Apple M1 ARM Performance With A 2020 Mac Mini

[pinoli]

This is my two cent on this whole thing.
DISCLAIMER: I have used Linux, Windows and macOS indifferently during my life. I used a MBP for at least 4 years as my main machine, but today I daily drive Windows. Also all my remote machines run on Linux.

I believe that the timing of the M1 has been carefully selected to inflict maximum damage to Intel and to dramatize the performance claims.
After browsing through all the benchmarks here and on other sites, I can easily say that the M1 is a good chip, a very good chip indeed.
At the same time I can say that Intel still has the lead in CPU performance for ultrabook with its 11th gen lineup (i7-1165G7/i7-1185G7), at least in many different workload scenarios.
Something to bear in mind here, is that Apple went all in and exploited the heck out of the new TSMC 5nm process, while the aforementioned Intel 11th gen CPUs are on the 10nm node (yes, I know these numbers are totally meaningless, but at the same time we indeed know that TSMC has a edge on Intel with its latest node).

At this time Intel is behind, it's having a lot of yield issues with its 10nm node. But it sill manages to take the lead, even in this dire situation.
Also, I believe that next iteration of M1 will not see the same jump we saw in respect to previous Intel-based Macs, because they already exploited the 5nm goodness a.k.a. lower power, higher frequencies, etc... Now that Apple is in TSMC yard, when Intel catches up, the only differentiator will still be ARM vs x86, so an architectural difference.

And finally, architecturally x86 and ARM serve very different purposes: yes, M1 will do for amazing ultrabooks, but I would like to see them used for AI/ML training now that don't even support eGPUs (and I think it has to do with limited PCie lanes implemented on the SoC). x86 will still win for big complex, power hungry tasks. Could there be comparison in performance per watts between the two architectures? Nope, ARM will always lead in this compartment.

And this is indeed how it's going: ARM CPUs are deployed on every low power device on earth, but the "low power" here is crucial.
Where power is not an issue, ARM does not scale as well as x86, as we can easily see with current performance (please don't mention Fugaku).

So what I will be EXTREMELY interested to see, is how Apple manage to scale these performances to iMacs and beefier Macs.
Also, I look forward to see NEXT YEAR performance for M2. From these numbers we could actually gather a lot of information on the trajectory Apple is on.

For now, at least to my eyes, the hype around this product is not justified.
The MacBook Air, MacBook Pro and MacMini all share the same CPU, and yet we have a "Pro" computer that differentiates only for a fan, a touchbar and two hour more of battery life. What's "Pro" about it?
All these macs took a step backward on IO ports: MacMini lost 10GBe option, very useful in many "Pro" scenarios. eGPUs will not work on any of these Macs, and I find this unacceptable (imagine all the people who shelled out in order to get a nice home/office setup... money wasted).
Also, Apple still sells the base models with 8GB of RAM and 256GB SSD for a very high price, a dubious choice to say the least.

These will do fine ultrabooks, but guess what: the iPad is cheaper, comparably faster, lighter, battery is insane, and you can basically do the same things you would do on a MacBookAir. I own an iPad pro and I love it, but I just opened Task Manager on Windows and memory utilization is at 26GB. How on earth am I going to get the same experience and performance on an iPad and consequently on a MacBook Air/MacBook Pro?

All of this is just to say that Apple, for me, is exiting the "Pro" market: I will be surprised if they could compete with Xeons or EPYC CPUs on the high end, for extreme workloads, something "pros", whatever that means, usually need. And now a single fan is what separates "light workloads" (as the MBA is maketed) from "professional workloads", I find this very amusing and sad at the same time.

[Jumbotron]

You need to relax your "never, never, never" mantra dude.
That sounds like a true fanatic.
And don't worry. Even your beloved Apple relies on the shoulders of other men.

i have never...never...never...not in a million years ever....owned ANY Apple product.

Prove me wrong. You do know we have had 50 years of personal computing data to back my claim ?

[ultimA]

blacknova writes...

So, the main question is - Will it run Linux?
As nice as OSX UI is, I'm not sure I'm up to running OSX. Well, that, and the first generation is sorely lacking on RAM department and general extensibility.


Answer....who cares and what's the point? A lot of the M1's functionality would be useless to Linux. The Neural Processor, Matrix Math processor, the Visual Processor, the HSA like High Bandwidth Memory architecture, etc. Linux is COMPLETELY unoptimized even if you could hack a working unit. There would be no software optimized for it either. Could GIMP take advantage of the M1 capabilities above and beyond any x86 version? What about Openshot? Libreoffice? Ardour? The answer is clearly no.

The only thing you would learn about Linux on M1 is that you could even do it in the first place as a hacker's exercise and how fast native Linux binaries could run on just the plain cores in the M1 SoC. That wouldn't CERTAINLY not give the whole or the holistic picture of what could be achieved with Linux and Linux apps on M1.

That's actually all very wrong. Most of the software (if not all) Michael just tested don't use those specialized features yet, as these software have not yet been optimized for the M1 either. And please don't tell us you were talking about kernel support because that doesn't change a thing. You cannot take software such as transcoding tools, PHPbench, C-Ray, or 7-zip etc. and make them use special silicon features simply by swapping out the kernel behind it. Kernel support may be needed, is surely not enough alone. And even Apple itself admitted that optimizing their software for the M1 will take years of effort, starting now, meaning the current software stack is still missing many-many optimizations for the M1.

With other words, there is a lot of sense and reward in making Linux run on this hardware, because the performance advantage we see in Michael's benchmarks is mostly just the unoptimized baseline already, and the optimized best-case should be even better.

[Grinness]

Non of which....image magic nor python....are native to nor optimized for M1

So what?

What do you want to compare?
Apple vs Pears?

The CPUs are not from the same year (as belong to completely different fabrication process and design iteration), do not have the same number of cores, nor the same number of threads, do not thermally throttle in the same way ....
To be honest this is a clusterf**k of comparison, if you want to be objective.

If you want to say Great Apple, let's do it all together and p(r)ay your new overlord

Amen

[blacknova]

Well. Yes. But you spend resources doing it. There is no magic in optimizing non-realtime stuff.
But this is as close as realtime as it gets, uses resources and sometimes gets away mere inches from native?
Are you telling me the native compiler is shittier than the x86->arm64 translation in terms of efficiency? Like massively so?

Consider this, normally on x86 you have native compiler->x86 ISA code->x86 CPU frontend CISC to RISC translator->internal RISC core execution path.
On new macs instead of CPU's CISC->RISC translator you have Rosetta.

Both CPU's frontend and Rosetta get optimized x86 code at their input. What differ is that they operate under different constraints, Rosetta can act as both AOT and JIT compiler and provided hot-spot optimization. Native translator run only in sequential mode and do not provide any additional optimizations.

So yeah, Rosetta can potentially run code at least as fast as on native processor, but in some cases it can provide better optimization than native front-end translator.

[MadeUpName]

That's why Apple has now and will forever more have the most optimized, powerful, and powerfully efficient software and hardware. The best the Windows, Google and Linux world can hope for and DEVELOP for is for ARM with it's native ecosystem of in die compute models such as in die DSP's, in die Neural Net Processors, in die Visual Processors, in die Advanced Matrix Math Processors, in die Storage Processors, in die FPGA's, etc. NOT ONE x86 CPU has ANY of these. NOT....ONE. Well...you could include the Trusted Zone ARM cores in AMD CPU's....but...really....is that helping your encoding ??

There is truth in what you say but don't sleep on AMD and Intel yet. AMD in particular has been making moves similar to what Apple has been with things like going to the infinity cache and moving early on BAR. AMD produces high end video cards and CPUs and makes APUs so further integration of those components like Apple has done isn't much of a stretch. But the big move was buying Xilinx. The potential for what could come from that is incredible, but we are probably 3-5 years away from seeing the benefits of that in consumer hardware. In the datacenter AMD is going to step on Intels air hose again with the Xilinx purchase.

Intel I am not as enthused about but they are finally realising that every one who makes a CPU is going to have to have serious graphics capability to go along with it. They do some interesting things that others don't like Optain. But I think they are going to have to buy some one to stay in the game.

Apple is going to face incredible pressure in about 5 years from the NVidia marriage with ARM. NVidia should be able to produce hardware that Apple can only dream about.

[Jumbotron]

I find it fascinating that Rosetta can be that fast.
This is instruction translation. And even so it's more than half the speed of native.

Either all x86 compilers are absolute rubbish, all or Rosetta is some alien magic sauce.

Other than that I'd want some more hard data compared to competing technology.
Transistors, caches, frequency, process technology, power budget.

16 Billion transistors

For the 4 High Performance ARM cores you have 320 KB of L1 Cache divided as such.... 192 KB of L1 Instruction 128 KB of L1 Data
12 MB of L2 Cache shared by the 4 Hi-Per Cores.

In addition to that cache setup for the high performance cores, there is a separate 192 KB of L1 Cache for the 4 additional Low Power Cores divided as such
128 KB Instruction and 64 KB Data. The 4 Low Power cores also share a 4MB L2 Cache.

So...altogether between the 8 cores making up the M1 SoC you have 2 MB of L1 Cache and a combined 16 MB of L2 Cache.

Incidentally, Anandtech found evidence during their test of the earlier A13 Apple SoC that Apple seems to use the iPhones Flash storage as a type of L3 cache when needed. I could not find any evidence of that with the M1 even though it is based on the upgrade of the A13 called the A14 or in the M1's extended performace case, the A14X. I'm not saying Apple doesn't use storage as a type of L3 cache in the Macbook. Just saying I can't find anyone writing about this.

Frequency of the Cores. The Hi-Perf cores run up to 3.2 GHz while the Low Power Cores runs up to 2 Ghz. The Thermals are 13.8 Watts for the Hi Performance cores and 1.3 watts for the Low Power Cores.

Process for the entire SoC is 5nm. It is the world's first laptop chip and desktop chip processed at 5nm. The A13 for the iPhone 11 last year was the world's first mobile 5nm chip.

The M1 GPU outperforms the Nvidia GTX 1050 Ti and destroys the Radeon RX 560. It is also just behind the Nvidia GTX 1650.

Finally, the M1 is fanless in the Macbook. Remember that and the thermals above when looking at it specs and how it is faster in Single Thread benchmarks than any Intel CPU including the i9 and is only bested by the newest AMD Ryzen 3 CPU's that have just been released.

Ohhh....BTW...16 GB of HBM LPDDR 4X RAM feeds every core on an AMD HSA like interconnect and architecture. This RAM is indeed HBM like found on certain AMD Radeon GPU cards, although it is NOT on die of the SoC but built into the SIP right beside the SoC diriectly tied with high speed interconnects....probably CCIX which is the ARM backed zero copy, cache coherent interconnect tech similar to Intel's CXL and AMD's Infinity Fabric. This allows Apple to achieve what AMD abandoned, which is HSA.

[mroche]

Flame away.

You’re going to give your money to Autodesk? Bad move

Cheers,
Mike

[Jumbotron]

There is truth in what you say but don't sleep on AMD and Intel yet. AMD in particular has been making moves similar to what Apple has been with things like going to the infinity cache and moving early on BAR. AMD produces high end video cards and CPUs and makes APUs so further integration of those components like Apple has done isn't much of a stretch. But the big move was buying Xilinx. The potential for what could come from that is incredible, but we are probably 3-5 years away from seeing the benefits of that in consumer hardware. In the datacenter AMD is going to step on Intels air hose again with the Xilinx purchase.

Intel I am not as enthused about but they are finally realising that every one who makes a CPU is going to have to have serious graphics capability to go along with it. They do some interesting things that others don't like Optain. But I think they are going to have to buy some one to stay in the game.

Apple is going to face incredible pressure in about 5 years from the NVidia marriage with ARM. NVidia should be able to produce hardware that Apple can only dream about.

Name me one Intel SoC or AMD SoC, even the PS 5 SoC that can do what the M1 does. All at 5nm. And at 13.8 watts max thermal?

And Nvidia.....LOL. Really? They'll just continue to license, at no charge perhaps as they'll own it, ARM designs and then completely blow out the thermals.

On the other hand ARM IS stepping up its game for licensees who don't want to spend all that money making custom high performance ARM SoCs like Qualcomm and Apple. The Neoverse initiative is just that. But everyone will be permanently behind Apple, on the consumer side that is, simply because compute performance has NEVER been just about the raw compute power of a CPU. It's all the way down the software stack from apps to compilers to kernel.

[MadeUpName]

You’re going to give your money to Autodesk? Bad move

Cheers,
Mike

If I was going to give my money to Autodesk it would indeed be a bad move.