Does Matlab perform well on AMD Ryzen?
Show older comments
Hi,
i want to buy a new CPU for Matlab but I am not sure if the new AMD Ryzen 1800X performs well on matlab. I know that the performance depends on the code you are running, but maybe there are some Ryzen owners out there who could post the results from the "bench" function.
Thanks. :)
3 Comments
Wes Johnson
on 4 Apr 2017
Edited: Wes Johnson
on 4 Apr 2017
I bought a system with the Ryzen 1800 processor (delivered on 3/30/17) for use with Simulink. Although the Ryzen has 8 cores and does well on the PassMark CPU test, it's slower than my i7-4790. A Google search shows that the Ryzen may have issues with the Windows 10 scheduler which limits is multi-threading abilities. I'm hopeful for a software patch...soon.
Mickaël Tits
on 22 Aug 2017
Hi, I'm thinking about buying a ryzen 7 too... Any news about a software patch so far ? Thanks!
Ned Flanders
on 20 Dec 2020
Edited: Ned Flanders
on 20 Dec 2020
Here's the result for an AMD Ryzen 5800x (integrated bench function - matlab 2020a)
Accepted Answer
Ned Flanders
on 19 Nov 2019
Edited: Ned Flanders
on 16 Sep 2020
AMD Ryzen/Threadripper CPUs perform excellent using Matlab since the 2020a Release, or if you perform a very simple tweak with the versions prior to 2020a.
Here is an example of several CPUs based on a matlab performance benchmark script you'll find in the link below.
As said, since Matlab 2020a, you'll have this performance out of the box. For Matlab versions prior to 2020a, please read the link below.
You should read it if you want to understand the background. If you ended up here because you are on an older Version on Matlab and have low performance with an AMD setup or you just want to apply the tweak, follow these steps: It will take less than a minute.
Solution 1 (Windows - no admin rights needed):
- Create a .bat file with the following lines to start Matlab in AVX2 Mode
@echo off
set MKL_DEBUG_CPU_TYPE=5
matlab.exe
This is straight forward. You open Notepad, copy and paste the above three lines and save the file as Matlab-AVX2. Notepad will save the file as Matlab-AVX2.txt. Now replace the extension ".txt" with ".bat".
If you double-click that file, Matlab will start the MKL in AVX2 Mode. If you start it the normal way, it will remain as always.
You can also download the .bat file from my highdrive if you trust me (which you of course should not, as I am a random guy in the Internet). If you delete the startup batch file provided in the download or the one you created yourself, its gone and your computer will be as it has been before.
Solution 2 (Windows - admin rights needed): If you are happy with the results (which you will be :-)), you should make the setting permanent by entering MKL_DEBUG_CPU_TYPE=5 into the System Environment Variables. This has several advantages, one of them being that it applies to all instances of Matlab and not just the one opened using the .bat file.
You can do this either by editing the Environmental Variables as shown above, or by opening a command prompt (CMD) with admin rights and typing in:
setx /M MKL_DEBUG_CPU_TYPE 5
Doing this will make the change permanent and available to ALL Programs using the MKL on your system until you delete the entry again from the variables.
LINUX: (Thanks to foreignrobot)
Simply type in a terminal:
export MKL_DEBUG_CPU_TYPE=5
and then run matlab from the same terminal.
For benchmarking, you can use this script:
Permanent solution for Linux:
echo 'export MKL_DEBUG_CPU_TYPE=5' >> ~/.profile
will apply the setting profile-wide, so you can launch it either through a terminal or the graphical launcher.
8 Comments
felicioharley
on 6 Jan 2020
Edited: felicioharley
on 6 Jan 2020
Hey Ned, this tweak (using the environment variable solution) will only improve Matlab or other softwares that use MKL (like CST electromagnetic simulation software, Numpy and Tensorflow) can benefit from it too? Thanks for sharing this.
Ned Flanders
on 6 Jan 2020
Hi fellicio,
setting the environmental variable applies to any software using the MKL.
Royi Avital
on 24 Jan 2020
It is a shame MathWorks, as one of the biggest customers of MKL doesn't ask Intel to stop this discrmination. I hope they will integrate this trick into next MATLAB or start supporting OpenBLAS until it matches and supress Intel MKL.
felicioharley
on 14 Feb 2020
I tested on an AMD Ryzen 5 3600. It works really well. Thanks again!
Michael Katz
on 28 Mar 2020
I ran the new "bench" on MATLAB R2020a, and the DEBUG workaround no longer appears to be necessary to get full performance on Ryzen and Threadripper. Has anyone else noticed this improvement in R2020a?
Ned Flanders
on 29 Mar 2020
Edited: Ned Flanders
on 29 Mar 2020
@Michael Katz
That's possible as Matlab intends to implement this workaround but as far as I am informed, it is not yet working in the current version of the 2020a release.
What is your result for 'getenv MKL_DEBUG_CPU_TYPE' in matlab and are you 100% that you haven't set the environmental variable or you are using the .bat file solution? Whats are your system specs?
Maybe we can help them troubleshooting the issue here.
Carlo Monjaraz
on 26 Jan 2021
Nice to see @Ned Flanders @Yair Altman in the same post. FYI: https://danieldk.eu/Posts/2020-08-31-MKL-Zen.html best greetings.
More Answers (3)
It depends if the used libraries call FMA3 commands. If so, the Ryzen crashes very fast.
The performance will depend on your programs. If writing to the disk is the bottleneck, the processor will not matter. If your code is parallelized, the number of cores rule. If memory is the limitation, buying a cheaper processor can mean, that you install more RAM. Then benchmarks of Matlab have a limited power only. Therefore I do not think that there is a general answer except for: The Ryzen is good!
As usual I mention, that you only have to wait some years to get a processor with the double performance, but writing efficient code can accelerate a program by a factor of 100 easily.
Robert Lee
on 10 May 2017
0 votes
Ryzen 1700 owner here, can't bench atm but I can update this later if you want one.
Ryzen is only worth it if you want to do parallel pool stuff (multi threaded), otherwise you may aswell get yourself a top end i3.
If you are wanting to do parallel pool stuff I'd be wary, I've had a lot of crashes doing parallel pool simulations with Ryzen that aren't occurring with intel CPUs - can't say with any certainty that Ryzen is to blame though.
2 Comments
Mickaël Tits
on 22 Aug 2017
Hello, I am thinking about buying a Ryzen 7 (1700, 1700x or 1800x). I wonder if you have some news about its use with Matlab ? Does it still crashes ? Did you do a benchmark ?
Georgi rozenman
on 2 Jul 2018
Edited: Walter Roberson
on 16 Sep 2020
All Ryzen 7 Cpus have weak 8 cores. If you'll be using MATLAB with general routines, without parallel pool, you'll be getting a worse performance than a slow core i3 since 1 core of Ryzen 7 is a lot weaker than 1 intel core.
Ryzen 7 performs well only when all 8 cores are used, therefore, it is very bad for MATLAB and any single core software.
Ned Flanders
on 21 Nov 2019
0 votes
This here is a pretty good and detailed article!
Categories
Find more on App Building in Help Center and File Exchange
on 20 Mar 2017
on 26 Jan 2021
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!
