is MATLAB support AMD RADEON?

My laptop has AMD RADEON graphics card,but when i am executing a CNN classification program ,requiring GPU, it is giving an error cuda driver can not be located since i doesn't have NVIDIA gpu .Please give me some solution of it .

Answer 1

[Walter Roberson]

The only solution is to change to a system with NVidia GPU.

There are two competing GPU interfaces: NVidia's proprietary system, and every else's OpenCL. Mathworks does not support OpenCL at all.

[Joss Knight]

It's not quite as simple as this. AMD use their own version of OpenCL, which isn't compatible with 'everyone else'. Each manufacturer has their own set of supported operations and their own compilation pipeline. So really, MATLAB 'just' supports one of the manufacturers, with the most compute-capable suite of devices and most sophisticated library support.

[Walter Roberson]

Thanks for the clarification, Josh. I have not had reason to look at OpenCL myself.

[Jerry Wang]

OpenCL is just a collection of headers. Matlab doesn't have to implement the manufacturer-specific features. There is no reason Matlab shouldn't support OpenCL considering a core chunk of their customers run Apple hardware.

[Walter Roberson]

Because of the differences in OpenCL support, optimizing graphics performance may depend greatly on manufacturer and model of the device. Not just in a "some models are faster/slower than others" but rather "these manufacturers/models do not have the capability to do things this way and we have to take other computational paths."

My understand, by the way, is that the portion lf Mathworks customers who use Apple hardware is diminishing (but the ones on Apple are still often passionate about Apple.) With Apple getting out of OpenGL and switching to Metal, which is not used by any other vendor, and with persistent rumours that Apple is getting out of the desktop business, it is not obvious how much longer Mathworks can afford to support Apple (said as someone who uses Apple and often gets frustrated as heck on MS Windows.)

[LeChat]

Do you folks at Mathworks know the proportion of Mac vs Linux vs Windows users in the Matlab community? In any case, you should be careful, with Python 3 and Julia growing, there is a chance bad marketing choices may have pretty bad consequences... Please don’t let Mac or Linux users down. You’d loose lots of people in education and academics.

[Walter Roberson]

I saw some approximate figures recently. The number of Android users has exceeded the number of Mac and iOS users by a fair margin. "market forces" would suggest dropping Mac and switching the effort to native Android.

However completely different market forces are pushing for cloud implementation in which the local component is just a web based display interface with ability read local files and display streaming media, giving up on much of the Windows or Mac or Linux implementation in favor of whatever was most effective to run in the cloud (possibly whatever flavor of Unix is most supported by aws and azure)

Talking about "marketing forces" goes both ways, as Mathworks is not a small company and the users were tend to hear from are the minority of where their money comes from.

[Royi Avital]

@Joss Knight, The roots of MATLAB's GPU Toolbox are well known: Jacket by AccelerEyes. The same company has a library called ArrayFire . It has been available for many years and it offers great capabilities built on top of OpenCL. They work beautifully on AMD, nVidia and Intel.

I think this time you can do the even more right thing to do. Embrace this open library and support it by resources. All your users will be greatful for that.

As @LeChat wrote, the worse thing to do is not to move forward. There are new kids in the block. You need to move forward faster.

[Walter Roberson]

@Royi Avital

As best I recall, AccelerEyes had some good technology, that worked on AMD as well as NVIDIA, but that Mathworks was already working on their own product, and therefore it could not be said that the "roots" of the toolbox were with AccelerEyes.

My understand is that Joss will not be able to comment on the matter, due to whatever the terms were of the settlement of the lawsuit (which I did read, at some point.) In the past, when I asked what happened, the "No comment"'s that I got had that pointed "I'm not permitted to talk about this" tone.

[Joss Knight]

Funny that this keeps coming up!

I can assure you there isn't a single line of Jacket code in MATLAB, and never was, indeed it would be illegal for there to be. It would be illegal for Jacket to contain MathWorks code too.

If you feel you have to use an AMD card and you want to use a C++ interface then feel free to use ArrayFire. MathWorks makes a very big product and sadly doesn't have the resources to provide every feature any customer has ever asked for. Not that I'm ruling anything out in future versions of course.

[Royi Avital]

@Joss Knight, OK. That's strange. The timeline between the purchase of Jacket and the introduction of the GPU support implies differently. Moreover, have a look at this:

https://i.imgur.com/qpNREKY.png

It is openly said to be purchased for that purpose.

@Walter Roberson, The Jacket library had a CUDA code path which was what I believe was the interest of MAthWorks. Moreover, since the purchase they couldn't even offer that functionality for MATLAB. They indeed had also an OpenCL code path which I guess is basically the first versions of ArrayFire which is a great library.

If their code wasn't used, I guessed the purchase was made to remove a competitor for the Toolbox.

@Joss Knight, Adding support for AMD is a big interest of MathWorks. In today market, where nVidia got their prices very high people look for alternative. PyTorch has support (On Linux) for AMD. TensorFlow support is on its way. I guess you can lead or be lead.

[Joss Knight]

MathWorks did not buy Jacket. There is no Jacket code in MATLAB. MATLAB's GPU support was developed completely independently of Jacket.

I'm sorry there is no native support for AMD in MATLAB. I can point you to the MEX interface and the out-of-process Python interface to help.

[Walter Roberson]

@Royi Avital

https://www.plainsite.org/dockets/kqa8bc2q/illinois-northern-district-court/the-mathworks-inc-v-accelereyes-llc-et-al/

The lawsuit was filed by Mathworks on July 9, 2011 and settlement finished October 26, 2012. Mathworks released its GPU support in R2011b, official release date September 1, 2011 -- and so had the product in beta testing for several months at the time the lawsuit was filed.

AccelerEyes press release claiming cooperation between the companies was dated December 12, 2012. https://arrayfire.com/tag/jacket/

So if you are claiming that the "roots" of Mathwork's GPU support is AccelerEyes Jacket, then you would have to be claiming that Mathworks illicitly used Accelereye's code for more than a year before agreement was reached -- either that or your use of the word "roots" would have to mean only something vague such as "Mathworks saw someone else doing it and was inspired to do something similar."

My memory of what John Melonakos posted as his side of the story, involved AccelerEyes demonstrating the technology to Mathworks and negotiating for cooperation between the two companies -- but as best I recall, at the time of the technical demonstration, Mathworks already had some GPU technology that was not as advanced as AccelerEyes was. That if cooperation had been successfully negotiated, that it would have given Mathworks a substantial improvement over what they had at the time -- but that Mathworks did have something of their own at the time. But that all was a decade ago, and I cannot promise that my memory is accurate about the details.

[Royi Avital]

@Walter Roberson, Actually it doens't matter that much. I made the assertion based on what I read on ArrayFire blog. If they are not accurate, then I'm sorry for being mistaken.

The important thing is that ArrayFire shows it is possible to have a great GPU library without a GPU vendor lock in based on OpenCL. It is something we expect MathWorks to have as well in MATLAB.

[Joss Knight]

MATLAB's first release of GPU support was in R2010b released in September 2010.

[SL: adding link to archive.org capture of release notes on the MathWorks website]

[Joss Knight]

MathWorks did not buy Jacket.

https://www.jonesday.com/en/practices/experience/2012/12/mathworks-wins-copyright-and-patent-infringement-cases

[Walter Roberson]

@Joss Knight

Ah, then the table at https://www.mathworks.com/help/parallel-computing/gpu-support-by-release.html is incomplete; it ends at R2011b.

[Joss Knight]

Yes, I don't know why it doesn't go back to the start. Perhaps it's just because we don't want the table to get too big.

[Nicholas Alexander]

Any plans to include AMD GPUs in the future?

[Walter Roberson]

No, there are no plans to include AMD in the future.

GPU use is mostly driven by Deep Learning these days, but AMD use for Deep Learning is only roughly 3% of researchers if I recall correctly. The greatest share of Deep Learning research work by far is Nvidia, but according to a white paper I read, second place is IBM, and AMD is far behind. If Mathworks is to add additional GPU lines then IBM hardware would be a much more natural target.

Answer 2

[Michal]

MATLAB use for GPU computing CUDA technology, which is supported only by NVIDIA cards.

Answer 3

[mengmeng y]

Why does amd's computer graphics card not support the operation of matlab?

I am very anxious to use matlab? But the online tutorial is not right. Is there any way to use matlab under amd?If you have a solution, can you send it to me, thank you very much.

[John D'Errico]

Please stop making multiple answers when you just want to ask a question. Learn to use the comments instead. I've edited two of your answers into one, that never should have been an answer in the first place.

As far as using MATLAB itself, you can still use it regardless of the graphics card. It is just GPU computing that will be problematic.

[Walter Roberson]

MATLAB can work with any graphics card that supports OpenGL 2.1 or later, but prefers OpenGL 3.3 or later. There are, though, some graphics cards such as Intel HD1000 graphics cards, for which the graphics drivers are known to have sufficient bugs that they are not usable (HD1000 has not been produced for several years now.)

Using gpuArray or some of the Deep Learning programs is all that requires NVIDIA graphics cards.