Power curve and efficiency of Apple N3

[Love-hate 🍏 relationship]

Might be better to wait until they lose the crown to start boasting/dooming.

Boasting ? If anything you're the one defending a company with a personal level of vehemence

I was pointing out that since his departure development plateaued, which is quite factual

 

[jeanlain]

I was pointing out that since his departure development plateaued, which is quite factual

You presented no fact supporting your argument.

 

[Love-hate 🍏 relationship]

You presented no fact supporting your argument.

True. I mean I thought the arch improvements were self explanatory . the lack of a bigger instruction set implementation (only this year did we see an increase) coupled with the fact that there has been no real improvements to the CPU other than higher clocks (since A14 )

 

[APCX]

Boasting ? If anything you're the one defending a company with a personal level of vehemence

I was pointing out that since his departure development plateaued, which is quite factual

Hmmm. That’s not quite how I see it. Have you been following this thread at all? The A17 represents a significant improvement interns of absolute performance (~400 GB6 points) and is more efficient simultaneously. It improvement in gpu performance and features has allowed it to have actual console level games on a phone for the first time, significantly increasing the addressable market for software developers.

To state development has plateaued, is not just not factual…it’s anti-factual.

You’ve posted a thread stating this and are surprised by any pushback. Instead spouting something about “personal level of vehemence”. Which to be honest I don't understand the meaning of. I don’t feel I’ve defended Apple. I’m just trying to discuss things as I see them. Everyone is free to disagree.

 

[APCX]

...fact that there has been no real improvements to the CPU other than higher clocks (since A14 )

What exactly do you mean by this? This seems to me to be counter-factual. The word “real’ here is key. It seems to be a weasel word used to evade any criticism. You can redefine what “real” means at will.

We have had: absolute performance improvements, efficiency improvements, ipc improvements, gpu improvements.

Are you expecting people to believe those things don't constitute a “real” improvement?

 

[leman]

I was pointing out that since his departure development plateaued, which is quite factual

No it’s not. Apple has delivered consistent performance improvement every year. And A17 is the widest commercial CPU architecture on the planet.

 

[Xiao_Xi]

A17 represents a significant improvement interns of absolute performance (~400 GB6 points) and is more efficient simultaneously.

Either you compare the SoCs at the same frequency or at the maximum frequency, but mixing results is misleading. If you compare them at the maximum frequency (Geekbench scores), the A17 is not more efficient than the A16. If you want to say that the A17 is more efficient (at the same frequency), you have to compare Geekbench scores at the clock frequency of the A16, i.e. IPC increase.

A17 is the widest commercial CPU architecture on the planet.

What about Cortex-X4?

 

[Confused-User]

Either you compare the SoCs at the same frequency or at the maximum frequency, but mixing results is misleading. If you compare them at the maximum frequency (Geekbench scores), the A17 is not more efficient than the A16.

I was going to say that that's not what "efficiency" means, but that's not quite right. I guess the best way to say it is that if you define it that way, you have a not very useful word.

In the context of looking at the SOC - which is what you're doing, and what this thread is about - the useful way of looking at efficiency is as a curve. Like, the way leman plotted it. That's because it can answer definitively a simple question: If you replace the A16 with the A17 while keeping EVERY other factor constant, what happens? In this case the answer is, you always save energy.

If you are talking about specific implementations of the SoC - like, say, the iP15P, then your definition is more useful, because it says something (not everything) about battery life. The 15Pro doesn't keep everything constant compared to the 14Pro - they changed virtually everything. So you have situations right now where the 15Pro may use more energy to accomplish the same human-level task (even though on a lower level it's actually doing more work, we don't notice or care). Most or all of those situations are likely due to bugs that will shortly be resolved. The big exception is in gaming, where there's open-ended demand for performance in the form of higher frame rates. For those, it's a judgement call what the chip does. You can opine on the choices of the people who made the device, and the OS it's running, and the application it's running. You don't get to use that to talk about the efficiency of the SoC, because that's simply false reasoning.

If you want to say that the A17 is more efficient (at the same frequency), you have to compare Geekbench scores at the clock frequency of the A16, i.e. IPC increase.

We have some nearly equivalent info. It shows small improvements in the A17. You can however also talk about efficiency at the same work level, as opposed to the same frequency, by comparing energy used. That also shows the A17 to be better.

What about Cortex-X4?

Vaporware so far, no?

I wouldn't expect too much from that anyway. ARM's performance claims for previous generations were greatly overstated.

 

[name99]

True. I mean I thought the arch improvements were self explanatory . the lack of a bigger instruction set implementation (only this year did we see an increase) coupled with the fact that there has been no real improvements to the CPU other than higher clocks (since A14 )

Since you clearly have no idea what you are talking about (in terms of changes to the SoC since A14) please go waste other people's time.
This is a nice forum for serious people talking about engineering issues. If you want to pick a fight, go do so with people who prefer fighting to learning.

 

[name99]

What exactly do you mean by this? This seems to me to be counter-factual. The word “real’ here is key. It seems to be a weasel word used to evade any criticism. You can redefine what “real” means at will.

We have had: absolute performance improvements, efficiency improvements, ipc improvements, gpu improvements.

Are you expecting people to believe those things don't constitute a “real” improvement?

Ignore him. When dog barks at you, you won't achieve anything by barking back.
He's clearly uninterested in understanding, he just wants to fight.

 

[Confused-User]

True. I mean I thought the arch improvements were self explanatory . the lack of a bigger instruction set implementation (only this year did we see an increase) coupled with the fact that there has been no real improvements to the CPU other than higher clocks (since A14 )

I refer you to my response to Retskrad: #187

Now in fact, by pulling in your horizon, you are being less ridiculous than he was, when it comes to the CPU cores. That doesn't make you any less wrong. The clock boosts are a significant architectural victory, not "just a node change", because clocking up a slow wide design (that still happens to be a performance champion) by 22%, without a loss of IPC, is no small feat. Further, we have good reason to believe that that 22% is limited by the phone form factor and power draw, and that there's a lot more headroom in the design, though we won't know that until the M3s ship.

You referred to "the CPU" and if you truly meant only the CPU cores, then that's that. But if you were talking about the SoC in general - which would make sense since Apple isn't just designing CPU cores - then your errors compound. There are substantial improvements throughout, especially in the GPU.

 

[Xiao_Xi]

I was going to say that that's not what "efficiency" means, but that's not quite right. I guess the best way to say it is that if you define it that way, you have a not very useful word.

What would you call the Geekbench points/W rate?

If you replace the A16 with the A17 while keeping EVERY other factor constant, what happens? In this case the answer is, you always save energy.

The @leman graph shows that A17 is more energy efficient than A16. But it is misleading to say that A17 scores 400 GB6 points more than A16 and is still more efficient, because it is not more efficient at Geekbench.

I wouldn't expect too much from that anyway. ARM's performance claims for previous generations were greatly overstated.

It doesn't matter how Cortex-X4 performs, only whether or not it can be considered a "commercial CPU architecture". If it is, then Cortex-X4 is wider than A17. Otherwise, A17 is the widest "commercial CPU architecture"".

 

[Confused-User]

What would you call the Geekbench points/W rate?

The @leman graph shows that A17 is more energy efficient than A16. But it is misleading to say that A17 scores 400 GB6 points more than A16 and is still more efficient, because it is not more efficient at Geekbench.

Almost right. It's not more efficient at GB6 at their top clocks. It's definitely more efficient at GB6 when they are both running at the A16's top clock.

That's why talking about efficiency as if it were a scalar figure is generally not that useful when you're talking about chips. Again, if you're talking about a particular implementation - say, the iP15P - then it can be a little less useless, as I mentioned earlier. But it's still likely to be deceptive. Chips are complicated!

It doesn't matter how Cortex-X4 performs, only whether or not it can be considered a "commercial CPU architecture". If it is, then Cortex-X4 is wider than A17. Otherwise, A17 is the widest "commercial CPU architecture"".

OK, sure. For me, if there's no shipping product, it doesn't count - there's no commerce going on yet. You could make an argument that the design (all that ARM really makes) is already sold, but with no actual testable chip available, I decline to accept that definition.

 

[leman]

What about Cortex-X4?

Yes, I forgot about X4. Disregarding the fact that it's not a shipping product yet, it indeed seems to roughly comparable to A17. ARM even claims one more instruction decode/dispatch per cycle. Curious to see how it performs once we have some silicon to test.

 

[leman]

Either you compare the SoCs at the same frequency or at the maximum frequency, but mixing results is misleading. If you compare them at the maximum frequency (Geekbench scores), the A17 is not more efficient than the A16. If you want to say that the A17 is more efficient (at the same frequency), you have to compare Geekbench scores at the clock frequency of the A16, i.e. IPC increase.

I feel that you are making this way too complicated. A17 will be approximately as fast as A16 when running at the same frequency (maybe a bit faster since the IPC did improve but let’s disregard this for the time being). So no, you don’t have to compare the Geekbench scores at the same frequency as they will be the same.

 

[theorist9]

View attachment 2282343

Thanks for posting the link to your data. I tried reproducing your graph of single-threaded power vs. frequency for the A17 using the data contained in these two files, and got different results from what's shown in your graph (see below), indicating I'm not working off the same data set (e.g., I have none of the data you show between ≈ 2.7 and 3.4 GHz, but I have a lot more data below 2.1 GHz).

I can understand the latter could come from additional data you added after you made the plot (or that you only included the performance core, while I included data for both), but that wouldn't explain the missing data.

Could you please update your files to add the missing data?

Also:

1) Did your plot include data for the performance cores only, or both the performance and efficiency cores? As I mentioned, I included both. [At the bottom I also show a plot for the performance cores only.]

2) I assumed the units of energy and time are joules and seconds, respectively, such that energy/time would give me watts. I also assumed that one gets hertz from cycles/time. Is that right, or are there some hidden correction factors I need to adjust for?

Here's the gap in the data set--I have nothing between these two frequencies (Hz on left, W on right):

If I plot the data for the performance cores only, I get this:

 

[falainber]

Hmmm. That’s not quite how I see it. Have you been following this thread at all? The A17 represents a significant improvement interns of absolute performance (~400 GB6 points) and is more efficient simultaneously. It improvement in gpu performance and features has allowed it to have actual console level games on a phone for the first time, significantly increasing the addressable market for software developers.

To state development has plateaued, is not just not factual…it’s anti-factual.

You’ve posted a thread stating this and are surprised by any pushback. Instead spouting something about “personal level of vehemence”. Which to be honest I don't understand the meaning of. I don’t feel I’ve defended Apple. I’m just trying to discuss things as I see them. Everyone is free to disagree.

There are some indications that higher GB score comes with higher phone temperatures which might indicate that Apple [erroneously] raised temperature limits for A17. If they had to throttle the processor to drop the temps, GB scores might drop.

 

[Confused-User]

There are some indications that higher GB score comes with higher phone temperatures which might indicate that Apple [erroneously] raised temperature limits for A17. If they had to throttle the processor to drop the temps, GB scores might drop.

Especially in the wake of Apple's weekend announcement, the chance that there is some bug that's lifting the expected limit of performance and heat is vanishingly small.

That's not to say that the 15P can't draw more power than the 14P. But that's not a bug and it won't be changed. There *are* apparently bugs, both in apps and in the OS, that are causing some heat/battery issues, but fixing them does not involve reducing performance, according to Apple.

 

[theorist9]

There are some indications that higher GB score comes with higher phone temperatures which might indicate that Apple [erroneously] raised temperature limits for A17. If they had to throttle the processor to drop the temps, GB scores might drop.

If that's the case, I would guess that would affect the MC GB scores only (or predominantly). It's also possible the higher temperatures are more due to GPU usage, in which case the CPU GB scores should not be directly affected by any adjustments Apple makes.

 

[leman]

There are some indications that higher GB score comes with higher phone temperatures which might indicate that Apple [erroneously] raised temperature limits for A17. If they had to throttle the processor to drop the temps, GB scores might drop.

The temperature limits are dictated by the thermal capability, Apple doesn't seem to impose any additional limits. This is evident from the fact that more aggressive cooling (e.g. using a piezoelectric element) results in higher scores. But the CPU will generate more heat under high load, so under normal circumstances (no additional cooler) the peak frequency can be only maintained for a short time.

 

[theorist9]

And here is a fitted prediction curve (using fourth degree polynomial, going higher doesn't improv the fit) for A17

View attachment 2282345

P.S. Looking at this graph I can't wonder whether there is some intent behind some crucial points. It it a coincidence that A17 Pro is clocked to peak exactly at 5Ghz....

I'm afraid I'm not following. Where are you getting that the "A17 Pro is clocked to peak exactly at 5Ghz"?

 

[Pressure]

I'm afraid I'm not following. Where are you getting that the "A17 Pro is clocked to peak exactly at 5Ghz"?

If we take the graph in good faith as a real representation then there is little room for clock increase after that point because the power requirements quickly goes to infinity.

Intel single P-cores can easily consume 71 Watts. Before the High-Performance cores in the A17 Pro all Apple Silicon CPU cores have been restricted to less than 7 Watts (even in the Mac Pro and Mac Studio). The graph indicates that the High-Performance cores in the A17 Pro can scale up to 5GHz at around 16 Watt per core, which is easily manageable in a desktop chip.

We don't know where on the graph that Apple will settle (or if it is possible for the cores to clock that high), so this is all just hypothetical.

 

[theorist9]

If we take the graph in good faith as a real representation then there is little room for clock increase after that point because the power requirements quickly goes to infinity.

Intel single P-cores can easily consume 71 Watts. Before the High-Performance cores in the A17 Pro all Apple Silicon CPU cores have been restricted to less than 7 Watts (even in the Mac Pro and Mac Studio). The graph indicates that the High-Performance cores in the A17 Pro can scale up to 5GHz at around 16 Watt per core, which is easily manageable in a desktop chip.

We don't know where on the graph that Apple will settle (or if it is possible for the cores to clock that high), so this is all just hypothetical.

There's actually nothing in the graph showing it peaks "exactly" at 5 GHz. The graph shows no hard wall at 5 GHz. I.e., it doesn't show this:

Nor can we infer such a hard wall from the data Leman displayed. Instead, we could extend his curve to 5.1, 5.2, 5.3 GHz, etc. 5 GHz is simply the point at which he chose to end the drawing.

 

[leman]

Thanks for posting the link to your data. I tried reproducing your graph of single-threaded power vs. frequency for the A17 using the data contained in these two files, and got different results from what's shown in your graph (see below), indicating I'm not working off the same data set (e.g., I have none of the data you show between ≈ 2.7 and 3.4 GHz, but I have a lot more data below 2.1 GHz).

Thanks for trying to replicate the results, @theorist9! I have now uploaded the R notebook and the rendered output to the GitHub repo, it shows exactly what I did

AppleSiliconPowerTest/analysis.qmd at main · mr-mobster/AppleSiliconPowerTest

Frequency and power sampling for Apple Silicon computers - mr-mobster/AppleSiliconPowerTest

github.com

https://mr-mobster.github.io/AppleSiliconPowerTest/analysis.html

I also tried a new approach to fitting the curves using constrained regression (to force the coefficients to non-negative range). This shows even more how arbitrary those results are as one gets wildly different numbers...

1) Did your plot include data for the performance cores only, or both the performance and efficiency cores? As I mentioned, I included both. [At the bottom I also show a plot for the performance cores only.]

The data I have shown before only shows the P-core (and I filter out any samples where P-core usage is very low to avoid sampling artefacts). Are you per chance aggregating the cycles/time across P- and E-cores? Cause that would obviously result in lower estimated frequency.

2) I assumed the units of energy and time are joules and seconds, respectively, such that energy/time would give me watts. I also assumed that one gets hertz from cycles/time. Is that right, or are there some hidden correction factors I need to adjust for?

This is correct.

I'm afraid I'm not following. Where are you getting that the "A17 Pro is clocked to peak exactly at 5Ghz"?

Sorry, typo, it was supposed to be "5 watts". What I meant is that the A17 Pro is clocked exactly so that at peak frequency the power consumption is very close to 5 watts. This appears to be a number chosen by a human rather than a random result. But maybe I'm interpreting too much into all of this.

 

[leman]

If we take the graph in good faith as a real representation then there is little room for clock increase after that point because the power requirements quickly goes to infinity.

No, that's not what the graph shows and it's also not what I meant.