CUDA simulation on Ada architecture - All CUDA-capable devices are busy

(4 posts) (2 voices)

f841r

Member
Joined: Feb '23
Posts: 3

Hey,

I'm trying to run the simulations in CUDA on a RTX 4090, but I always get the error "All CUDA-capable devices are busy or unavailable.", even though other things using CUDA (PyTorch, ...) work fine. Running the simulations with MATLAB GPU works. I also compiled the source code myself with sm_89 set (+ DYNAMIC linking), but the error persists. Does anybody know how to solve this?

Thanks for your help!
Fabian

General Infos:
GPU: RTX 4090
Nvidia drivers: 535.86.05
CUDA Version: 12.2 (also tried with 11, but changed nothing)
Code tested: example_at_focussed_bowl_3D.m (model=4)

Posted 9 months ago #
tnie

Member
Joined: Apr '22
Posts: 9

I ran into the same problem with precompiled binaries. Going to try to compile from source and try again.

Posted 9 months ago #
tnie

Member
Joined: Apr '22
Posts: 9

The v1.3 C++ binaries released here are precompiled to support up to Turing. For anything above (3090 is Ampere, 4090 is Ada Lovelace), you need to compile from source and add the CUDA architecture to the binary.

I just compiled from source to make it work on my 4090.

Posted 9 months ago #
f841r

Member
Joined: Feb '23
Posts: 3

It was a problem with my cuda installation. I now installed the recommended drivers over apt and cuda 11.8 using the run file with the --toolkit option. Now everything works with the self-compiled binaries.

Posted 9 months ago #

Reply

You must log in to post.