Apple's M1 native support

Are there any updates on when will Apple's M1 chip be natively supported by Matlab?

Any chance this could be on realease 2021b?

I had to go back to an Intel based Mac due the frequent crashes (3 to 4 times a day) experienced while running Matlab 2021a on a 16 GB 8-core M1 Macbook Pro with Big Sur.

I understand that the native version is under development, but it's been months since so.

Could someone please provide some insights on this? It's really causing some problems to some of us ...

Thanks in advance.

[Truman]

Okay Mathworks - what's the story. When wil M1 native support be relased. Mathematica has had native M1 support from day one. All major image editors have native M1 support. Where is Mathworks on this. Will 2021b provide native support to M1? Matlab along with the toolboxes are not inepensive products. Time to give the users a schedue - don't you think.

[Rik]

I don't really see the use of complaining here. Contact support instead.

Anyone who knows will be under an NDA, so the only people who can comment are those who don't know.

From what I understand, Rosetta2 offers most of the performance already. Frequent crashes can probably be fixed by updates (R2021a u4 is available since this week(?), you could check the changelog).

[Walter Roberson]

As I wrote in https://www.mathworks.com/matlabcentral/answers/641925-is-matlab-supported-on-apple-silicon-macs#comment_1547845

I expect R2023b for the fully native version.

That's my estimate. Taking into account Polyspace, which has quite different interal processes.

[Walter Roberson]

@Truman

What is your evidence for the claim that Mathematica has had native M1 support "from day one" ?

According to https://doesitarm.com/app/wolfram-mathematica/ the first Mathematica release with native Mathematica support was 12.3.1, which checks with the Mathematica release notes https://www.wolfram.com/mathematica/quick-revision-history.html which announced native M1 support for 12.3.1 in July 2021.

Wolfram has a chart showing what parts of Mathematica 12.0* and 12.1* work with M1, but that was under Rosetta 2. For example they indicate 12.0 was supported, but 12.0 was released in 2019, which was before engineering samples of the M1 were available to the public, so it would have been very difficult for the M1 to have been supported "natively" in the 2019 time frame.

[Royi Avital]

I think the challange here would be supporting a new and different BLAS / LAPACK library. But once it is done, if the API is general enough (Like a trampoline approach) then it means the support for Ryzen CPU's can be better as well (By using AMD's BLIS instead of MKL).

[Hauke Fath]

BLAS is pluggable already -- setting 'BLAS_VERSION=/usr/lib/libblas.so' in the evironment will use the system's (Arch Linux) OpenBLAS instead of what Mathworks ships on our Ryzen machines.

Answer 1

[Selahattin Imrohoroglu]

I have not run any of my Matlab code yet but executing bench gives the following for my MacBook Pro M1 Max with 64GB memory:

[Michael Brown]

Two points:

1. I'm impressed with your graphics (2D 3D) numbers. Why are they so much better than what I am seeing on my mac?
2. These bench tests do not really show the strengths of the new mac architecture. I use a stress test of a large-scale inverse problem (millions of data and tens of thousands of parameters). I can push problems beyond limits of physical memory. With intel macs pushed to use the matlab-induced swap space of 10s of GB, the intel CPU would be slowed to 20-100% rather than 600% with all cores engaged. I could get the calculation to complete if I waited hours. On the M1 macbook pro the same problem slows to 400-500% rather than running at 800-900% (as it does for smaller calculations) and completes in 5 minutes. Having the memory bandwidth of 200 GB/s and SSD access at 5000-6000 MB/s makes an enormous difference (compared to 20 GB/s for DDRM4 ram and 2000MB/s for older SSD's)

[Selahattin Imrohoroglu]

Sorry for the late response. I guess the 2D and 3D numbers are better due to the newer, faster and multi-core GPUs?

I ran similar code with my 2018 MacBook Pro (2.9 GHz 6-core Intel i9, 32 GB memory), the new iMac (Apple silion, 32 GB memory) and the new M1Max MacBook Pro with 64 GB memory, and I get scores within 20-30% of each other, except for graphics output. I guess the real difference would come from very demanding and memory intensive processes.

[Peter Somhorst]

Performance is not the real issue. Regular crashes are. I'm giving a course using Simulink in three weeks, and I'm consider dusting off my old 2015 MacBook Pro because 2021b is not stable enough on my M1 MacBook Pro.

[Robin Thomas]

I can't run Matlab at all on my MacBook Air with the M1 chip. It simply crashes too often.

[Kadir Gunel]

Matlab still does not work on M1 chips natively. It first transforms the code from x86 via Rosetta. So, even you are happy with the quickness, expect to have more.

@Robin Thomas Did you update the Mac Os ? I am running flowlessly on M1.

[Rodrigo Cerda]

I've been running Matalab 2021b (Update 3 - Feb 2022) for a couple of weeks now, natively on my MB Pro M1. That is, without installing Rosetta. It works better than when I was using Rosetta. It's faster and I haven't had any crashes so far.

[Jorge Ignacio Cisneros Saldana]

Have you tried parallel processing? does it work?

Answer 2

[Oliver Woodford]

MATLAB R2022a still doesn't run natively on Apple silicon. The release notes say "On Apple silicon Macs, MATLAB runs using the Rosetta 2 environment." and "A future release of MATLAB will run natively on Apple silicon."

[Truman]

Yep, Mathworks seems to be a day late and a dollar short. Just about every other serious application from image editing S/W such as Adobe and Capture One to Mathematica has been supporting Apple Silicon natively. Mathworks is still making promises and not delivering. Pretty sad, actually.

[Rik]

This may sound like a snide remark, but I intend it as a serious question: why do you care?

Shouldn't you only be concerned with the actual performance? I understand the performance of Rosetta 2 is very good, so the performance you get now should be mostly fine. Mathworks is making promises that they're working on it. Did you expect a day 1 release?

[Andrew Kacmarcik]

At this point for me it is less about the "performance" of Rosetta 2 vs. Native and more about the stability and reliability of the program. I still am running into issues with copy and paste just not working sometimes on an M1 Pro and that can be a pretty big annoyance regardless of how fast the program actually runs.

[Rik]

I have that as well, and I pretty much only work on Windows. So native vs Rosetta 2 will probably not change that (although I haven't had it so far on R2022a, but it's still early days).

[Peter Somhorst]

@Rik, your comment is not very helpful. No, people didn't expect a day 1 release, but we're at day 496 now and as was pointed out: plenty professional tools have moved to native support, many of which within the first 100 days. Rosetta 2 is good, but it's far from native. Many projects have time as a bottleneck, so every bit of performance counts. Stability is another concern, as stated before. Besides, people don't pay hefty licenses fees for inefficient software while the alternatives have had native support for more than a year.

[Spencer Kraisler]

@Rik Not only would a native build be faster, it would allow future matlab and simulink updates to take further advantage of apple silicon. Franky I think Mathworks has taken too long to make a native build. Matlab is paid software, so I do expect really amazing performance. Although Rosetta 2 is great, I am getting sub-intel i7 performance on an m1 max. :(

[Rodrigo Cerda]

I totally agree with Peter and Spencer. It's been way too long for Mathworks to release an M1 native version. And we don't even have a clear roadmap or estimated date for it. The only official statement is that is currently under development. A little more empathy with all of us would be greatly appreciatted ...

[Rik]

They just published a native beta for R2022a, so you can try that now.

[Rik]

@Vishnu Baldew Click on the link I posted, then follow the link on the official answer provided by Mathworks.

[Jonathan]

Mathematica was native quite a while ago. Day 1 isn't Day 1 for developers; they get access to hardware and compilers well before the actual hardware is released to users. I wonder if this is related to Mathworks' lazy decision to go with Java to make MATLAB multi-platform instead of just paying to write for each platform natively.

[Walter Roberson]

Jonathan, you would be mistaken. I named exact dates and linked evidence in https://www.mathworks.com/matlabcentral/answers/641925-is-matlab-supported-on-apple-silicon-macs#comment_1187918 and I recommend that you read the timeline there.

[Umut Sayin]

Hi,

I can assure you that 2022a still runs on Intel. There is no such thing as installing or not installing Rosetta, it is part of MacOS. It is literally written on the requirements page that it uses Rosetta 2

https://www.mathworks.com/support/requirements/matlab-mac.html

Answer 3

[Michael Brown]

I'm running R2021b on a new MB Pro M1 with 32 GB and macOS 12.01. So far MATLAB is simply running as intended and seems to handle large numerical inverses that push MATLAB to 29 GB of ram usage better than the prior intel MB Pro. The times required for calculations tend to be nearly a wash - within 20-30% - even with this non-native version. It has not crashed. Below are the results of the bench command (these are not as good as the intel MB Pro - but they are not terrible). I am eagerly awaiting the native version

[Rodrigo Cerda]

Hi Michael.

How long have you been running Matlab on your M1 MB Pro?

My major concern are the frequent crashes -several times a day- that I had on release 2021a when running it on my M1 MB Pro. Actually, I had to switch back to an older Intel MB Pro to avoid them.

Thanks and regards,

RC

[Michael Brown]

I've had the MB Pro M1 for two weeks and have been nearly continuously running scripts with linear/non-linear inverse problems and lots of 2D-3D plotting. I typically run problems that take ~5 minutes each and have 800-900% parallel utilization with near 30 GB of ram used by matlab. So far (using R2021b) matlab has not crashed once. Version R2021a on the intel mac had crash-inducing bugs. I was frustrated prior to the release of 2021b which was much better on the intel mac and seems stable on the M1 mac

Answer 4

[Shaohui Liu]

I am running 2021b on a MBP with M1 Pro chip and 32G RAM. Regular jobs look fine but some Simulink-based tasks have some weird bugs.

[Rodrigo Cerda]

I've downloaded the latest update of R2021b and my M1 MB Pro hasn't crashed as it used to.

It seems like a much more stable release.

I'm performing large matrix operations (several millions of elements) with the Paralel Computing Toolbox and it's working just fine.

Thanks!

[Robin Thomas]

So folks recommend 2021B for the M1 chip?

[j paterson]

@Robin Thomas I haven't had any Matlab crashes with 2021B on a Macbook Air (M1 Big Sur). I've been using it for a few months. I would suggest you try it if you're on an earlier version

[bibin]

i tried 2022b ..it crashes and lag like 5 year old windows