Question Incredible Apple M4 benchmarks...

[Hitman928]

That is very possible. Could you be more specific?

@SarahKerrigan pointed out that relying on a strong correlation between benchmarks to make the case that one can be used in substitute of the other is a very weak argument (especially with a statistically insignificant number of data points) and I agreed. So my posts were to show that not only does their own predictive model break when analyzing additional CPUs, but that the correlation argument is basically meaningless as almost any CPU focused benchmark will show extremely strong correlation with SPECint 1t. I used Cinebench r20 as an example because it is such a poor substitute for SPECint, but still shows near perfect correlation anyway.

 

[roger_k]

@SarahKerrigan pointed out that relying on a strong correlation between benchmarks to make the case that one can be used in substitute of the other is a very weak argument (especially with a statistically insignificant number of data points) and I agreed. So my posts were to show that not only does their own predictive model break when analyzing additional CPUs, but that the correlation argument is basically meaningless as almost any CPU focused benchmark will show extremely strong correlation with SPECint 1t. I used Cinebench r20 as an example because it is such a poor substitute for SPECint, but still shows near perfect correlation anyway.

Thank you! You are right, I should have paid better attention. I full agree with everything you say here.

 

[poke01]

It's more memory bandwidth constrained so Apple wins due to gobs of it available. Should perform better with 12-channel DDR5-8800 on Threadripper/Epycs.

The base M3 is nipping at the heels of the M3 Max. So I don’t believe it’s due to memory bandwidth, in fact I don’t think that plays a huge role. It helps but not that much.

M3 is a dual channel LP5 SoC. The M3 P core is very powerful. There are mobile AMD and Intel parts that score much lower and they also have dual channel memory.

 

[okoroezenwa]

15k MT https://browser.geekbench.com/v6/cpu/6138581

 

[igor_kavinski]

15k MT https://browser.geekbench.com/v6/cpu/6138581

Anyone know why it's hard to get both ST and MT high? It seems to be one or the other. What could be preventing the user from getting 3900 ST in this case?

 

[Nothingness]

Anyone know why it's hard to get both ST and MT high? It seems to be one or the other. What could be preventing the user from getting 3900 ST in this case?

Their cooling system perhaps doesn't last long enough for that

 

[Eug]

15k MT https://browser.geekbench.com/v6/cpu/6138581

Impressive. And it's 0.01 GHz underclocked too.

BTW, as far as I can tell scrolling through the results, without crazy cooling methods it seems reasonable expectations for GB 6.3 scores are roughly as follows:

M4 9-core: ~3700 / 13300, with occasional ones hitting 3800 / 13400 or so.
M4 10-core ~3700 / 14600, with occasional ones hitting 3800 / 14700 or so.

That means real world speed up for 9-core to 10-core is <10%.

Back in the Intel days, I used to say that if the CPU speed upgrade to the next step up wasn't more than about 15%, it probably wouldn't be worth the extra cost for most mainstream consumers.

Apple makes it a memory and storage decision though, since it only pairs 9-core with 8 GB plus 256-512 GB, and 10-core with 16 GB plus 1-2 TB, the latter at a US$600 price premium.

 

[igor_kavinski]

Apple makes it a memory and storage decision though

Someone forgot to tell them that what works for phones need not necessarily apply to Macs...

 

[Eug]

Someone forgot to tell them that what works for phones need not necessarily apply to Macs...

Well, iPads aren't Macs. As many have complained about before, iPadOS is a variant of iOS and is not macOS.

(Both iOS and iPadOS are derived from OS X, but that's a different discussion.)

 

[SpudLobby]

You should read my original Reddit post. Everything you wrote here has either been argued or agreed upon in the Reddit post already.


I'm certain that they do for their consumer CPUs. It's the #1 benchmark for AMD and Intel and unfortunately, Cinebench dominates the minshare space for CPU benchmarking. Things have turned a bit more since I made the Reddit post a few years ago. Quite often, you'll see people on the internet actually cite Nuvia's medium post (for example, @Eug cited it just yesterday) as well as bring up the fact that Cinebench R23 is optimized for x86 and not ARM when comparing between the two architectures.


The person I quoted claims that Cinebench R20 scores are a close predictor of SPEC scores. I'd like to see the data.

You can barely compare Cinebench results to Blender. They don't correlate as much as you think.

I'm not sure how a rendering benchmark became THE mainstream CPU benchmark over the last 6 or 7 years. Most people who use Cinema4D would use a GPU renderer. Same for Blender.

I guess people need a CPU benchmark to justify buying 16 or more consumer CPUs.

I dislike the Cinebench wanking but because of Embree and

The base M3 is nipping at the heels of the M3 Max. So I don’t believe it’s due to memory bandwidth, in fact I don’t think that plays a huge role. It helps but not that much.

M3 is a dual channel LP5 SoC. The M3 P core is very powerful. There are mobile AMD and Intel parts that score much lower and they also have dual channel memory.

View attachment 99148

fwiw, strictly speaking Apple runs 8x16bit channels on it’s RAM, but no I doubt it’s About memory

The newest version (Cinebench 2024) has increased the scene size so that it does not fit in the cache of the most CPUs, and they also fixed a performance issue with their AVX emulation on Apple CPUs.

And of course, the newest Cinebench is being heavily criticized because suddenly Apple Silicon tops the charts, and that is obviously physically impossible, so the benchmark must be flawed.

Yeah the CB2024 fix for NEON is awesome and hilarious RE: making everyone mad.

 

[poke01]

fwiw, strictly speaking Apple runs 8x16bit channels on it’s RAM

Yep, it’s 8 channel. In the end the memory bandwidth is 120GB/s compared to 400GB/s on the M3 Max

 

[SpudLobby]

Yep, it’s 8 channel. In the end the memory bandwidth is 120GB/s compared to 400GB/s on the M3 Max

yeah the memory bandwidth isn’t why Apple’s cores are beating AMD’s in whatever benchmark that was people were complaining about, I agree there (that’s cope).

 

[SpudLobby]

https://twitter.com/x/status/1790354528770084945

Btw this seems about right to me. M2 and M3 are around 7-9.5W depending on the test, so.

 

[SpudLobby]

They’re certainly pumping the frequency more than the node gives them from N3B to N3E.

 

[poke01]

https://twitter.com/x/status/1790354528770084945

Btw this seems about right to me. M2 and M3 are around 7-9.5W depending on the test, so.

Does the use of SME increase power usage in GB6?


Also we need more info here? What was M3 avg watts in SPEC int?

 

[Eug]

https://twitter.com/x/status/1790354528770084945

Btw this seems about right to me. M2 and M3 are around 7-9.5W depending on the test, so.

Let me see if I understand this:

CPU Geekbench 6.3 single-core - 3700+, 11 Watts
CPU Geekbench 6.3 multi-core - 14500+, 32 Watts
GPU Wild Life Extreme - 8800, 28 Watts (29.6 W total minus 1.5 W idle)
Low power mode GPU Wild Life Extreme - 3800, 5.5 Watts

Is this correct?

In the comments the statement was made that M3 was also 30+ Watts with Geekbench multi-core.

Also, I believe this would be M4 low power mode in Geekbench:

iPad16,3 - Geekbench

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

browser.geekbench.com

9-core 2.09 GHz, 1839 / 6846

With these types of scores, most entry level users who aren't gaming but need maximum battery life (while traveling for example) could run in low power mode all the time and not even notice much of a difference.

1839 single-core is roughly 75-80% of M1
6846 multi-core is roughly 75-80% of M1

 

[SpudLobby]

Does the use of SME increase power usage in GB6?

View attachment 99229
Also we need more info here? What was M3 avg watts in SPEC int?

those aren’t even the full SpecInt measurements and look separate to me. I’m skeptical the A17 Pro is doing 3.62W. Fwiw, Andrei had said he thinks the 3.62/7W is nonsense, but the 11W M4 makes sense, which I also suspect based on historical trends.

And SME is only used for one subtest which wouldn’t be enough to crowd out the results.

More likely they just actually draw around 11W platform which is still a crazy result for the ST performance, just still creeping upwards.

 

[SpudLobby]

11W could be too high, but I doubt it’s 7W based on pushing frequency even further than the node gave. I’d love to be wrong though and LPDDR5x is a power win too, but based off that A17 Pro @ 3.62W — no way the M4 is 7W whatever methodology that is.

 

[SpudLobby]

one thing to keep in mind re M4 is that the fundamental curve has likely still improved for the ST or especially the MT. The trouble is just that you can’t granularly select frequencies other than LPM.

 

[SpudLobby]

one thing to keep in mind re M4 is that the fundamental curve has likely still improved for the ST or especially the MT. The trouble is just that you can’t granularly select frequencies other than LPM.

But like take an M4 in ST with 3.2GHz? That thing will have + 15% IPC and should also be like - 30-35% power even for the platform vs M1 — you have more efficient RAM, way better process. Could see it hitting a 2400-2600 GB6 at like 3.5-4.5W range easy. Would be a great result.

 

[SpudLobby]

That’s what sucks about MacOS and iOS. We can’t actually measure as well as Windows or even Android at given frequencies — harder to measure full power curves or even experience them.

 

[mikegg]

11W could be too high, but I doubt it’s 7W based on pushing frequency even further than the node gave. I’d love to be wrong though and LPDDR5x is a power win too, but based off that A17 Pro @ 3.62W — no way the M4 is 7W whatever methodology that is.

Running GB6 ST on my M1 Pro and eyeballing powers, the package power is 1w - 5w for most of the run.

11w would be more than double that.

 

[SarahKerrigan]

those aren’t even the full SpecInt measurements and look separate to me. I’m skeptical the A17 Pro is doing 3.62W. Fwiw, Andrei had said he thinks the 3.62/7W is nonsense, but the 11W M4 makes sense, which I also suspect based on historical trends.

And SME is only used for one subtest which wouldn’t be enough to crowd out the results.

More likely they just actually draw around 11W platform which is still a crazy result for the ST performance, just still creeping upwards.

Per Geekbench 6 Internals, SME is used in three different subtests (Photo Library, Object Detection, and Background Blur.)

 

[Eug]

Per Geekbench 6 Internals, SME is used in three different subtests (Photo Library, Object Detection, and Background Blur.)

That makes sense, since all three subtests got a significant boost in M4 vs. M3.

These features are very important on both pro Macs and consumer Macs. I'm referring to the function's general category, and not specific code per se. Nonetheless, even though I am just looking at this from the outside, I am getting increasingly convinced that these speedups Apple Silicon is seeing in Geekbench are going to be very useful in actual shipping code going forward.

I've talked about this before, but for example, Apple Photos ships with every Mac and all photos are scanned in a background service for face recognition, etc. For those with sync'd iPhones, that means all of those many thousands of photos are scanned automatically not only on their iPhones but also on their iPads and Macs. This is not exactly a fast process so anything that speeds this up and minimizes power usage would be welcome.

Object detection is also integral to Photos and Final Cut Pro, both for face and pet recognition and for the automatic rotoscoping feature. (Forgive me if I don't have the terminology totally correct.) Apple also has a cut-and-paste feature for people in photos, where you can automatically cut out those people and use them in other apps like iMessage.

Background Blur is also an integral feature. For example, this is the basis for Apple's fake portrait mode photos. Using object detection and edge detection to recognize and trace humans, they blur out the background for fake bokeh. Furthermore, with newer iPhones, the portrait photos are no longer permanently portrait mode still photos but are in fact regular photos, both Live Photos and still photos, but with depth metadata captured along with it so that you can create fake portrait photos with background blur after the fact. In this context, you no longer are just doing this on the iPhone, but also on iPads and Macs, so it makes a lot of sense for these functions to be as fast as possible across the board.

With regards to portrait mode, my impression is that not only has the quality of the object detection improved with newer versions of iOS, it has also improved with newer versions of the iPhone. To me this implies that there are both software and hardware components to the quality of this feature. IOW, on my iPhone 12 Pro Max, I have better portrait mode pictures now with the latest iOS than I did when I first got the phone 4 years ago, but it's not as good as people with the iPhone 15 Pro Max running the same iOS version.

 

[SpudLobby]

Running GB6 ST on my M1 Pro and eyeballing powers, the package power is 1w - 5w for most of the run.

11w would be more than double that.

Lmao. Here we go, muh PowerMetrics from senttoschool. This isn’t accurate, Geekbench is a composition of different tests, and power metrics doesn’t even measure package power anymore. It just measures “CPU power”.

You want to include the whole platform. That’s what Andrei does, that’s what people who aren’t lobotomized dull iSheep do. Your M1 Pro is drawing much more than 5W to begin with, and Apple has pushed up power.

Even an iPhone draws 4-5W for GB5/6 or Spec dude.