Error using gpu acceleration

I have trained network model in simulink, I am trying to perform GPU acceleration.

But, I am gettin the following error:

Value 'sm_86' is not defined for option 'gpu-architecture'

I have dowloaded cuda. Still, the error doesn't go

[Walter Roberson]

That error seems to be associated with the RTX 3060 or RTX 3080, and CUDA 11.0 or CUDA 11.1; in some contexts people are reporting success with CUDA 11.2 (but in other contexts it seems to still be a problem.)

Which CUDA do you have installed?

[Shaw]

I have tried on Cuda version 11.0 and 11.5 on RTX 3050Ti

[Walter Roberson]

https://githubmemory.com/repo/jeffheaton/docker-stylegan2-ada/issues/1

"ok going in circles but I believe the root of this issue is related to misalignment of cudatoolkit + cuDnn https://developer.nvidia.com/rdp/cudnn-download "

It looks like you might need to be part of the Developer's Program to download cuDnn (which is the CUDA Deep Learning Network)

[Joss Knight]

Which version of MATLAB are you using? I believe installing CUDA won't substitute for the version MATLAB is using unless you force it to. So presumably you are using a CUDA 10 version of MATLAB (R2020b or earlier)?

Edit: I see that CUDA 11.0 doesn't support compiling for cc8.6, so no released version of MATLAB will work. It seems you need support from the Coder team to determine how to prevent cc8.6 arch being used in the compiler options or alternatively to force Simulink to use your installed version of CUDA instead of the shipping version.

[Walter Roberson]

@Joss Knight According to the Release information they entered, R2021b.

Answer 1

[Hanumanth Hanumantharayappa]

The issue seems to be the same as the one captured in the below bug report. Please get the update mentioned in the below bug report and that should resolve the issue. Let me know if you face any problems.

https://www.mathworks.com/support/bugreports/2506291

[Shaw]

The bug fix seems to be for 2021a, I am using version 2021b

[Chao Luo]

This issue is GPU acceleration is using the CUDA toolkit shipped with MATLAB, and it does not support sm_86, which is picked up from properties of the GPU device.

Workaround:

Open configset, navigate to Simulation Target -> GPU Acceleration, specifiy Custom compute capability with '-arch=sm_80', or in MATLAB command window, set_param(modelName, 'SimGPUCustomComputeCapability', '-arch=sm_80');

[han wang]

Hi Chao,

Thanks for explanation. Your method works for me. I am using 2021B.

[Rik]

Comment posted as flag by @han wang:

GPU Simulink