Question Incredible Apple M4 benchmarks...

[IGBT]

Incredible Apple M4 benchmarks suggest it is the new single-core performance champ, beating Intel's Core i9-14900KS — results of 3,800+ posted...Apple's M4 processors have become convincing leaders in the Geekbench single-core leaderboard. Several scores of roughly 3,800 points have appeared in the Geekbench online database over recent hours. This is significant as single-core benchmark scores of this magnitude put clear blue water between the M4 and Intel’s flagship Core i9-14900KS. A little Geekbench database checking shows that, in single-threaded tests, Apple's M4 outpaces Intel's power-hungry desktop champ by about 16%. https://www.tomshardware.com/pc-com...-4777-BC4E-5B196F46C7D1&utm_source=SmartBrief

 

[WelshBloke]

We don't talk about the i9-14900ks any more!

 

[Jan Olšan]

The GB6 score is just artifact of SME extension being used in GB 6.3, which increases performance by 125% in one single subtest (and in singel thread only, it mostly seem to disappear in multithread). The averaged final score is mostly not representative of everything else for that reason.

But basically the processor just has at best ~3% higher IPC possibly just due to faster memory used*, and 9% higher clocks. The rest is the AI acceleration schananigans via SME.
Basically it should have ~12% higher single-thread performance than M3. You would get about 3450 points ST score without that single SME-skewed subtest. Which is still good of course.
If you tested with GB5, you would probably get a more grounded comparison to M3 (or other CPUs).

* it's probably the same core as M3, although some refresh-level fixes/changes to the core are not ruled out.

 

[igor_kavinski]

Good journalism there! Report through the lies and smoke!

 

[poke01]

If you tested with GB5, you would probably get a more grounded comparison to M3 (or other CPUs).

This is the highest GB5 score I found so far.

SC: 2641.

iPad16,4 - Geekbench

Benchmark results for an iPad16,4 with an ARM processor.

browser.geekbench.com

M4 is probably the fastest CPU for now. I know the Core i9s will the e cores disabled will score in the 3000s in GB5 but I consider stock scores considering the M4 you know doesn’t guzzle power.


It’s massive feat considering Intels nor AMD CPUs can’t manage that single core in laptops or tablets.
You need a desktop i9 to get to M4 levels.

Regarding SME, since Apple using an industry standard ARM extension and the world is an AI craze right now, I wouldn’t be surprised if ARM vendors also start supporting SME soon.

Edit: there’s still so much we don’t know about about M4. It’s PPW and it’s area.

Also here SPEC. A bit older but you can probably speculate where Intel and Apple lie in for M4 Max and Arrow Lake.

 

[gdansk]

This composite ST score says it is 50% faster than M2. For some reason I don't think that's representative of the performance difference.
Apple themselves chose to say it's 50% faster in multi-threaded (Affinity Photo 2 benchmark) for a reason.

 

[mavere]

About 38% faster in SC in GB5 comparing the high score above against a high-scoring m2 iPad entry. https://browser.geekbench.com/v5/cpu/compare/22466510?baseline=22468481

+10% IPC over 2 years, I guess.

 

[Doug S]

The GB6 score is just artifact of SME extension being used in GB 6.3, which increases performance by 125% in one single subtest (and in singel thread only, it mostly seem to disappear in multithread). The averaged final score is mostly not representative of everything else for that reason.

If you remove the Objection Detection test that's affected by SME it is still scoring in the mid 3600s.

 

[Jan Olšan]

If you remove the Objection Detection test that's affected by SME it is still scoring in the mid 3600s.

That sounds like basing the guesstimating math from some very high M3 score.
Primate Labs lists a reference score of just 3087 for M3, I was starting from that.
Possible problem is that that's a MacOS score.

 

[thunng8]

The GB6 score is just artifact of SME extension being used in GB 6.3, which increases performance by 125% in one single subtest (and in singel thread only, it mostly seem to disappear in multithread). The averaged final score is mostly not representative of everything else for that reason.

But basically the processor just has at best ~3% higher IPC possibly just due to faster memory used*, and 9% higher clocks. The rest is the AI acceleration schananigans via SME.
Basically it should have ~12% higher single-thread performance than M3. You would get about 3450 points ST score without that single SME-skewed subtest. Which is still good of course.
If you tested with GB5, you would probably get a more grounded comparison to M3 (or other CPUs).

* it's probably the same core as M3, although some refresh-level fixes/changes to the core are not ruled out.

How did you calculate 3450? It is approx 3660 without it. Remember is it the geometric mean calculated, not just a mean

 

[Nothingness]

How did you calculate 3450? It is approx 3660 without it. Remember is it the geometric mean calculated, not just a mean

Yeah, though it's a bit more complex:

Geekbench 6 provides two composite scores: single-core and multi-core. These scores are
computed using a weighted arithmetic mean of the subsection scores. The subsection scores
are computed using the geometric mean of the scores of the workloads contained in that
subsection.

https://www.geekbench.com/doc/geekbench6-benchmark-internals.pdf

 

[jeanlain]

The GB6 score is just artifact of SME extension being used in GB 6.3, which increases performance by 125% in one single subtest (and in singel thread only, it mostly seem to disappear in multithread). The averaged final score is mostly not representative of everything else for that reason.

Sorry for being lazy to check myself, but the M4 isn't faster than the intel GPU in other tests?

EDIT: it apparently is.

 

[jeanlain]

If you remove the Objection Detection test that's affected by SME it is still scoring in the mid 3600s.

M4 can detect objections faster, and reject arguments that its scores are biased. Impressive.

 

[poke01]

what’s the Intel equivalent of SME?

 

[mikegg]

what’s the Intel equivalent of SME?

AVX, which is already used in Geekbench 6. So this is just ARM chips catching up with x86 in SIMD acceleration through CPU instructions.

Previously, Apple used CoreML for these tasks which could use either the AMX, NPU, or GPU. GB6 did not use CoreML for CPU tests. Now, the AMX has been moved to SME CPU instruction sets, making it far easier for programs to utilize it through the CPU.

So in my opinion, this is just Apple/ARM catching up to AVX where previously, they had a disadvantage. The scores between x86 and ARM in GB6.3 are still valid.

If people think M4's scores are only/too high because of SME, what do people think about Geekbench already having optimizations for for AVX but not Apple's AMX processor previously?

The GB6 score is just artifact of SME extension being used in GB 6.3, which increases performance by 125% in one single subtest (and in singel thread only, it mostly seem to disappear in multithread). The averaged final score is mostly not representative of everything else for that reason.

But basically the processor just has at best ~3% higher IPC possibly just due to faster memory used*, and 9% higher clocks. The rest is the AI acceleration schananigans via SME.
Basically it should have ~12% higher single-thread performance than M3. You would get about 3450 points ST score without that single SME-skewed subtest. Which is still good of course.
If you tested with GB5, you would probably get a more grounded comparison to M3 (or other CPUs).

* it's probably the same core as M3, although some refresh-level fixes/changes to the core are not ruled out.

Even so, regardless of what people here think about SME, the M4 beats the 14900KS in 14/16 ST tests. https://browser.geekbench.com/v6/cpu/compare/6059033?baseline=6050769

And for those who somehow want to discredit M4 based on low IPC improvement over M3, who cares if perf/watt is drastically better? A design can optimize for more IPC or higher clock speeds. In this case, M4's design seems to emphasize higher clocks. Now if power also went up drastically, then we have some issues. But that's unlikely given how thin the iPad Pro is.

This is a 10-15w max SoC vs a 500w max CPU. For some perspective, the entire M4 iPad Pro is thinner than Intel's 14900KS package.

 

[poke01]

AVX, which is already used in Geekbench 6. So this is just ARM chips catching up with x86 in SIMD acceleration through CPU instructions.

Previously, Apple used CoreML for these tasks which could use either the AMX, NPU, or GPU. Now, the AMX has been moved to SME CPU instruction sets, making it far easier for programs to utilize it through the CPU.

So in my opinion, this is just Apple/ARM catching up to AVX where previously, they had a disadvantage. The scores between x86 and ARM in GB6.3 are still valid.

Even so, regardless of what people here think about SME, the M4 beats the 14900KS in 14/16 ST tests. https://browser.geekbench.com/v6/cpu/compare/6059033?baseline=6050769

And for those who somehow want to discredit M4 based on low IPC improvement over M3, who cares if perf/watt is drastically better? A design can optimize for more IPC or higher clock speeds. In this case, M4's design seems to emphasize higher clocks. Now if power also went up drastically, then we have some issues. But that's unlikely given how thin the iPad Pro is.

For some perspective, the entire M4 iPad Pro is thinner than Intel's 14900KS package.

Thanks for the detailed response.

I been reading this piece and it goes thru Intel AMX and Apple AMX (which has now been replaced by SME)

Contrasting Intel AMX and Apple AMX

 

[mikegg]

If you remove the Objection Detection test that's affected by SME it is still scoring in the mid 3600s.

I sense that Apple Silicon has given a lot of non-Mac users angst. AMD/Intel consumer users can no longer say they have the fastest CPU in the world since M1 came out and that annoys them.

 

[poke01]

My other question is, why is Apple scoring real high in the Object Detection subsection when compared to Intel if Intel also has matrix operations?

Is Apples implementation of matrix ops better?

 

[Nothingness]

what’s the Intel equivalent of SME?

It's AMX.
NEON and SVE are the equivalent of AVX*

 

[roger_k]

The GB6 score is just artifact of SME extension being used in GB 6.3, which increases performance by 125% in one single subtest (and in singel thread only, it mostly seem to disappear in multithread). The averaged final score is mostly not representative of everything else for that reason.

But basically the processor just has at best ~3% higher IPC possibly just due to faster memory used*, and 9% higher clocks. The rest is the AI acceleration schananigans via SME.
Basically it should have ~12% higher single-thread performance than M3. You would get about 3450 points ST score without that single SME-skewed subtest. Which is still good of course.
If you tested with GB5, you would probably get a more grounded comparison to M3 (or other CPUs).

* it's probably the same core as M3, although some refresh-level fixes/changes to the core are not ruled out.

Not really. The IPC improvements range from 0% in a few tests to 5-7% in some overs and >10% for a few more. Regardless, M4 beats any currently released desktop CPU for every GB6 subtest, which is the point OP is making. Frankly, the IPC gains could be negative for all that matters, what’s important is that Apples 5W passively cooled core confidently outperforms the latest and greatest enthusiast and workstation desktop market can muster.

 

[igor_kavinski]

that annoys them.

No, what annoys us is Apple's walled garden. If they would let us run Windows and our current games library with proper DX12/Vulkan drivers, why would we be angry?

 

[igor_kavinski]

@adroc_thurston , how annoyed are you by M4?

 

[amosliu137]

For gb5

iPad16,3 - Geekbench

Benchmark results for an iPad16,3 with an ARM processor.

browser.geekbench.com

 

[jeanlain]

If they would let us run Windows and our current games library with proper DX12/Vulkan drivers [...]

You speak as if Apple just had to give a permission.
But they would have to develop, maintain and support Windows drivers for MX SoCs, including GPUs and all... all that for a low percentage of users who would still be unhappy with the performance due to translation of X86 code.

 

[poke01]

Okay, these new scores are bonkers. Both for GB5 and GB6.

iPad16,4 - Geekbench

Benchmark results for an iPad16,4 with an ARM processor.

browser.geekbench.com

I believe these large scores are the scores with the help of some external cooling. So we can expect these scores in a Mac mini or MacBook Pro.

I bet Geekerwan, tested using a larger cooler