Thursday, May 9th 2024

Apple M4 Chip Benchmarked: 22% Faster Single-Core and 25% Faster Multi-Core Performance

Yesterday, Apple launched its next-generation M4 chip based on Apple Silicon custom design. The processor is a fourth-generation design that brings AI capabilities and improved CPU performance. First debuting in an iPad Pro, the CPU has been benchmarked in Geekbench v6. And results seem to be very promising. The latest M4 chip managed to score 3,767 points in single-core tests and 14,677 points in multi-core tests. Compared to the M3 chip, which scores 3,087 points in single-core and 11,702 in multi-core tests, the M4 chip is about 22% faster in single-core and 25% faster in multi-core synthetic benchmarks.

Of course, these results are not real-world use cases, but they give us a hint of what the Apple Silicon design team has been working on. For real-world results, we have to wait a little longer to see reviews and results from devices such as MacBook Pro and MacBook Air, which should have better cooling and possibly better clocks for the chip.
Sources: Geekbench v6, via Vadim Yuryev on X
Add your own comment

51 Comments on Apple M4 Chip Benchmarked: 22% Faster Single-Core and 25% Faster Multi-Core Performance

#26
Space Lynx
Astronaut
Mac Mini M4 is going to be an impressive device as long as they do proper airflow in it.

If my education discount will bring it down to $499 at launch I will def go that route for my work station. The m3 was that price at launch for my discount level, but I decided to hold off.
Posted on Reply
#27
AnotherReader
bencrutzimpressive due to the SME, the highest jump, look at the details of the GB score.
CPU IPC increase is actually meh

hint: object detection
There's a post on Anandtech's forums comparing the M4 to the A17. It seems that the improvements in non SME accelerated tests are primarily due to clock speed. Code compilation, i.e. the Clang subtest, sees practically no increase in IPC.
Posted on Reply
#28
Tek-Check
Guwapo77M4 has a great performance increase of 22-25%, but its nothing compared to Zen 5's 10% increase. *insert clown emoji*
Let us know when you finally learn the basic difference between IPC increase and performance increase. Thank you.
Posted on Reply
#29
ikjadoon
Chrispy_Apple chips are actually good. They're state-of-the-art ARM.

My issue is with Geekbench being used.

Most of the M1, M2, M3, M4 performance can be attributed to Apple forcing developers to recompile their software specfically for their silicon. They ditched compatibility to gain performance which is a perfectly acceptable trade-off for Apple's walled garden. ARM performance and IPC is a known quantity; Apple are bolstering it with high-speed interfaces to fast storage and fast RAM, a decent power budget and state of the art manufacturing for the best performance and the best performance/Watt. A lot fhe M-silicon's success is down to Apple's software ecosystem and clout with Developers, but this also isn't just your basic tablet; These things have high-end platform componentry and a level of integration that x86 hardware vendors will never get close to. Even if it's just ARM, it's the best example of ARM you'll ever see.
Every CPU uArch engineer uses Geekbench. There's nothing re: Apple or Arm "forcing" Geekbench to "recompile" for their software or "software ecosystem" nor "clout with developers". SME is an Arm-standard extension. We can analyze Apple's cores just like any x86 core. Apple isn't reinventing what a CPU does or how it works.

You'll never hear Intel and AMD make the excuse that "Geekbench is for Apple". In reality, Intel also uses Geekbench. AMD also uses Geekbench.

TBH, people don't realize how incredibly widespread, well-tested, and well-understood Geekbench is for CPU designers. When Intel calculated its Alder Lake +19% IPC uplift, Intel calculated it using only six benchmark suites:
  • SYSmark 25
  • CrossMark
  • PCMark 10
  • SPEC CPU 2017
  • WebXPRT 3
  • Geekbench 5
Geekbench familiarity is a preferred experience in Intel's job descriptions for power-perf engineers:




AMD relies heavily on Geekbench for commercial PC testing:



When AMD touted Zen4 had +29% faster 1T perf than Zen3? 100% Geekbench.




AMD, though, gets a little embarrassed about Geekbench when it needs to compare to Apple (for good reason: AMD will lose handily. These are corporations. They won't use the XYZ benchmark externally if the XYZ benchmark results look bad for them.).

AMD, when comparing to Intel Meteor Lake: "Geekbench proves we are faster!"
AMD, when comparing to Apple M2: "Whoa, we're losing. Quick, find another benchmark."

TL; DR: Anyone can complain, but CPU designers rely on industry-standard benchmarks like Geekbench for a reason: reproducibility, relevance to consumers, and excellent for cross-OS uArch tests.

There is nothing niche about Geekbench and there's no specific bias with Arm CPUs when using Geekbench.

EDIT: fixed links
Geofrancisgeekbench is a joke.

I want to see some classic benchmarks that actually use the CPU cores and not the bunch of accelerators attached to it.
FYI, Geekbench 6 is a CPU core-only benchmark. It uses zero accelerators, e.g., nothing except the CPU instructions themselves.

www.geekbench.com/doc/geekbench6-benchmark-internals.pdf
Posted on Reply
#30
Darmok N Jalad
Quantify things how ever you'd like, but I've owned an M2 15 MBA for about a year now, and it's hands-down the best laptop I've ever used, and I have been doing this for a few decades. I also get issued pro-grade laptops at work every couple years, and I have the ability to level-up the options. It's not even close in terms of total experience. Dead silent operation, solid build quality, only need to plug it in once a week. I never thought such a device was possible with the way most modern mobile chips consume the power and put off heat. Do I do insane workloads? No, but what I do run (photo edits), runs very fast, with no complaints from me. Even Paul Thurrott had to compliment his M3 MBA, which I never thought I'd see the day.
Posted on Reply
#31
Geofrancis
ikjadoonFYI, Geekbench 6 is a CPU core-only benchmark. It uses zero accelerators, e.g., nothing except the CPU instructions themselves.
My point is it's a collection of benchmarks rather than a single task, and unless I'm missing something, there is no breakdown of each task. intel were caught manipulating benchmarks like these many times by just getting the benchmark company to run the tests more favourable to intel more often. apple have added a lot of on cpu accelerators, we have no idea what geekbench benchmarks are taking advantage of this hardware so how can we tell if its CPU performance or some onboard math or video accelerator giving it the better score.

with tests like pi or prime numbers it's much harder to manipulate, especially if its opensource.


Another way to look at it is apple CPUs are designed for apple products. So rather than add lots of math extensions like x86 done with SSE SSE4 etc, they just added a specific accelerator for their specific formats. so instead of video extensions they just add a full video converter so as long as its apple software and you use those formats it will be faster on the apple computer. they have done this with a lot of functions like storage, video, audio, cameras, cryptography, AI. so this is why im sceptical about performance gains in geekbench because you cannot tell if its raw CPU performance or one of the accelerators doing the work. especially when you look at something like Cinebench, its raw compute and the m3 pro around the same performance as a Intel Core i5-13500H or ryzen 5800H.
Posted on Reply
#32
Guwapo77
Tek-CheckLet us know when you finally learn the basic difference between IPC increase and performance increase. Thank you.
Until you learn basic math... understand that there is a very decent IPC increase here as well. It all didn't come from a frequency increase alone.
Posted on Reply
#33
Tek-Check
Guwapo77Until you learn basic math... understand that there is a very decent IPC increase here as well. It all didn't come from a frequency increase alone.
Math is different when you say 22-25% performance increase in specific benchmark and compare it to 10% IPC increase on Zen5.

Face value comparison math in this case is meaningless. Got it?
Posted on Reply
#34
Lew Zealand
AnotherReaderThere's a post on Anandtech's forums comparing the M4 to the A17. It seems that the improvements in non SME accelerated tests are primarily due to clock speed. Code compilation, i.e. the Clang subtest, sees practically no increase in IPC.
5.4% IPC increase when excluding Object Detection. 11.3% when including OD. So:

1/2 the claimed performance increase from clock speed
1/4 the claimed performance increase from Object Detection alone
1/4 the claimed performance increase from other IPC/architectural improvements

and a bit more from some multicore enhancements.
Posted on Reply
#35
Darmok N Jalad
Space LynxMac Mini M4 is going to be an impressive device as long as they do proper airflow in it.

If my education discount will bring it down to $499 at launch I will def go that route for my work station. The m3 was that price at launch for my discount level, but I decided to hold off.
The mini is cooled adequately, and even though it has a fan, you'll be hard-pressed to hear it running. I had the M1 mini first, and I don't think it ever pulled more than 30w from the wall.
Posted on Reply
#36
ikjadoon
GeofrancisMy point is it's a collection of benchmarks rather than a single task, and unless I'm missing something, there is no breakdown of each task. intel were caught manipulating benchmarks like these many times by just getting the benchmark company to run the tests more favourable to intel more often. apple have added a lot of on cpu accelerators, we have no idea what geekbench benchmarks are taking advantage of this hardware so how can we tell if its CPU performance or some onboard math or video accelerator giving it the better score.

with tests like pi or prime numbers it's much harder to manipulate, especially if its opensource.


Another way to look at it is apple CPUs are designed for apple products. So rather than add lots of math extensions like x86 done with SSE SSE4 etc, they just added a specific accelerator for their specific formats. so instead of video extensions they just add a full video converter so as long as its apple software and you use those formats it will be faster on the apple computer. they have done this with a lot of functions like storage, video, audio, cameras, cryptography, AI. so this is why im sceptical about performance gains in geekbench because you cannot tell if its raw CPU performance or one of the accelerators doing the work. especially when you look at something like Cinebench, its raw compute and the m3 pro around the same performance as a Intel Core i5-13500H or ryzen 5800H.
Geekbench has a public methodology (though this is the GB6.0 version) with a breakdown of each task and if any ISA-specific extensions are tapped into: www.geekbench.com/doc/geekbench6-benchmark-internals.pdf

Nothing is Apple-specific and in fact, Geekbench refused to support Apple's older AMX ISA extensions. Geekbench can't tap into accelerators outside the CPU; ISA extensions, yes, accelerators, no, just like virtually every CPU benchmark.

//

Unfortunately PiFast and wPrime are Windows only, but a quick search showed that the M1 / M2 / M3 do well in an open source Pi benchmark.

Apple's silicon does too well across too many domains for it to be any accelerator fluke or "Apple is gaming the benchmark!" Of course, some poor reviewers don't understand what is a CPU benchmark and what is an SoC benchmark, but across virtually all CPU benchmarks, the M1 / M2 / M3 perform exceedingly well and are close to or exceed Intel & AMD in 1T.

//

Re: video decoders / encoders: everyone (AMD, Intel, Qualcomm, Apple) ships HW-accelerated IP for video encode / decode, though. That isn't Apple-specific. PC reviewers can, and most do, toggle their benchmarks to use the CPU-only whenever some app allows you to choose. There have been a few silly reviews, though, where someone doesn't understand how to use Handbrake.

//

Cinebench:

IMO, that is still impressive. You're comparing CPUs with over +50% more power available.

As Cinebench's current version doesn't have many mega-reviews yet, these are all collated: M3 Pro results & power, i9-13980HX results, i7-12700K & 7700X results, 155H results. The i7-14700HX and i5-1235U, you need to use NBC's search bar below the charts to "Add an additional device", as the device reviews don't include newer tests.

I wanted to add more ~30W TDP tests, but Notebookcheck hasn't added many newer laptops yet, either. :(

R24 1TR24 nT
Apple M3 Pro (6+6) - 27W143 - 100%1059 - 100%
Intel i5-1235U (2+8) - 20W PL193 - 65%406 - 38%
AMD 7840HS (8+0) - 45W TDP104 - 72%923 - 87%
Intel Core Ultra 7 155H (6+8+2) - 20W TDP106 - 74%733 - 69%
Intel i7-14700HX (8+12)101 - 71%1003 - 95%
AMD Ryzen 7700X (6+0)116 - 83%1070 - 103%
Intel i7-12700K (8+4)114 - 81%1169 - 112%
Intel i9-19380HX (8+16)125 - 90%1665 - 160%


At its power level, the M3 Pro is doing extremely well, IMO.


Posted on Reply
#37
Guwapo77
Tek-CheckMath is different when you say 22-25% performance increase in specific benchmark and compare it to 10% IPC increase on Zen5.

Face value comparison math in this case is meaningless. Got it?
And there is still more IPC gain in this 22% performance increase...got it? :banghead::banghead::banghead:


M3 Single Core - 3217 - browser.geekbench.com/v6/cpu/search?utf8=✓&q=Apple+M3
M4 Single Core - 3824 - browser.geekbench.com/search?utf8=✓&q=ipad16,5


Its all because of the 300 mhz boost increase. :roll::roll::roll:
Posted on Reply
#38
Minus Infinity
NhonhoDo you know why Apple still hasn't put an AV1 encoder in its CPUs?
They have: From Apple's own webpage

A new advanced Media Engine includes support for AV1 decode, providing more power-efficient playback of high-resolution video experiences from streaming services.
Posted on Reply
#39
Guwapo77
Minus InfinityThey have: From Apple's own webpage

A new advanced Media Engine includes support for AV1 decode, providing more power-efficient playback of high-resolution video experiences from streaming services.
Encoder was what he was looking for. M3 had a decoder as well.
Posted on Reply
#40
Tek-Check
Guwapo77And there is still more IPC gain in this 22% performance increase...got it?
You don't know this, do you?
Posted on Reply
#41
pat-roner
Space LynxMac Mini M4 is going to be an impressive device as long as they do proper airflow in it.

If my education discount will bring it down to $499 at launch I will def go that route for my work station. The m3 was that price at launch for my discount level, but I decided to hold off.
I'm very much looking forward to trying one of those. Love the form factor and so much power within such a small enclosure. I hope it gets a redesign soon though - with perhaps some front facing usb-c's
Posted on Reply
#42
Chrispy_
ikjadoonEvery CPU uArch engineer uses Geekbench. There's nothing re: Apple or Arm "forcing" Geekbench to "recompile" for their software or "software ecosystem" nor "clout with developers". SME is an Arm-standard extension. We can analyze Apple's cores just like any x86 core. Apple isn't reinventing what a CPU does or how it works.

You'll never hear Intel and AMD make the excuse that "Geekbench is for Apple". In reality, Intel also uses Geekbench. AMD also uses Geekbench.

TBH, people don't realize how incredibly widespread, well-tested, and well-understood Geekbench is for CPU designers. When Intel calculated its Alder Lake +19% IPC uplift, Intel calculated it using only six benchmark suites:
  • SYSmark 25
  • CrossMark
  • PCMark 10
  • SPEC CPU 2017
  • WebXPRT 3
  • Geekbench 5
Geekbench familiarity is a preferred experience in Intel's job descriptions for power-perf engineers:




AMD relies heavily on Geekbench for commercial PC testing:



When AMD touted Zen4 had +29% faster 1T perf than Zen3? 100% Geekbench.




AMD, though, gets a little embarrassed about Geekbench when it needs to compare to Apple (for good reason: AMD will lose handily. These are corporations. They won't use the XYZ benchmark externally if the XYZ benchmark results look bad for them.).

AMD, when comparing to Intel Meteor Lake: "Geekbench proves we are faster!"
AMD, when comparing to Apple M2: "Whoa, we're losing. Quick, find another benchmark."

TL; DR: Anyone can complain, but CPU designers rely on industry-standard benchmarks like Geekbench for a reason: reproducibility, relevance to consumers, and excellent for cross-OS uArch tests.

There is nothing niche about Geekbench and there's no specific bias with Arm CPUs when using Geekbench.

EDIT: fixed links



FYI, Geekbench 6 is a CPU core-only benchmark. It uses zero accelerators, e.g., nothing except the CPU instructions themselves.

www.geekbench.com/doc/geekbench6-benchmark-internals.pdf
Thanks for the solid info, I guess it's more manufacturers exploiting Geekbench rather than Geekbench itself being shit.

My only hate of Geekbench is based on the simple experience of years of independent reviews proving that Geekbench score doesn't have anything to do with real-world results.
Posted on Reply
#43
pat-roner
Chrispy_Thanks for the solid info, I guess it's more manufacturers exploiting Geekbench rather than Geekbench itself being shit.

My only hate of Geekbench is based on the simple experience of years of independent reviews proving that Geekbench score doesn't have anything to do with real-world results.
Isn't that usually the fact though, Benchmarks not actually mimicking real world use? Not that all of them are like that though. It's just a really objective way of comparings stuff under the same load.
Posted on Reply
#44
Dredi
Chrispy_Thanks for the solid info, I guess it's more manufacturers exploiting Geekbench rather than Geekbench itself being shit.

My only hate of Geekbench is based on the simple experience of years of independent reviews proving that Geekbench score doesn't have anything to do with real-world results.
Especially in phone tests it used to be a Wild West with a lot of shenanigans done by phone manufacturers to make it run faster. The only real issue with the test is that it is too short, and thermal limits do not necessarily realise.

My experience with the apple chips is that they are very good. We got some at work and they trash basically everything at everything in the same device category. Especially so in applications where the accelerators are utilized and battery life matters. It's not an absurd difference, but a clearly noticeable one.

Should there be a valid game support ( ~steam os) and possibility to add external graphics cards, I would switch to these even for my desktop gaming pc. And that is coming from someone happily building PCs for over two decades.
Posted on Reply
#45
Guwapo77
Tek-CheckYou don't know this, do you?
I'm sure chip makers want to increases to IPC without any performance gains. Miss me with the semantics.
Posted on Reply
#46
mrnagant
Let me use MacOS on it and I'd be interested. iPadOS is the most limiting factor of all this power and capability. Multitasking on iPadOS is essentially non-existent. Half my apps run in phone window mode and the other half are just stretched out phone apps.

Would be sweet to be able to use it as a tablet, and then at my workstation mag clip it into a dock with multiple monitors and accessories. Heck, I wouldn't even really be opposed if it switched between iPadOS and MacOS on the fly depending if you were in docked mode or tablet mode.

Microsoft Surface has the capabilities I want, but then Apple has the hardware I want.
Posted on Reply
#47
Space Lynx
Astronaut
pat-ronerI'm very much looking forward to trying one of those. Love the form factor and so much power within such a small enclosure. I hope it gets a redesign soon though - with perhaps some front facing usb-c's
Apple will never do front facing USB slots on this device. the whole point of Apple is image/clean look. I am alright with this.

for $499 a m4 Mac Mini is going to be great bang for buck regarding work stuff.
Posted on Reply
#48
Darmok N Jalad
Space LynxApple will never do front facing USB slots on this device. the whole point of Apple is image/clean look. I am alright with this.

for $499 a m4 Mac Mini is going to be great bang for buck regarding work stuff.
We’ll see. They’ve gotten a bit more port friendly with the departure of Ive. Still could be better, and the question is how willing they are to stop reusing the classic mini chassis and invest in a new design. The mini is the affordable Mac, so Apple often scrimps there for better margins.
Posted on Reply
#49
Tek-Check
Guwapo77I'm sure chip makers want to increases to IPC without any performance gains. Miss me with the semantics.
You were initially challenged on comparing wrong numbers, apparently due to misunderstanding the difference between IPC and generic performance in synthetic benchmark. There are also real life workloads, but that's yet another topic and more numbers.

M3->M4 ST/MT leaked increase is 22% and 25%. What is IPC increase? We don't know at the moment. Could be ~10%
Zen3-Zen4 ST/MT increase was 29% and 48% for the top SKU, at 14% IPC increase.
Zen4-Zen5 ST/MT increase is unknown. Leaks on IPC increase suggest 10-15%, which could bring ST 20-30% and MT 30-40%. We will know soon.
Posted on Reply
#50
Minus Infinity
Guwapo77Encoder was what he was looking for. M3 had a decoder as well.
Sorry too quick on the trigger. I thought it was new feature they added seeing it was highlighted.
Posted on Reply
Add your own comment
Jun 2nd, 2024 20:08 EDT change timezone

New Forum Posts

Popular Reviews

Controversial News Posts