Thursday, November 12th 2020

Apple's M1-Based MacBook Air Benchmarked

When Apple announced that they are going to switch their Mac lineup from Intel-based x86 processors to the custom "Apple Silicon," everyone was wondering how the new processors will look and perform. To everyone's luck, Apple has just a few days ago announced its first Apple Silicon custom processor for MacBook. The M1, as the company calls it, is their first processor designed for higher-power and performance tasks The M1 features eight CPU cores (four high-performance and four-high efficiency) paired with eight cores dedicated to the graphics. On the die, there is also a 16-core neural engine made to accelerate machine learning tasks found in the new applications.

Today, we are getting the first GeekBench 5 CPU benchmarks that showcase just how far Apple has come with its custom design. What we have is the M1 processor found in MacBook Air. This Mac model features a passive cooling system, cooling a CPU with a base frequency of 3.2 GHz. The system scored 1719 points in the single-core result, and 6967 points in the multi-core result. The single-threaded results measure itself with some of the highest-end offerings from Intel and AMD, while the multi-threaded results are very good given the mix and match of small and big cores.
Source: GeekBench 5
Add your own comment

116 Comments on Apple's M1-Based MacBook Air Benchmarked

#1
Sandbo
These are actually very good numbers for a laptop:
My AMD Ryzen 3600X desktop is ST 1251 and MT 6973.
A laptop with AMD 4900HS is ST 1091 and MT 7075.
More: browser.geekbench.com/processor-benchmarks/

I guess at the beginning the software support might be limited, but this suddenly makes Apple's laptop a high performance option, especially at $999.
I had been discouraged by the fact that paying a premium for mac is giving me overpriced, outdated spec, guess this has changed now.
Posted on Reply
#2
Metroid
Incredible st performance.
Posted on Reply
#3
Rahnak
Geekbench is worthless for comparing between different architectures.
Posted on Reply
#4
Vya Domus
Here is why Geekbench is dogshit :



Every time other SoCs inch closer to Apple's the app gets updated and a chasm appears again between their chips and everyone else's.

In 4.4 there was less than 10% difference between A13 and Exynos 990, in 5 that somehow became a colossal 50%. This happens every single time a new version appears, without exception. Only an idiot would take these numbers for granted and not realize that this benchmark is always optimized specifically for Apple's chips.
Posted on Reply
#5
sweetsuicide
Rahnak
Geekbench is worthless for comparing between different architectures.
I totally agree with you. This is AArch64, not the day-to-day application performance.
Posted on Reply
#6
dyonoctis
Let's just wait for some leaks of that new R23 cinbench :). Or just next week even.
Posted on Reply
#7
windwhirl
Vya Domus
Here is why Geekbench is dogshit :



Every time other SoCs inch closer to Apple's the app gets updated and a chasm appears again between their chips and everyone else's.

In 4.4 there was just 10% difference between A13 and Exynos 990, in 5 that somehow became a colossal 50%. This happens every single time a new version appears, without exception. Only an idiot would take these numbers for granted and not realize that this benchmark is always optimized specifically for Apple's chips.
That's starting to look like another Userbenchmark
Posted on Reply
#8
Fourstaff
Wonder if there are any other cross platform benchmarks we can use with high degree of confidence. A lot of people seem to dislike Geekbench.
Posted on Reply
#9
thevoiceofreason
Fourstaff
Wonder if there are any other cross platform benchmarks we can use with high degree of confidence. A lot of people seem to dislike Geekbench.
You mean like for example SPECint, here showing Apple's previous A14 chip powering Iphone 12?



I'm very curious about Cinebench R23 results of the M1.
Posted on Reply
#10
Fourstaff
thevoiceofreason
You mean like for example SPECint, here showing Apple's previous A14 chip powering Iphone 12?



I'm very curious about Cinebench R23 results of the M1.
There are also people who like to say SPECint is not representative of real world usage.
Posted on Reply
#11
lemonadesoda
Real-world figures needed. Without a doubt Apple is using whatever PR and benchmarking tricks they can in order to make M1 look good. The question is: todays application software, how does it perform? And I don't mean having the whole CPU and memory space dedicated to one application, but making it work with multitasking or memory constrained typical use situation: e.g. Right now I have 50 tabs open in Chrome instance 1, another 50 or so tabs in Chrome instance 2, Chromium with another 30 tabs there, skype, MS Office, Adobe Acrobat, and then some. How the OS handles memory once it has so shoe-horn into virtual, and how even simple applications can get bogged down with multiple instances, is what is of interest to me. Then, battery life on that.

I'm hopeful M1 will deliver. But a synthetic benchmark optimised for strange workloads where there are different codepaths depending on what CPU registers/functions are available, isn't going to cut it with me.
Posted on Reply
#12
Blueberries
I've never used Geekbench myself but that ST performance
Posted on Reply
#13
Dredi
Fourstaff
There are also people who like to say SPECint is not representative of real world usage.
And what are the usual reasons to say so?
The different subscores come from very relevant performance metrics, like compression/decompression, h264, code compiling etc.
Posted on Reply
#14
Fourstaff
Dredi
And what are the usual reasons to say so?
The different subscores come from very relevant performance metrics, like compression/decompression, h264, code compiling etc.
Real life benchmarks uses a mixture of workloads etc etc. Everyone have their preferred reason to disregard a certain benchmark, but once you have enough of them there is really not much left but to accept.
Posted on Reply
#15
Vya Domus
Dredi
And what are the usual reasons to say so?
The different subscores come from very relevant performance metrics, like compression/decompression, h264, code compiling etc.
Is code compilation relevant to most people ?
Posted on Reply
#16
Imsochobo
Fourstaff
There are also people who like to say SPECint is not representative of real world usage.
Yes, It's not.
It's a integer benchmark.

"designed to test exclusively the integer performance of the system. "

AVX, AVX2, AVX512 is not included at all.
Posted on Reply
#17
Vya Domus
Imsochobo
Yes, It's not.
It's a integer benchmark.

"designed to test exclusively the integer performance of the system. "

AVX, AVX2, AVX512 is not included at all.
Incidentally or not Apple's cores are known for featuring a very wide integer pipeline unlike most other CPUs out which usually have wider floating point pipelines. What's also interesting is that SPECint is Apple's darling for a very long time, they used to make comparisons based on it since PowerPC days.
Posted on Reply
#18
Dredi
Fourstaff
Real life benchmarks uses a mixture of workloads etc etc. Everyone have their preferred reason to disregard a certain benchmark, but once you have enough of them there is really not much left but to accept.
Vya Domus
Is code compilation relevant to most people ?
You can just dismiss the sub-benchmarks that you don’t find relevant and see who comes on top (it’s apple).
Imsochobo
Yes, It's not.
It's a integer benchmark.

"designed to test exclusively the integer performance of the system. "

AVX, AVX2, AVX512 is not included at all.
Then look at other benchmarks from the SPEC suite, like SPECfp, which can be compiled with whatever AVX support you wish.
Posted on Reply
#19
pat-roner
Say what you want about Geekbench, but I found this quote from Andrei interesting;

There’s been a lot of criticism about more common benchmark suites such as GeekBench, but frankly I've found these concerns or arguments to be quite unfounded. The only factual differences between workloads in SPEC and workloads in GB5 is that the latter has less outlier tests which are memory-heavy, meaning it’s more of a CPU benchmark whereas SPEC has more tendency towards CPU+DRAM.

I can only say that I'm really excited about the new M1 chips. As a lover of technology I'm all for new technology, and seeing people hate on a product they will never buy is beyond me.

PCMasterrace amirite?/s
Posted on Reply
#20
Vya Domus
Dredi
You can just dismiss the sub-benchmarks that you don’t find relevant
That has to be ironic, SPECint has absolutely nothing to do with the real world, it is purely a synthetic benchmark :


00.perlbenchCPerl Programming LanguageDerived from Perl V5.8.7. The workload includes SpamAssassin, MHonArc (an email indexer), and specdiff (SPEC's tool that checks benchmark outputs).
401.bzip2CCompressionJulian Seward's bzip2 version 1.0.3, modified to do most work in memory, rather than doing I/O.
403.gccCC CompilerBased on gcc Version 3.2, generates code for Opteron.
429.mcfCCombinatorial OptimizationVehicle scheduling. Uses a network simplex algorithm (which is also used in commercial products) to schedule public transport.
445.gobmkCArtificial Intelligence: go playingPlays the game of Go, a simply described but deeply complex game.
456.hmmerCSearch Gene SequenceProtein sequence analysis using profile hidden Markov models (profile HMMs)
458.sjengCArtificial Intelligence: chess playingA highly-ranked chess program that also plays several chess variants.
462.libquantumCPhysics: Quantum ComputingSimulates a quantum computer, running Shor's polynomial-time factorization algorithm.
464.h264refCVideo CompressionA reference implementation of H.264/AVC, encodes a videostream using 2 parameter sets. The H.264/AVC standard is expected to replace MPEG2
471.omnetppC++Discrete Event SimulationUses the OMNet++ discrete event simulator to model a large Ethernet campus network.
473.astarC++Path-finding AlgorithmsPathfinding library for 2D maps, including the well known A* algorithm.
483.xalancbmkC++XML Processing


Nearly all of those are completely irrelevant to real world workloads performed by your average user. Even for me, someone who knows what all of those tests entail it still means nothing.
pat-roner
and seeing people hate on a product they will never buy is beyond me.
Huh ? Wouldn't it be strange if I were to hate on something that I will buy ? :kookoo:
Posted on Reply
#21
pat-roner
Vya Domus
Huh ? Wouldn't it be strange if I were to hate on something that I will buy ? :kookoo:
You are hating on something brand new. Never before has something like this been released with so tight integration of SW and HW with ARM.
For all we know hating on this, might be like hating on the iPhone when it was released.
Posted on Reply
#22
Flanker
I guess the only benchmark to trust is the one that actually does something we use our computers for. e.g. something we does a lot of video encoding will want to see how quickly this CPU encodes a 1 hr long video
Posted on Reply
#23
Vya Domus
pat-roner
You are hating on something brand new. Never before has something like this been released with so tight integration of SW and HW with ARM.
I don't know why you feel the need to defend Apple when I actually said nothing about them. I only spoke against Geekbench and offered a pretty clear example on why it should be disregarded completely since it generates discrepancies from version to version than can only be explained by tailoring the code for specific uarchs. And I don't need to say why that's a big no no if you want to pretended that a certain benchmark is a reliable way to compare processors.
Posted on Reply
#24
Punkenjoy
Well still, Apple would received way less hate if they were either sellings these chips whitout the full package and had a way better policy on "right to repair"

i don't hate device itself, altought i am neither a fan of IOS and Mac OS, but i really hate Apple as a company. Their reps are a bunch of arroguant pricks that think you don't need what you think you need...

These chips are certainly impressive, but real world usage test will be needed. How much it consume? does it throttle ? etc..
Posted on Reply
#25
Dredi
Vya Domus
That has to be ironic, SPECint has absolutely nothing to do with the real world, it is purely a synthetic benchmark :


00.perlbenchCPerl Programming LanguageDerived from Perl V5.8.7. The workload includes SpamAssassin, MHonArc (an email indexer), and specdiff (SPEC's tool that checks benchmark outputs).
401.bzip2CCompressionJulian Seward's bzip2 version 1.0.3, modified to do most work in memory, rather than doing I/O.
403.gccCC CompilerBased on gcc Version 3.2, generates code for Opteron.
429.mcfCCombinatorial OptimizationVehicle scheduling. Uses a network simplex algorithm (which is also used in commercial products) to schedule public transport.
445.gobmkCArtificial Intelligence: go playingPlays the game of Go, a simply described but deeply complex game.
456.hmmerCSearch Gene SequenceProtein sequence analysis using profile hidden Markov models (profile HMMs)
458.sjengCArtificial Intelligence: chess playingA highly-ranked chess program that also plays several chess variants.
462.libquantumCPhysics: Quantum ComputingSimulates a quantum computer, running Shor's polynomial-time factorization algorithm.
464.h264refCVideo CompressionA reference implementation of H.264/AVC, encodes a videostream using 2 parameter sets. The H.264/AVC standard is expected to replace MPEG2
471.omnetppC++Discrete Event SimulationUses the OMNet++ discrete event simulator to model a large Ethernet campus network.
473.astarC++Path-finding AlgorithmsPathfinding library for 2D maps, including the well known A* algorithm.
483.xalancbmkC++XML Processing

Nearly all of those are completely irrelevant to real world workloads performed by your average user. Even for me, someone who knows what all of those tests entail it still means nothing.




Huh ? Wouldn't it be strange if I were to hate on something that I will buy ? :kookoo:
Uh, ok. What would you deem relevant? Boot up time? How long does it take to launch candy crush?

Please elaborate.
Posted on Reply
Add your own comment