Power curve and efficiency of Apple N3

[leman]

There has been a lot of discussion about the higher power consumption of A17 Pro in the new iPhone and what this might mean for the new Macs. The only in-depth review (by Geekerwan) does not provide in-depth data, so many details were unclear. I wrote a simple tool that loads up CPU cores on Apple devices and measures their frequency and power consumption in single and multi-core scenarios. You can find the source code and for the tool here: https://github.com/mr-mobster/AppleSiliconPowerTest and I would be very thankful for additional device reports.

The current data is very preliminary (notably, I am missing results for M2 and A16), but I think it's already interesting to look at what we have right now. Here are the samples of thread counters, arranged in a power/frequency graph.

As you can see, even with this small data sample, the power curves emerge fairly cleanly (A14 is weird a bit, I might need to redo it several times). I don't really want to interpret too much into the data right now, but I think there are few interesting bits here:

- A17 does use significantly more power than the previous A-series in the usual operational range
- A17 is more efficient that either A14 and A15
- A17 and A16 are very close
- A17 runs about 3% faster than A15 at 2.5watts, about 7% faster than A15 at 3.5 watts, and about 15% faster than M1 at ~5 watts
- A17 consumes 1.5 watts less to reach the same frequency than A15/M2 at the peak (that's > 20% improvement in efficiency!)

It kind of looks to me that A17 is designed to be more efficient in the higher frequency and power range (as paradoxically as it might sound), above 3Ghz. Contrast this with A15 which appears to focus the efficiency gains in the 2.8-3Ghz area. Looking at this, I am even more inclined to speculate that A17 cores have been designed for desktop use first and foremost as the gains over 5N products improve the higher we go. I wouldn't be surprised if the Macs will use the technology to reach very high performance while still staying below industry average per-core power consumption.

One can also see how A15 and M2 resp. A14 and M1 results line up, betraying the fact that they use the same CPU core.

And just for fun, here is a predicted power curve for A17 core up to 5Ghz. I used a polynomial of fourth degree to fit it, and as you can see, that fit is TIGHT. According to this, Coll should consume 10 watts as 4.5 Ghz (take it with a grain of salt, obviously)

I will update this thread and this post as more data is coming in. Please consider submitting the results for your machine, it would help a lot!

Update 27.09

Added a bunch of M2 and M1 devices as well as A16 and updated the discussion. Thanks for all the fine folks who contributed with the data!

P.S. I was not able to confirm Geekerwan's crazy 15W multi-core power consumption. In my tests it was between 7-10 watts. I assume that they run their tests using an active cooler (their benchmarks results are also higher than what people are normally getting). The A17 obviously can go higher, but that doesn't seem to be normal or intended operation.

P.P.S. The graphs shown here are my intellectual property and you do not have my permission to reproduce them for commercial purposes. If you want to cover this in your YouTube channel, blog, journal etc. contact me for permission.

 

[Xiao_Xi]

A17 is more efficient that either A14 and A15

How can you tell if one SoC is more efficient than another?

P.P.S. The graphs shown here are my intellectual property and you do not have my permission to reproduce them for commercial purposes. If you want to cover this in your YouTube channel, blog, journal etc. contact me for permission.

I would add a watermark to the graphic.

 

[leman]

How can you tell if one SoC is more efficient than another?

Lower power draw at the same frequency. Observe how the blue curve (A17) is below everything else.

 

[leman]

I would add a watermark to the graphic.

Too much hassle and the topic is niche enough.

 

[Analog Kid]

Fantastic chart! Thanks for the effort putting this together!

Any way you can replace the GHz axis with work performed? If there's a way to get computation per watt rather than GHz per watt, it would give a better estimate of true efficiency.

A14 and A15 appear to follow essentially the same curve aside from that weirdness at the top end, which isn't surprising at the same node. Interesting that the A series seems to scale frequency and the M1 seems to have discrete frequencies.

 

[Xiao_Xi]

Lower power draw at the same frequency.

The graph shows that A17 needs less voltage, not that it is more efficient. Consumption is proportional to Voltage^2 and Frequency.

 

[leman]

Any way you can replace the GHz axis with work performed? If there's a way to get computation per watt rather than GHz per watt, it would give a better estimate of true efficiency.

My testing is very simple and its only purpose is to load up the core. For the purpose of this workload the IPC is identical between the architectures (it makes one crips line a rise the board). In addition, comparing work is a bit tricky as threads get migrated between P and E cores on MC workloads. But I can do it if you want to see it.

Interesting that the A series seems to scale frequency and the M1 seems to have discrete frequencies.

Desktop has fewer thermal constraints, that all

 

[Analog Kid]

In addition, comparing work is a bit tricky as threads get migrated between P and E cores on MC workloads. But I can do it if you want to see it.

I'm grateful for the data you've shared already so I don't want to sign you up for more work.

There's already really interesting information here, but it's somewhat un-normalized for chip size and architecture. When I see the A14 and A15 coincident on the chart, though, I can't help but be curious if those two lines would be more distinct if we were measuring work and if the A17 would be yet further to the right of the others. Educated guesstimates can fill in some of that detail.

I guess I'm assuming the true efficiency improvement from generation to generation is more than the chart can show...

Desktop has fewer thermal constraints, that all

Agreed, but the distinct behaviors are so very clear in your chart. It's fun to see.

(it makes one crips line a rise the board)

I tried, but I can't untangle this... 😄

 

[Boil]

P.P.S. The graphs shown here are my intellectual property and you do not have my permission to reproduce them for commercial purposes. If you want to cover this in your YouTube channel, blog, journal etc. contact me for permission.

Inbound MaxTech video...!

Desktop has fewer thermal constraints, that all

I am trying to imagine the honking large cooler that would be needed for the Mn Extreme as depicted in that recent patent you linked...?

Seems a prime candidate for a Mac Cube chassis...! ;^p

 

[quarkysg]

The graph shows that A17 needs less voltage, not that it is more efficient. Consumption is proportional to Voltage^2 and Frequency.

So @leman's graphs implies that A17 needs lower voltage to drive the CPU at the same frequency compared to the A16, which then implies better efficiency?

 

[TechnoMonk]

When you get to M3 Max and Ultra, the difference could be huge at higher frequencies.

 

[Tyler O'Bannon]

Fantastic. Well done. This is great info and insight into the A17 Pro and the direction Apple may push all silicon. Very interesting info.

 

[jdb8167]

notably, I am missing results for M2 and A16

As I posted on a different site, if no one gets you the M2 power curve numbers before the weekend, I’ll run the test program on my M2 MacBook Air when I have time this weekend.

 

[Shirasaki]

So, Apple is quietly switching silicon design focus from mobile to desktop now? That’s a very interesting trend indeed, and as long as they can maintain their nominal battery life on iPhone, better performance is always welcome.

 

[bnumerick]

Just posted some results from a A16 Bionic, 14" M2 Max, and M2 Ultra.

 

[leman]

The graph shows that A17 needs less voltage, not that it is more efficient. Consumption is proportional to Voltage^2 and Frequency.

I’m not sure how you understand efficiency, but in my book if something can faster at lower power it’s more efficient.

Sure, efficiency is about work, not frequency, but IPC is comparable across architectures so frequency is decent enough proxy.

 

[eilavid]

Excellent chart.

 

[EntropyQ3]

The graph shows that A17 needs less voltage, not that it is more efficient. Consumption is proportional to Voltage^2 and Frequency.

This is true as a first order approximation, but it also means that the two are directly connected and by measuring one you effectively measure the other.

One of the things that does complicate the relationship is that increasing temperatures will contribute to a drop in efficiency, i.e. the A17 Pro could conceivably show a bit better behaviour in a thermally less demanding environment, flattening the upper end of the curve somewhat.

Thanks to leman for doing this! It’s so much better to have numbers from a method you know and understand, and where limitations have a chance of being assessed.

 

[v0lume4]

Very interesting findings. Great work!!

 

[leman]

One of the things that does complicate the relationship is that increasing temperatures will contribute to a drop in efficiency, i.e. the A17 Pro could conceivably show a bit better behaviour in a thermally less demanding environment, flattening the upper end of the curve somewhat.

I think the fact that phones are thermally constrained is a blessing for this kind of experiment, since we get a nice range of frequencies and power consumption samples. As the phone throttles more, the frequency decreases and we can get a glimpse into the actual curve. Note how M-series results are tight clusters of dots instead of lines, since they don't throttle (different groups correspond to high/low power as well as difference in clock between single and multi-core runs). The later also gives us some idea about the accuracy of the measurements and variance of results.

It’s so much better to have numbers from a method you know and understand, and where limitations have a chance of being assessed.

If there is one thing I am teaching my students is that one should look at distributions and not point estimates. If one only reports averages, there are a lot of interesting things one is missing. It always pays off to collect as detailed data as possible and aggregate later.

For example, I noticed when doing multicore benchmarks is that Apple moves every thread between efficiency and performance cores. I will need to investigate it further, but so far what I'm seeing is that around 20% of the time for every thread is spent running on the E-core!

 

[leman]

Updated the charts with M2 and A16, thanks to all the fine people who contributed with the data!

 

[leman]

Any way you can replace the GHz axis with work performed? If there's a way to get computation per watt rather than GHz per watt, it would give a better estimate of true efficiency.

Here you go! Please treat this data with a good degree of scepticism — my test is very simplistic and does not represent any useful work (it's just a nested loop with an int divide and some branches), it's just for loading up the CPU core. Also, the system appears to migrate the threads between P and E cores when running multiple threads, introducing additional variation to the mix. Since I have no way to distinguish work done on P or E core, I add them together for each thread.

But I think that the basic result is the same as in the power/frequency graph. A17 is more efficient at higher power usage than the rest, except for A16 (which is as efficient as A17).

And here a little graph that shows how all these u-archs behave identically for my workload. Fun fact: I did this for my old A13 as well, and it has half the throughput per Ghz. I guess A14 introduced an additional integer divide unit and OOO cores essentially unroll the loop.

 

[boak]

So no efficiency gains from A16 to A17, performance gains came directly from frequency increase, node process gains went to GPU and other modules?

 

[leman]

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

Take this with a grain of salt obviously, but given how good the fit is, I'd say we have even more evidence that A17 is designed with desktop use in mind. I think 4.5 Ghz at 10 watts should at least be achievable, and that would give desktop M3 GB6 single of > 3600

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, or that the curve hits 4.5 Ghx exacly at 10 watts and 5Ghx at approx 15 watts? These are all important psychological points.

 

[leman]

So no efficiency gains from A16 to A17, performance gains came directly from frequency increase, node process gains went to GPU and other modules?

No efficiency gains from A16 to A17... but I think it's much more interesting to compare A17 to M2. A16 was a one-time design aimed at optimising performance and efficiency in a mobile phone, and Apple tweaked the structure sizes and used an optimised N4 node to get there. I am getting more and more convinced that A17 P-cores instead are developed for the desktop and essentially continue where A16 stopped, but with a wider frequency range in mind.

There was some preliminary die analysis showing that the A17 cores shrunk in size compared to previous designs, so cost factor could also be in play.