Sidebar Sidebar
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

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

E

exoticSpice

Suspended
Original poster
Jan 9, 2022
1,242
1,953
With the advent of Zen 4 we can finally have a comparison between these two architectures as they are both based on TSMC 5nm.

We can see how they compare in perf/w and clock freq and also the power consumption per core and under load.

Zen 4 analysis is limited to desktop and laptop CPUs. Since Zen 4 is not launched yet I am keep on updating this post

M1 Ultra 16P and 4E cores:
Cinebench R23 - Single Core: 1536 , Multi Core: 24190
Geekbench 5 - Single Core: 1795 , Multi Core: 24038

M2 4P and 4E cores:
Cinebench R23 - Single Core: 1585 , Multi Core: 8745
Geekbench 5 - Single Core: 1937 , Multi Core: 8991

Ryzen 9 5950X(TSMC 7nm) 16 cores:
Cinebench R23 - Single Core: 1684, Multi Core: 24835
Geekbench 5 - Single Core: 1725, Multi Core: 17106



I added the last gen AMD desktop flagship to also compare how AMD improved from Zen 3 to Zen 4.
The best comparsion to M2 will be the 7800U and for the M1/M2 Ultra it would the Ryzen 9 7950X.

Zen 4 line up:

7950X:

7900X:

7800X:

7600X:
 
senttoschool said:
Cinebench should not be used to gauge general CPU performance for Apple Silicon (and for any CPU really):
https://www.reddit.com/r/hardware/comments/pitid6
Click to expand...
Look Cinebench is also unoptimised for AMD cpus. We use every CPU benchmark there is. Feel free to suggest your own. The more points of data the better.

I did also add Geekbench. Later I will add Blender CPU tests, handbrake and also linux kernel complie times.
and gaming later on.
 
  • Haha
Reactions: AlexMac89
Kazgarth said:
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.
Click to expand...
Real world testing is the alternative. Opening apps, measuring efficiencies over each generation. For example, encoding a video, doing something on the CPU akin to what a user would do.

Synthetics just gauge a specific overview and should never be used for black and white CPU to CPU comparison.
 
  • Like
Reactions: Miicat_47 and AlexMac89
Kazgarth said:
Geekbench uses NPU for machine learning tests, which makes it not pure CPU benchmark at all.
Click to expand...

Are you sure about that? From what I understand they use a tensorflow model (MobileNet) running on the CPU.

jav6454 said:
Real world testing is the alternative. Opening apps, measuring efficiencies over each generation. For example, encoding a video, doing something on the CPU akin to what a user would do.
Click to expand...

How are you going to do it in a way that allows systematic, consistent performance comparison? And video encoding is just one specific domain that is only relevant for very few users and hardly a predictor of overall CPU performance in an age of dedicated hardware video encoders.

jav6454 said:
Synthetics just gauge a specific overview and should never be used for black and white CPU to CPU comparison.
Click to expand...

Synthetic benchmarks have been out of fashion for many years — for reasons you mention. Modern benchmarks like Geekbench or SPEC are not synthetic benchmarks: they are derived from real-world workloads that aim to capture different kind of tasks a user might run.

exoticSpice said:
It's more optimised for Intel CPUs. Cinebench uses a lot of Intel frameworks
Click to expand...

Cinebench works just fine on AMD. It's just not a general-purpose benchmark.
 
  • Like
Reactions: Tagbert
jav6454 said:
Real world testing is the alternative. Opening apps, measuring efficiencies over each generation. For example, encoding a video, doing something on the CPU akin to what a user would do.

Synthetics just gauge a specific overview and should never be used for black and white CPU to CPU comparison.
Click to expand...
Black and white CPU comparisons are a fools game anyway.
The problem with application tests is that the results aren’t transferable. Even if you look at performance in a single app, lets say Photoshop, you will find that relative performance varies depending on what specific tasks you use it for. Not only that, but it will also vary with for instance the size of the problem set, et cetera.

We simply cannot measure ”general performance” with a degree of precision that is adequate for the small increments in performance that todays lithographic and architectural engineers struggle hard to eke out. It can’t be done.
 
  • Like
  • Haha
Reactions: harold.ji, throAU, AlexMac89 and 1 other person
What about PassMark's CPU Mark? Is it problematic to use this in comparing AS with AMD/Intel x86?

1660553951938.png

www.passmark.com

PassMark PerformanceTest - macOS System Benchmark Software

Benchmark the speed of your macOS devices, then compare the result to other devices. Includes disk, Memory and CPU tests.
www.passmark.com www.passmark.com
 
Last edited:
Last edited:
  • Wow
Reactions: exoticSpice
exoticSpice said:
Look Cinebench is also unoptimised for AMD cpus.
Click to expand...
This is just simply not true. In fact, AMD brought Cinebench into the mainstream as a benchmark during Zen2's product announcement. Cinebench is disproportionately kind to the Zen architecture.
 
Last edited:
exoticSpice said:
It's more optimised for Intel CPUs. Cinebench uses a lot of Intel frameworks
Click to expand...

Intel frameworks like Embree are not a problem per se, when you compare x86 architectures - as both AMD and Intel support SSE4.2/AVX2. However on ARM architectures the AVX instructions are simply mapped to NEON - which is highly inefficient.
 
EntropyQ3 said:
Black and white CPU comparisons are a fools game anyway.
The problem with application tests is that the results aren’t transferable. Even if you look at performance in a single app, lets say Photoshop, you will find that relative performance varies depending on what specific tasks you use it for. Not only that, but it will also vary with for instance the size of the problem set, et cetera.

We simply cannot measure ”general performance” with a degree of precision that is adequate for the small increments in performance that todays lithographic and architectural engineers struggle hard to eke out. It can’t be done.
Click to expand...
That is also true. I for one just look at efficiencies based of a combined result of real world against prior gen and synthetics.
 
EntropyQ3 said:
Black and white CPU comparisons are a fools game anyway.
The problem with application tests is that the results aren’t transferable. Even if you look at performance in a single app, lets say Photoshop, you will find that relative performance varies depending on what specific tasks you use it for. Not only that, but it will also vary with for instance the size of the problem set, et cetera.

We simply cannot measure ”general performance” with a degree of precision that is adequate for the small increments in performance that todays lithographic and architectural engineers struggle hard to eke out. It can’t be done.
Click to expand...
I recommend simply following the standard industry practice of using whichever benchmarks make your favored processor look better than the others. 😄
 
  • Like
  • Haha
  • Love
Reactions: Dismayed, roman.stapunov, SymeonArgyrus and 5 others
theorist9 said:
I recommend simply following the standard industry practice of using whichever benchmarks make your favored processor look better than the others. 😄
Click to expand...
Indeed, that is the forum war approach. 😀
Or for that matter what benchmarketing is, note how the manufacturers (=advertising revenue) supply helpful ”guidelines” for how to benchmark their products.

Luckily, the engineers working on improving the products aren’t stupid, and do as best they can in terms of weighing gains vs. cost while trying to avoid corner case pitfalls and introducing vulnerabilities. Also, these days targeting specific perceived-to-be-critical application areas also makes sense as opposed to primarily going for general purpose tweaking. Which, incidentally, makes sensible cross architecture comparisons even more difficult.
 
  • Like
Reactions: harold.ji and theorist9
leman said:
Why? It's not like those will be any faster than the desktop CPUs...
Click to expand...
They are (particularly Phoenix) more conceptually similar to Apples offerings. Monolithic build, substantial GPU (1536x2xclock GFLOPs) resources backed by on-chip LLC + LPDDR5 main memory, quite modest power draw target, similar process (5nm TSMC variant).
Laptop wise, they mostly share targets, but AMD seems to prefer not offering these processors for the desktop, which would be AM5 in this case. I hope that changes, as these puppies are quite attractive, perhaps too attractive for the market partitioning game. Not due this year however, presumably Apple will have refreshed their offerings by the time Phoenix comes around.
 
Last edited:
  • Like
Reactions: Xiao_Xi
leman said:
Why? It's not like those will be any faster than the desktop CPUs...
Click to expand...
What is the value of a raw performance comparison between two SoCs with such different consumptions?
www.techpowerup.com

AMD Ryzen 9 7950X Specs

Raphael, 16 Cores, 32 Threads, 4.5 GHz, 170 W
www.techpowerup.com www.techpowerup.com

exoticSpice said:
We can see how they compare in perf/w and clock freq and also the power consumption per core and under load.
Click to expand...
Phoenix and Dragon Range will have better performance/consumption than Raphael. Why make a comparison against a worse SoC?
 
You must log in or register to reply here.
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.
Top Bottom
Back