After uninstalling CUDA toolkit still MATLAB responds for gpuDevice() !!

Hi

Accidently I installed CUDA 10 in ubuntu 16.04.

Later on I came to know that matlab supports CUDA 9.1.

So i unistalled CUDA from ubuntu. But still it responds to gpuDevice.

I am raising this issue because I am worried if I install CUDA 9.1, MATLAB may not be able to detect new CUDA.

Is there any environment variable or something, I have to change.

Answer 1

[Edric Ellis]

MATLAB and Parallel Computing Toolbox need the CUDA driver to be installed on your system, not the CUDA toolkit.

I admit it is very easy to get confused about the distinction between these two components, let me try to explain.

The CUDA driver is a frequently-updated low-level library that connects your GPU hardware into the rest of your computer. It provides the lowest-level support for your GPU - for example, it provides the ability to allocate memory on the GPU, and to launch kernels. However, the driver does not contain any specific algorithms - for example, it does not contain an FFT implementation.

Any given CUDA driver supports all CUDA versions up to and including the version current at the time of release of that driver. It is almost always the right thing to do to install the latest available CUDA driver from NVIDIA, since newer drivers are compatibile with older applications that use CUDA. (There are rare exceptions when this compatibility is inadvertently broken)

The CUDA toolkit comprises a CUDA compiler (nvcc), and support libraries for FFT and BLAS (MTIMES etc.). Any given CUDA application (such as MATLAB) is built using a specific version of the CUDA toolkit. In the case of MATLAB, we ship with MATLAB the portions of the CUDA toolkit that are needed for execution of gpuArray code. For example, we ship the cufft and cublas libraries. You can see these in the <matlab>/bin/<arch> directory.

Putting all this together - for most usage of gpuArray, you do not need to install any version of the CUDA toolkit, since MATLAB already brings along the pieces it needs. However, we cannot bring along the CUDA driver, so you must install that - and the only requirement there is that the driver is at least as recent as the version of the toolkit shipping in the version of MATLAB you're using.

The time you need to install your own CUDA toolkit is when you want to build CUDA sources to use within MATLAB, either using parallel.gpu.CUDAKernel or mexcuda. In that case, to ensure correct binary compatibility, it is important that you install precisely the same version of the CUDA toolkit as is being used by MATLAB. You can see this version information by looking at the ToolkitVersion property of the gpuDevice object. (The DriverVersion of your gpuDevice tells you the latest version of the CUDA toolkit supported by the CUDA driver installed on your machine.)

One final note: NVIDIA have a page about compatilibity of drivers and toolkits here.

[Tish Sheridan]

Also - if you're looking for toolkit versions, see this: MATLAB GPU Support by Release.

[JAI PRAKASH]

Wow, a very useful info.

Thanks Edric

I have some fundamental questions still

1. How can I install only CUDA drivers not toolkit?
2. For GPU coder mex creation, I think it needs nvcc support. Would only CUDA drivers are sufficient for mex creation?
3. As I said in my last post, I uninstalled everyting with below commands following uninstall guidelines. Then also gpuDevice() shows toolkit and driver version in MATLAB? So I doubt there is something still to do, to remove any footprints of CUDA from ubunutu 16.04.

sudo /usr/bin/nvidia-uninstall

sudo apt-get --purge remove <all cuda packages> (one by one)

[Edric Ellis]

1. I'm afraid I'm not that familiar with Linux package managers, so I can't really comment on that sort of thing. I think you can get just a driver alone from this page. I think the toolkits often bundle in a "developer driver" too, but you can use a release driver.
2. The GPU Coder documentation states that you need to install a CUDA toolkit. Note that it's generally fine to have multiple toolkits installed on your system - but you only ever have one driver. I would definitely recommend installing the toolkit corresponding to the toolkit reported by gpuDevice.
3. Yes, the gpuDevice reported DriverVersion reports the most recent version of CUDA supported by the driver. The ToolkitVersion reports the version of the CUDA toolkit with which MATLAB was built, and which MATLAB ships the libraries for.

[Joss Knight]

Linux drivers are distributed (on that page that Edric linked) as a runfile (that you just run from the command line) if you select "Linux 64 bit", and as a .deb or other if you select a specific Linux distribution.

[JAI PRAKASH]

After setting required environment variables it is working all good.

https://www.mathworks.com/help/gpucoder/gs/setting-up-the-toolchain.html

Thanks for your support.