CUDA seems not to be supported

My GPU is NVIDA GTX 1660, and I succeeded in updating my CUDA toolkit to 12.6, as confrimed by nvcc --version

C:\Users\Administrator>nvcc --version

nvcc: NVIDIA (R) Cuda compiler driver

Copyright (c) 2005-2024 NVIDIA Corporation

Built on Thu_Sep_12_02:55:00_Pacific_Daylight_Time_2024

Cuda compilation tools, release 12.6, V12.6.77

Build cuda_12.6.r12.6/compiler.34841621_0

Yet in MATLAB 2024a

>> gpuDevice

ans =

CUDADevice with properties:

Name: 'NVIDIA GeForce GTX 1660 Ti'

Index: 1

ComputeCapability: '7.5'

SupportsDouble: 1

GraphicsDriverVersion: '560.94'

DriverModel: 'WDDM'

ToolkitVersion: 12.199999999999999

MaxThreadsPerBlock: 1024

MaxShmemPerBlock: 49152 (49.15 KB)

MaxThreadBlockSize: [1024 1024 64]

MaxGridSize: [2.147483647000000e+09 65535 65535]

SIMDWidth: 32

TotalMemory: 6442123264 (6.44 GB)

AvailableMemory: 5378138112 (5.38 GB)

CachePolicy: 'balanced'

MultiprocessorCount: 24

ClockRateKHz: 1770000

ComputeMode: 'Default'

GPUOverlapsTransfers: 1

KernelExecutionTimeout: 1

CanMapHostMemory: 1

DeviceSupported: 1

DeviceAvailable: 1

DeviceSelected: 1

ToolkitVersion: 12.199999999999999 remains so. Is there anything I can do for MATLAB to recognise and use the latest toolkit? Cheers.

[Star Strider]

Contact Support

Answer 1

[Joss Knight]

MATLAB does not use the toolkit that you download, it installs its own version. You cannot change it by downloading the toolkit. What is it you are trying to do?

Answer 2

[Walter Roberson]

There is a difference between the toolkit version and the nvcc version. The toolkit version is essentially the driver version that is used all the time; the nvcc version is used only when you specifically ask to compile (for example you use mexcuda )

[Joss Knight]

Note: mexcuda uses the toolkit that MATLAB ships and does not use the version you download unless you specify MW_NVCC_PATH.

The toolkit does also come with a GPU driver but its main purpose is to provide libraries of algorithms and the compilation toolchain, all of which are shipped with your copy of MATLAB.

To install the latest driver you should find it on the NVIDIA driver downloads site, as explained on our GPU Computing Requirements page. You should not install the CUDA toolkit.

[Theron FARRELL]

Thanks for your demysification, Joss. My status quo is that I am going to generate GPU code for my neural network built and trained in MATLAB 2024a. It will then be integrated into my VC2022IDE to build a C++ executable. Other code in my VC2022 utilises CUDA 12.6 toolkit, hence will I have to change and specify MATLAB's toolkit version via MW_NVCC_PATH to CUDA 12.6 as well? If needed, how to do it? Cheers a lot.

[Joss Knight]

It sounds like you are mixing a GPU Coder code generation strategy with a PCT mexcuda build strategy? That no doubt makes things a bit muddled. Do you actually need to do it this way? Typically, generating code is not necessary within a MATLAB workflow; it is targeted at deployment to an embedded device. Any verification within MATLAB via MEX would typically be done through Coder apps and commands rather than through mexcuda (see for instance https://uk.mathworks.com/help/gpucoder/gs/gpucoder-verification.html)

If you need to use GPU Coder, perhaps because you are enhancing a C++ AI model with your own extensions in C++, then a number of options are available to you. But first - can you not recompile your existing code with CUDA 12.2?

To change your toolkit path to an unsupported version of the toolkit, you need to do some advanced things, so let's try something else first.

[Theron FARRELL]

Hi Joss, It is not that complex. I simply want to generate my neural network via GPU coder into a .h and lib (also .bin files that I suppose contain parameters of the network) that will be called into my VC project. Other libs such as VTK, ITK in the project are built on CUDA 12.6 toolkit, which will not be recompiled on CUDA 12.2. So my question is whether such .h and .lib generated on CUDA 12.2 will co-work with the rest of my project?

An additional question: in the generared folder for libs, such as codegen\lib\GenESNNCUDACode, there are many .h, .cpp etc, will it be sufficient for me to copy the ticked .h, .lib and .bin files as shown in the first picture, in my VC project to call the neural network, or will I need to copy all .h, .hpp, .cu, and .cpp files, including those under codegen\lib\GenESNNCUDACode\interface, which are shown in the second picture, as well? Could you offer me a general rule?

Many thanks.

[Joss Knight]

This is not my area of expertise, so I'm not sure what those files are for. I doubt there's any harm in including them (I doubt they take up much space, and they will be stripped if you are using static linking) or excluding them (you'll get a build error if you need them).

Your libraries built using CUDA 12.2 should work perfectly fine with the rest of your project. The question is whether the other-way-round will work, namely, trying to build an executable mex function using CUDA 12.2 which links against libraries built using CUDA 12.6. I think that will work, as in nvcc won't compain, although there is a small risk of runtime errors because minor versions promise to be backwards compatibly not forwards compatible (they are supposed to be binary compatible across all the 12.x libraries, but there's nothing to stop someone using a new 12.6 API in a library that isn't available at runtime because the version of, say, cublas loaded at runtime is the one MATLAB loads, the 12.2 one, which maybe doesn't have that API). Still, if you're having issues I'd just rebuild the rest of your project using CUDA 12.2.

If you want to change the behaviour of mex you can use mexcuda -v to find your options file, copy it, edit it to remove checks for CUDA 12.2, and then use the -f option to target the modified options file. But this is advanced. And regardless, you do incur some risk forcing libraries built against the CUDA 12.6 runtime to run alongside libraries built with the CUDA 12.2 runtime (i.e. MATLAB's libraries), so if you can just rebuild against 12.2 across the board your life is going to be simpler.

[Theron FARRELL]

Hi Joss, cheers a lot for your expounding everything. which is helpful and nice. Merry X'mas and Have a Happy 2025!