[CPU only] Apple M1/2(Max/Ultra) (TSMC 5nm) vs AMD Zen 4 (TSMC 5nm) - Technical Analysis

[theorist9]

I've not been able to find enough data to compare AS's efficiency to Z4's. I assume someone will do a direct comparison in the near future. For now, the only datapoint I have is CB R23, which shouldn't be used for AS. Having said that, Anandtech measured the M1 Max's package power when running CBR23 MT, and got 34W with a score of 12,375. On Eco Mode, according to ArsTechnica, the package power of the 7950X seems to be 90W, so if we use that and Anadtech's Eco mode CB R23 measurement of 31,308, we have:

M1 Max = 12,375 pts/34W = 364 pts/W
7950X Eco Mode = 31,308 points/90W = 348 pts/W

So on a multithreaded CPU benchmark that is suboptimized for AS, the M1 Max is 5% more efficient than the 7950X in Eco Mode. This suggests that AMD is catching up with AS in efficiency, at least for MT tasks; if true, this would be an impressive achievement for AMD.

Note also that, as @leman pointed out, the M1's package power includes RAM, while the 7950X's doesn't.

As far as process nodes go, both the M1 and Z4 are using 5 nm. But it's possible (indeed, it seems likely) Z4 is using a newer 5 nm variant.

Sources

Apple's M1 Pro, M1 Max SoCs Investigated: New Performance and Efficiency Heights

www.anandtech.com

AMD Zen 4 Ryzen 9 7950X and Ryzen 5 7600X Review: Retaking The High-End

www.anandtech.com

Ryzen 7600X and 7950X review: Zen 4 starts off expensive but impressive

Thermal behavior is a bit odd, but there’s lots to like about these new CPUs.

arstechnica.com

 

[leman]

Having said that, Anandtech measured the M1 Max's package power when running CBR23 MT, and got 34W with a score of 12,375.

Don’t forget also that M1 package power includes RAM power usage as well, unlike x86 CPUs.

 

[theorist9]

Don’t forget also that M1 package power includes RAM power usage as well, unlike x86 CPUs.

Thanks, I'll add that caveat to my post. Any way to estimate what the RAM power usage would be for the 7950X when running CB R23 in MT mode?

 

[leman]

Thanks, I'll add that caveat to my post. Any way to estimate what the RAM power usage would be for the 7950X when running CB R23 in MT mode?

No idea, but wouldn’t it be better to just subtract the RAM from the M1 figure? At least we have precise measurements there.

 

[theorist9]

No idea, but wouldn’t it be better to just subtract the RAM from the M1 figure? At least we have precise measurements there.

Sure, but what are they? I didn't see any in the Anandtech article from which I got the package power measurements. Note that you'd need CB R23-specific RAM power consumption values for the M1 Max (or some way to reasonably estimate them), since RAM power consumption depends on how RAM-heavy the task is.

Also curious how they're able to tease out the RAM power consumption, given that it's integrated with the package.

 

[leman]

Sure, but what are they? I didn't see any in the Anandtech article from which I got the package power measurements. Also curious how they're able to tease out the RAM power consumption, given that it's integrated with the package.

I just tried to run a test myself but it seems that Apple has changed the powermetrics tool in some of the recent updates (I am on latest Ventura beta). It now only reports power consumption of CPU clusters, GPU and the ANE. It used to also report DRAM power but that counter seems to be gone. I also don't see any counter for total SoC power but I don't think it was every there (I assume package power as used by benchmarks was simply CPU+GPU+ANE+DRAM).

FWIW, I get an average of 29W on CPU clusters running CB23, with the resulting score of 12133. This is identical to the total SoC power reported by the tool as the DRAM is not included anymore. I have no idea whether this includes the other parts of the SoC like the cache and other stuff. Anyway, this review seems to imply that the CB23 CPU power usage for 8/16 Zen4 running at 65W mode is 56Watts (the CPU power in the screenshot). So your math becomes (I've taken the Zen4 values from the article I quoted):

M1 Max = 12133 pts/29W = 418.4 pts/W
7950X Eco Mode = 29500 points/56W = 526.8 pts/W

Which makes Zen4 25% more efficient in this test (looking at the CPU values only).

It's a shame we don't have a way to measure the power consumption of the system cache and the chip fabric.

 

[theorist9]

I just tried to run a test myself but it seems that Apple has changed the powermetrics tool in some of the recent updates (I am on latest Ventura beta). It now only reports power consumption of CPU clusters, GPU and the ANE. It used to also report DRAM power but that counter seems to be gone. I also don't see any counter for total SoC power but I don't think it was every there (I assume package power as used by benchmarks was simply CPU+GPU+ANE+DRAM).

FWIW, I get an average of 29W on CPU clusters running CB23, with the resulting score of 12133. This is identical to the total SoC power reported by the tool as the DRAM is not included anymore. I have no idea whether this includes the other parts of the SoC like the cache and other stuff. Anyway, this review seems to imply that the CB23 CPU power usage for 8/16 Zen4 running at 65W mode is 56Watts (the CPU power in the screenshot). So your math becomes (I've taken the Zen4 values from the article I quoted):

M1 Max = 12133 pts/29W = 418.4 pts/W
7950X Eco Mode = 29500 points/56W = 526.8 pts/W

Which makes Zen4 25% more efficient in this test (looking at the CPU values only).

It's a shame we don't have a way to measure the power consumption of the system cache and the chip fabric.

Where are you seeing the 56W? Everything I'm seeing for 65W Eco mode socket power/package power is 88W (about the same as the 90W figure reported by Ars Technica):

"65W TDP – which is 88W package power – had the chip running at just under 3.9GHz average."

 

[leman]

Where are you seeing the 56W? Everything I'm seeing for 65W Eco mode socket power/package power is 88W:

"65W TDP – which is 88W package power – had the chip running at just under 3.9GHz average."
View attachment 2086419

Look at the second screenshot at the page (one showing detailed stats of running the 65W mode)

 

[leman]

P.S. One thing that really struck me is the very low average instructions per cycle executed of CB23. The tool reports average of 2.7 instructions per cycle (which I assume is over active CPU cores). Running JetStream2 gives between 4 and 5.5 instructions per cycle depending on the test. This again goes to show that running Cinebench the Apple CPU is suffering from some very bad backend utilisation.

 

[theorist9]

Look at the second screenshot at the page (one showing detailed stats of running the 65W mode)

The 2nd screenshot isn't Eco Mode (even though the article labels it as such). It shows 142W PPT. Please don't make me hunt for it -- just post a screenshot showing what you're seeing.

 

[leman]

The 2nd screenshot isn't Eco Mode. Don't make me hunt for it -- just post a screenshot showing what you're seeing. View attachment 2086446

Sorry, I just didn’t want to link to the screenshot directly, it’s part of the article after all. Anyway, it’s this image

 

[theorist9]

Sorry, I just didn’t want to link to the screenshot directly, it’s part of the article after all. Anyway, it’s this image

Getting a bit confused. It now seems that, contrary to what you said before, the 90W package power Ars Technica measured for the 7950X *does* include the RAM (plus iGPU) (you also mentioned "ANE"; what is that?). I.e., it's:

M1 Max:
package power (Anandtech): 34 W; CPU power (Leman): 29W
7950X:
package power (kitguru/Ars Technica): 88W/90W; CPU power (kitguru): 56W.

Also, I think it's perfectly OK to link an image in an article, so long as you give full attribtion to the author. So let's just say this image is from an article by Luke Hill (kitguru),

AMD Ryzen 9 7950X & Ryzen 7 7700X ‘Zen 4’ Review - KitGuru

AMD launched its Ryzen 5000 – Zen 3 – desktop processors back in late 2020. Coupled with the venerab

www.kitguru.net

 

[leman]

Getting a bit confused. It now seems that, contrary to what you said before, the 90W package power Ars Technica measured for the 7950X *does* include the RAM. I.e., it's:

M1 Max:
package power (Anandtech): 34 W; CPU power (Leman): 29W
7950X:
package power (kitguru): 88W; CPU power (Kitguru): 56W.

RAM is not part of the Zen4 package so it is not powered by the socket and thus there is no way it’s included in the measurement. Package power is not just the Zen CPU itself, it’s also the IO die with the cache and memory controllers, the chip bus, the PCI bus, likely some power supporting circuitry, various auxiliary units and who knows what else. The way I understand this is that all this supporting stuff draws almost 30W just to let the CPU do its job. And we do know that IO die on Zen4 is particularly power-hungry (I think I might have seen figures like 20 watts quoted somewhere?)

Frankly I start thinking that a poster (forgot the name unfortunately) who wrote the other day that all this power reporting is be meaningless might be right. Even if we assume that the counters are accurate (and it’s far from obvious that they are!), they only give us the part of the story. Like, powermetrics does not expose the socket power draw and now it removed the DRAM, meaning we have no way knowing how much power the chip pulls. We can compare the CPU only, but then we are ignoring the cache/interconnect power usage which are all important contributors to power usage. I mean, if Zen3 cache+interconnect needs 30W and Apples only needs 5, that’s also part of the overall efficiency, right? I’m a bit at a loss here. It does seem like the only prudent course of action is to treat these numbers as very imprecise indicators and trends and just say “no idea” if the difference is less than some generous constant factor.

 

[theorist9]

RAM is not part of the Zen4 package so it is not powered by the socket and thus there is no way it’s included in the measurement. Package power is not just the Zen CPU itself, it’s also the IO die with the cache and memory controllers, the chip bus, the PCI bus, likely some power supporting circuitry, various auxiliary units and who knows what else. The way I understand this is that all this supporting stuff draws almost 30W just to let the CPU do its job. And we do know that IO die on Zen4 is particularly power-hungry (I think I might have seen figures like 20 watts quoted somewhere?)

Frankly I start thinking that a poster (forgot the name unfortunately) who wrote the other day that all this power reporting is be meaningless might be right. Even if we assume that the counters are accurate (and it’s far from obvious that they are!), they only give us the part of the story. Like, powermetrics does not expose the socket power draw and now it removed the DRAM, meaning we have no way knowing how much power the chip pulls. We can compare the CPU only, but then we are ignoring the cache/interconnect power usage which are all important contributors to power usage. I mean, if Zen3 cache+interconnect needs 30W and Apples only needs 5, that’s also part of the overall efficiency, right? I’m a bit at a loss here. It does seem like the only prudent course of action is to treat these numbers as very imprecise indicators and trends and just say “no idea” if the difference is less than some generous constant factor.

Maybe someone will do a wall power comparison of the Studio and a 7950X build running Eco Mode, with the same size RAM and SSD, and driving the same monitor (to equalize the iGPU demand). You could then subtract their respective fan power consumptions, in case they're significantly different.

 

[leman]

Maybe someone will do a wall power comparison of the Studio and a 7950X build running Eco Mode, with the size RAM and SSD, and driving the same monitor (to equalize the iGPU demand). You could then subtract their respective fan power consumptions, in case they're significantly different.

Yeah I was thinking the same. You don’t even need to make the systems too similar, just looking at the difference between idle and load power should be indicative.

 

[name99]

Whats the alternative ?

Geekbench uses NPU for machine learning tests, which makes it not pure CPU benchmark at all.

Till then Cinebench R23 is the only real cross-platform CPU-only test.

(a) The alternative is Geekbench (adequate) or SPEC
Or admit your ignorance.
A bad benchmark (and the undeserved confidence that goes with it) is worse than no benchmark.

On the plus side, on the internet it can sometimes be hard to see who is clueless. Loud talk about Cinebench helps solve that problem...

(b) WTF are you talking about?

 

[theorist9]

Whats the alternative ?

Geekbench uses NPU for machine learning tests, which makes it not pure CPU benchmark at all.

Till then Cinebench R23 is the only real cross-platform CPU-only test.

ML performance contributes only about 3% to the GB overall score (it's one of 11 individual tests making up the FP subscore, and that in turn contributes 30% to the overall score). [Each GB subscore is calcuated from the geo mean of the individual scores; then the overall score is calculated from the weighted arithmetic mean of the subscores, where it's 5% crytpo, 30% FP, and 65% integer. Hence approximately 3% for ML.]

 

[leman]

Geekbench uses NPU for machine learning tests, which makes it not pure CPU benchmark at all.

Where did you get that idea from?

 

[Gerdi]

I just tried to run a test myself but it seems that Apple has changed the powermetrics tool in some of the recent updates (I am on latest Ventura beta). It now only reports power consumption of CPU clusters, GPU and the ANE. It used to also report DRAM power but that counter seems to be gone. I also don't see any counter for total SoC power but I don't think it was every there (I assume package power as used by benchmarks was simply CPU+GPU+ANE+DRAM).

FWIW, I get an average of 29W on CPU clusters running CB23, with the resulting score of 12133. This is identical to the total SoC power reported by the tool as the DRAM is not included anymore. I have no idea whether this includes the other parts of the SoC like the cache and other stuff. Anyway, this review seems to imply that the CB23 CPU power usage for 8/16 Zen4 running at 65W mode is 56Watts (the CPU power in the screenshot). So your math becomes (I've taken the Zen4 values from the article I quoted):

M1 Max = 12133 pts/29W = 418.4 pts/W
7950X Eco Mode = 29500 points/56W = 526.8 pts/W

Which makes Zen4 25% more efficient in this test (looking at the CPU values only).

It's a shame we don't have a way to measure the power consumption of the system cache and the chip fabric.

Why are you even using CPU power instead of socket power? Do you even know what is part of this CPU power measurements - like which caches are taken into consideration?
At least for socket power we do know, that it contains everything what the CPU needs to run plus perhaps DRAM in case of M1.
Even the Kitguru article you are referring to is calculating performance/package power for their efficiency rating.

 

[leman]

Why are you even using CPU power instead of socket power? Do you even know what is part of this CPU power measurements - like which caches are taken into consideration?
At least for socket power we do know, that it contains everything what the CPU needs to run plus perhaps DRAM in case of M1.
Even the Kitguru article you are referring to is calculating performance/package power for their efficiency rating.

Exactly my point. There is no way to measure socket power because Apple doesn’t provide this info. Probably never did.

 

[kvic]

I don't trust that Linux propaganda website. AMD uses Geekbench for Single Core in their benchmarks and so does Intel. So I will also use Geekbench to compare CPU pref.

You can complain about his lack of spell check, skill in sentence composition. Phoronix is a propaganda site to Linux as MR a propaganda site to Apple. Phoronix owner a propagandist for Linux as some active members here propagandists for Apple. At least that's my view if you ask me.

 

[leman]

You can complain about his lack of spell check, skill in sentence composition. Phoronix is a propaganda site to Linux as MR a propaganda site to Apple. Phoronix owner a propagandist for Linux as some active members here propagandists for Apple. At least that's my view if you ask me.

Photonix maintains a collection on obscure, irrelevant benchmarks many of which are not optimized for macOS. It’s a useful website of course if you use the information critically, but the way how some posters quote it is just cherry picking results to make Apple look bad.

 

[kvic]

Photonix maintains a collection on obscure, irrelevant benchmarks many of which are not optimized for macOS. It’s a useful website of course if you use the information critically, but the way how some posters quote it is just cherry picking results to make Apple look bad.

Phoronix has nothing to do with Apple and its ecosystem as far as I'm aware. Benchmark about Apple products rarely show up there. Phoronix is a Linux enthusiastic site, focussing on Linux, open source software. The site owner happens to be behind a benchmark tool (which is a collection of other benchmarks) and hence he uses it a lot for comparison between different Linux kernels, sometimes between Linux & Windows, rarely care about Apple.

EDIT: his benchmark tool is not "obscure" since it's open source. You can check it out if you care. But as I said Phoronix seems not interested in Apple at all.

 

[altaic]

Phoronix has nothing to do with Apple and its ecosystem as far as I'm aware. Benchmark about Apple products rarely show up there. Phoronix is a Linux enthusiastic site, focussing on Linux, open source software. The site owner happens to be behind a benchmark tool (which is a collection of other benchmarks) and hence he uses it a lot for comparison between different Linux kernels, sometimes between Linux & Windows, rarely care about Apple.

I think that was his point.

 

[kvic]

I think that was his point.

My point should be clear in my initial response to exoticSpice. If not, I could elaborate.