors plus the R1 chip. The R1 chip is best understood as specialized accelerator peripheral for the Mac, since visionOS and its apps run on the M2, not the R1. And if they had to do it, I doubt there would be any technical challenge in using R1 with M3 instead of M2.

I don't think supply plays any role here either, as I doubt they're anticipating the first generation becoming a high volume hit the way iPhone did. Vision Pro is way too expensive and niche for that.

That said, I agree with your prediction that Vision Pro will launch with M2. What Apple needs is production units in developer and early adopter hands, ASAP. They don't need to disrupt the schedule even a little bit by integrating the next generation SoC; that's not going to make or break the product.

 

[leman]

BTW, has it been mentioned that since A17 has external display support, M3 is likely to support two displays?

 

[diamond.g]

BTW, has it been mentioned that since A17 has external display support, M3 is likely to support two displays?

Wouldn't you mean 3 displays? M1/M2 technically support 2 displays already right?