may I use R2019b with CUDA 9.1

Table 1. CUDA Toolkit and Compatible Driver Versions

CUDA Toolkit Windows x86_64 Driver Version

CUDA 10.2.89 >= 441.22

CUDA 10.1 (10.1.105 general release, and updates) >= 418.96

CUDA 10.0.130 >= 411.31

CUDA 9.2 (9.2.148 Update 1) >= 398.26

CUDA 9.2 (9.2.88) >= 397.44

CUDA 9.1 (9.1.85) >= 391.29

CUDA 9.0 (9.0.76) >= 385.54

CUDA 8.0 (8.0.61 GA2) >= 376.51

CUDA 8.0 (8.0.44) >= 369.30

CUDA 7.5 (7.5.16) >= 353.66

CUDA 7.0 (7.0.28 >= 347.62

For my GeForce GT540M the last driver version is 391.35 so I should use CUDA Toolkit 9.1. Is Matlab R2019b compatible with this CUDA Toolkit version?

Answer 1

[Marcantonio Sommacal]

Then just to resume in other words, with the GeForce GT540M I should use Matlab R2017b with CUDA Toolkit 8.0

[Jason Ross]

That is correct -- there is no support for Fermi cards beyond that toolkit version / MATLAB release.

Answer 2

[Jason Ross]

You can generally upgrade a driver to a newer version, the CUDA toolkit a newer driver supports is backwards-compatible with previous versions.

It doesn't work the other way -- if your driver is too old and you are trying to use a MATLAB that has a dependency on a newer toolkit version, an error will be produced telling you to update your driver. Looking at this page it looks like you will get that error. I confirmed on the nvidia page that the latest driver only supports CUDA 9.1.

Also note that you only need to install the driver to have MATLAB use CUDA. Installations of the toolkit and SDK are not required.

Answer 3

[Marcantonio Sommacal]

I went to the page "GPU Support by Release" you suggested. In corrispondance of my Matlab Release (R2019b) it seems clear that there is no support for Fermi gpu architecture which is that of the GeForce GT540M installed in my laptop. So what about backwards compatibility?

[Jason Ross]

nVidia provides backwards compatibility for older toolkit drivers in newer drivers. So this provides backwards compatibility for software developed using older toolkits with newer drivers. This implicitly assumes that the hardware is also supported by nVidia. For example, if you had a Titan V GPU, you can upgrade to a newer driver (which as of this reply supports 10.2) and older releases of MATLAB using earlier releases should continue to function correctly.

In the case of the GPU in your laptop, Fermi entered "legacy" status in April 2018 (meaning no enhancements, e.g. support for newer toolkit versions) and the legacy support phase ended this month. This means that Fermi GPUs are unsupported by newer toolkit versions.

Answer 4

[Marcantonio Sommacal]

I am licensed for Matlab R2019b and Toolbox for parallel processing and partial differential equation. Is there the possibility of having Matlab R2017b with some discounts with the same toolboxes?

[Jason Ross]

You would need to contact Sales / Customer Service to get an answer to that.

[Walter Roberson]

If you are already licensed for R2019b then you are automatically authorized to use any previous release under the same license. Just download and install and activate.

[Marcantonio Sommacal]

Thank you

[Fadi Alsuhimat]

I use 2018 & 2016 matlab, I have same problem,

with Intel HD graphic 3000 what I have to do?

[Walter Roberson]

Intel HD 3000 graphics is not a GPU, and cannot be used for GPU processing. Also, MATLAB only supports NVIDIA cards.

If you have a Fermi series NVIDIA card, then it will work up to R2017b. If you are not using a student license, then a license for any newer release of MATLAB (such as R2018a) entitles you to use earlier releases (such as R2017b.)

(Well, student licenses do as well, but students often cannot get access to the older release.)