What microarchitecture choices have made M2 less efficient than M1?

[Xiao_Xi]

M2 seems to be less efficient than M1 although M2 uses a better TSMC node than the M1. What microarchitecture choices have made M2 less efficient than M1?

Notebookcheck wrote in its M2 MBA review:

We ran Cinebench R23 Single & Multi on both machines and logged the package power, which allows us to compare the efficiency. The numbers clearly show that the performance increase is based on higher consumption figures (+42 % @single-core, ~41 %@multi-core), which means the efficiency is significantly worse compared to the M1 SoC. However, even the M2 is still clearly ahead of Intel and AMD in terms of single-core efficiency. The multi-core advantage over Intel’s Alder Lake chips is also still impressive, but AMD is not that far off with the new Ryzen 7 6800U.

Apple MacBook Air M2 Entry Review – A very good, but too expensive daily MacBook

Notebookcheck reviews the new Apple MacBook Air M2 Midnight in the base-level spec with 8 GB RAM and 256 GB SSD storage.

www.notebookcheck.net

 

[leman]

From my entirely naive perspective: M2 is clocked higher but the process is not that different, which puts M2 at a less advantageous point in the efficiency curve. Still, looking at the notebookcheck results the M2 is still more efficient than M1 when running at the same package power.

In other terms, efficiency is not linear, but is a function of frequency. It would be very interesting to see the power/frequency curve of these systems.

 

[Xiao_Xi]

Has anyone tried reverse engineering the M2 CPU/GPU like people did with the M1?

Icestorm Overview

Firestorm Overview

Apple M1 GPU ISA

M1 Explainer 070.pdf

drive.google.com

 

[Zest28]

Because of higher clock speeds probably.

 

[mi7chy]

Apple attempted to make it more performant especially on the iGPU side to catch up with the competition but lack of active cooling negated any benefit. If you're coming from M1 it's better to wait for 3nm M3 with better designed MBA.

 

[diamond.g]

Apple attempted to make it more performant especially on the iGPU side to catch up with the competition but lack of active cooling negated any benefit. If you're coming from M1 it's better to wait for 3nm M3 with better designed MBA.

Nah active cooling, or lack there of isn't the culprit. Notice how they helpfully included the M2 Macbook Pro and it's score isn't that much better than the M2 Air.

 

[MrGunny94]

From my entirely naive perspective: M2 is clocked higher but the process is not that different, which puts M2 at a less advantageous point in the efficiency curve. Still, looking at the notebookcheck results the M2 is still more efficient than M1 when running at the same package power.

In other terms, efficiency is not linear, but is a function of frequency. It would be very interesting to see the power/frequency curve of these systems.

Yep... nail in the head.

Can't wait to see a 3nm M3..

 

[leman]

Apple attempted to make it more performant especially on the iGPU side to catch up with the competition

What are you even talking about. Even M1 is significantly better than any x86 iGPU. How is it Apple trying to catch up?

 

[Xiao_Xi]

M2 is clocked higher but the process is not that different, which puts M2 at a less advantageous point in the efficiency curve.

How is it possible that an increase of 0.3 GHz results in a 25% efficiency loss? Could part of the efficiency loss be due to LPDDR5? Does LPDDR5 consume more than LPDDR4? How much?

Apple MacBook Air M2 Entry Review – A very good, but too expensive daily MacBook

Notebookcheck reviews the new Apple MacBook Air M2 Midnight in the base-level spec with 8 GB RAM and 256 GB SSD storage.

www.notebookcheck.net

Apple MacBook Air 2020 M1 Entry Review: Apple M1 CPU humbles Intel and AMD

The M1 based entry-level Macbook in our extensive technical test.

www.notebookcheck.net

 

[Gerdi]

How is it possible that an increase of 0.3 GHz results in a 25% efficiency loss? Could part of the efficiency loss be due to LPDDR5? Does LPDDR5 consume more than LPDDR4? How much?

View attachment 2034775

Apple MacBook Air M2 Entry Review – A very good, but too expensive daily MacBook

Notebookcheck reviews the new Apple MacBook Air M2 Midnight in the base-level spec with 8 GB RAM and 256 GB SSD storage.

www.notebookcheck.net

View attachment 2034777

Apple MacBook Air 2020 M1 Entry Review: Apple M1 CPU humbles Intel and AMD

The M1 based entry-level Macbook in our extensive technical test.

www.notebookcheck.net

Look, the efficiency loss is mostly due to voltage Pdyn = Cdyn * V^2 * f. But it does not really matter, because when the M2 is throttling down to sustainable power, Voltage will be much reduced and efficiency is back up again.

In any case, the architectural constant, which impacts power is Cdyn. You however conclude, that because Pdyn is going up, it is because of Cdyn going up - which is probably not the case.

 

[Gerdi]

Apple attempted to make it more performant especially on the iGPU side to catch up with the competition but lack of active cooling negated any benefit. If you're coming from M1 it's better to wait for 3nm M3 with better designed MBA.

You make no sense, there is no competition Apple needs to catch up to.

Even worse, you mixing up 2 issues, which are inversely related. The efficiency is lower because of high voltage and power _despite_ passive cooling.

 

[diamond.g]

You make no sense, there is no competition Apple needs to catch up to.

For an iGPU this is true (I am assuming the 680M isn't as performant). For a GPU, eh, it does get outperformed (at the cost of all the heat/power needed) by others.

 

[Gerdi]

For an iGPU this is true (I am assuming the 680M isn't as performant). For a GPU, eh, it does get outperformed (at the cost of all the heat/power needed) by others.

Assuming you mean CPU in your second sentence - it does not make much sense either. In order to outperform the competition you just design a SoC in the according power range. A 9W base SKU is not intended to outperform the competition, which is using 3x the power. For this Apple has the Pro/Max SoCs as the next higher power range.
However it is supposed to outperform any competition in any passively cooled device - which it easily does.

With other words as long as your efficiency is highest, you will outperform the competition at any similar power level.

 

[diamond.g]

Assuming you mean CPU in your second sentence - it does not make much sense either. In order to outperform the competition you just design a SoC in the according power range. A 9W base SKU is not intended to outperform the competition, which is using 3x the power. For this Apple has the Pro/Max SoCs as the next higher power range.
However it is supposed to outperform any competition in any passively cooled device - which it easily does.

With other words as long as your efficiency is highest, you will outperform the competition at any similar power level.

Oh I meant GPU. Sadly in the PC world "big" iGPU's aren't popular. The closest would be the Series X APU, but that isn't an off the shelf option for OEMs. So they settle for CPU/dGPU options for more graphics performance. Maybe Apple will get the industry to change on that front, but I kind of doubt it (mostly because nvidia doesn't have a CPU to integrate a GPU with). Intel has just recently started making GPU's that are not 100% hot garbage (they are only 60%, lol) and they don't plan on back porting those changes to iGPU's until something like 2024.

 

[Xiao_Xi]

But it does not really matter, because when the M2 is throttling down to sustainable power, Voltage will be much reduced and efficiency is back up again.

If the benchmark causes M1 and M2 to lose efficiency by pushing them into the burst zone, wouldn't M1 exhibit the same behavior as M2?

By the way, if MBA can dissipate 10W, neither M1 nor M2 is thermally throttled at R23 single, only M2 at R23 multi as results show.

For an iGPU this is true (I am assuming the 680M isn't as performant).

M2 GPU is more performant than 680M.

 

[mi7chy]

Notice how they helpfully included the M2 Macbook Pro and it's score isn't that much better than the M2 Air.

You're confusing performance and efficiency. Aside from throttling, M2 is more performant than M1 but M1 is more efficient than M2. The fan version of M2/M1 is more performant than fanless.

Cinebench R23

Furthermore, 6800U is more performant than M2 and M1 but they didn't test the 6800U at ~15W where it shines. They also conveniently left out the power consumption metrics for Blender so if it's similar to 28W and 21.2W respectively then the 6800U has more performance per watt even outside of its sweet spot power zone.

Blender

6800U sweet spot at 15W where it delivers ~91% of the performance at half the power draw of 30W.

 

[mi7chy]

M2 GPU is more performant than 680M.
View attachment 2034915

That's useless synthetic benchmark. Show benchmarks for God of War, Cyberpunk 2077, Doom Eternal, etc. Disable fake MSAA 8x anti-aliasing on MacOS though. There's a reason why AMD APU is king of consoles and the new portable PC console segment like Steam Deck.

Fake MSAA 8x on MacOS

Real MSAA 8x

 

[NT1440]

That's useless synthetic benchmark. Show benchmarks for God of War, Cyberpunk 2077, Doom Eternal, etc. Disable fake MSAA 8x anti-aliasing on MacOS though. There's a reason why AMD APU is king of consoles and the new portable PC console segment like Steam Deck.

Fake MSAA 8x on MacOS

Real MSAA 8x

Performance only counts in games, eh?

 

[diamond.g]

Performance only counts in games, eh?

The interesting thing (I swear this isn't a tangent) is the 680M and M2 have the same TFLOP rating, so for compute they should be similar. The M2 has a bandwidth advantage though, so I assume that helps in some workloads.

 

[1096bimu]

M2 seems to be less efficient than M1 although M2 uses a better TSMC node than the M1. What microarchitecture choices have made M2 less efficient than M1?

Notebookcheck wrote in its M2 MBA review:

Apple MacBook Air M2 Entry Review – A very good, but too expensive daily MacBook

Notebookcheck reviews the new Apple MacBook Air M2 Midnight in the base-level spec with 8 GB RAM and 256 GB SSD storage.

www.notebookcheck.net

It isn't less efficient, you're not interpreting this correctly.
Efficiency drops exponentially as you go higher in power and therefore performance. But you measure efficiency at same power/performance, not peak power/performance.
It is therefore perfectly possible for A to be more efficient than B at peak performance, but B more efficient than A at same performance.

For example Car A uses 5L/100km at 100km/h, Car B uses 6L/100km at 100km/h, but car A has top speed 250km/h car B has top sped 200km/h. You drive both cars to top speed and conclude B is more efficient.

 

[leman]

There's a reason why AMD APU is king of consoles and the new portable PC console segment like Steam Deck.

Yeah, it’s cheap. That’s pretty much the main reason.

 

[diamond.g]

Yeah, it’s cheap. That’s pretty much the main reason.

The development toolset is also easier than what was available during the PPC days.

 

[Gerdi]

If the benchmark causes M1 and M2 to lose efficiency by pushing them into the burst zone, wouldn't M1 exhibit the same behavior as M2?

Not sure if I understand the question. M2 happens to run with higher frequency and voltage when not thermally restricted - hence it will be less efficient - but it has not much to do with architectural efficiency (remember the formula in my last post).

By the way, if MBA can dissipate 10W, neither M1 nor M2 is thermally throttled at R23 single, only M2 at R23 multi as results show.

Yet, even under single core, the M2 uses higher voltage and frequency and is thus less efficient in this scenario.

What I was saying is, once the M2 is throttled down to the same power level as the M1 - and the M2 is still faster than the M1, then it must be also more power efficient.

 

[NT1440]

Not sure if I understand the question. M2 happens to run with higher frequency and voltage - hence it will be less efficient - but it has not much to do with architectural efficiency (remember the formula in my last post).


Yet, even under single core, the M2 uses higher voltage and frequency and is thus less efficient in this scenario.

What I was saying is, once the M2 is throttled down to the same power level as the M1 - and the M2 is still faster than the M1, then it must be also more power efficient.

But if it gets the given job faster and returns to a lower power state faster, is that less efficient?

 

[Gerdi]

But if it gets the given job faster and returns to a lower power state faster, is that less efficient?

It does not really matter, if you interpret the formula I did post above correctly. Given 2 identical cores, one runs with higher frequency and voltage, it will never get the efficiency back by finishing faster. Because you only gain time back, which is linear to frequency, while the power shows a cubic increase. Pdyn = Cdyn * V^2 * f
With other words, you always loose efficiency if you increase voltage - independent how you calculate.

Another conclusion from this formula is, that power is always increasing faster than frequency (and performance does not increase faster than frequency) - hence the lower the voltage the higher the efficiency.