Matlab is unable to find a supported GPU device, but my drivers are installed

I just installed the Parallel Computing Toolbox and tried to use the gpuArray function (on MATLAB R2021a) but recieved the following error:

Error using gpuArray
Unable to find a supported GPU device. For more information on GPU support, see GPU Support by Release.

Then I tried the gpuDevice function and got the same error. I looked at the GPU Support by Release link and tried to find other MATLAB Ask posts on this issue. I found the following post, and I tried the troubleshooting suggestions. Below are the suggestions and then the output I get on my computer (Ubuntu 22.04.2 LTS):

• "Run "nvidia-smi" in a command line and see if the GPU is reported correctly. It should look something like this:

+-----------------------------------------------------------------------------+

| NVIDIA-SMI 440.33.01 Driver Version: 440.33.01 CUDA Version: 10.2 |

|-------------------------------+----------------------+----------------------+

| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |

| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |

|===============================+======================+======================|

| 0 Tesla K80 Off | 00000000:04:00.0 Off | 0 |

| N/A 37C P0 57W / 149W | 0MiB / 11441MiB | 0% Default |

+-------------------------------+----------------------+----------------------+

| 1 Tesla K80 Off | 00000000:05:00.0 Off | 0 |

| N/A 43C P0 70W / 149W | 0MiB / 11441MiB | 51% Default |

+-------------------------------+----------------------+----------------------+

+-----------------------------------------------------------------------------+

| Processes: GPU Memory |

| GPU PID Type Process name Usage |

|=============================================================================|

| No running processes found |

+-----------------------------------------------------------------------------+"

This is what I get from running that command:

As you can see, my GPU is even showing a process from Matlab... so why can't Matlab recognize it?

• "To see if the system sees the GPU, use:

% lspci | grep -i nvidia

04:00.0 3D controller: NVIDIA Corporation GK210GL [Tesla K80] (rev a1)

05:00.0 3D controller: NVIDIA Corporation GK210GL [Tesla K80] (rev a1) "

This is what I get from running that command:

• "If it's working, check the permissions on your GPU device, it should look something like the following. I've seen cases where the GPU is only accessible to root.

% ls -l /dev/nv*

crw-rw-rw- 1 root root 195, 0 Jan 16 16:17 /dev/nvidia0

crw-rw-rw- 1 root root 195, 1 Jan 16 16:17 /dev/nvidia1

crw-rw-rw- 1 root root 195, 255 Jan 16 16:17 /dev/nvidiactl

crw-rw-rw- 1 root root 243, 0 Jan 16 16:17 /dev/nvidia-uvm

crw-rw-rw- 1 root root 243, 1 Jan 16 16:17 /dev/nvidia-uvm-tools"

This is what I get from running that command:

• Check and see if the nvidia module is loaded, and that "nouveau" is not.

% lsmod | grep nvidia

nvidia_uvm 917504 0

nvidia_drm 45056 2

nvidia_modeset 1110016 1 nvidia_drm

nvidia 19894272 2 nvidia_modeset,nvidia_uvm

drm_kms_helper 155648 2 mgag200,nvidia_drm

drm 360448 10 mgag200,ttm,nvidia_drm,drm_kms_helper

ipmi_msghandler 49152 2 nvidia,ipmi_si

% lsmod | grep nouveau

(there should be no output)

This is what I get from those commands:

So basically my system passed all these tests, yet MATLAB still cannot detect my GPU. Any clues would be greatly appreciated.

[Walter Roberson]

Hmmm, in theory the K80 should still work. It appears to have been CUDA 3.7, and R2023a supports CUDA 3.5 and later, so it should still squeak by.

However the Linux complicates matters; I have no idea what adjustments would be needed for Linux.

[Matt J]

Did you try rebooting?

[Mike Croucher]

I don't think Drew is using a K80. K80 is mentioned when he copies and pastes advice from the docs. His own GPU is a GTX 4080 (I'm jealous).

I'm not in the GPU team so can't say for sure but my guess is 'Your copy of MATLAB is too old'. You're using R2021a and the GTX4080 is very shiny-new! Can you update to R2023a?

[Walter Roberson]

@Mike Croucher oh, you are right, I misread that !

[Walter Roberson]

You might need to enable forward compatibility; see https://www.mathworks.com/help/releases/R2021b/parallel-computing/gpu-support-by-release.html#mw_aebb8064-40ea-447f-ac9a-ec875be27224

[Drew Grant]

Hey @Matt J, @Mike Croucher, and @Walter Roberson! Thank you all for your help/suggestions! I got it to work!

I'm assuming I just had some export/library path issues because I ended up just reinstalling all my NVIDIA/cuda/cudnn packages with the following linux commands in the terminal:

(Note after fixing this, I realize this post was likely better suited for NVIDIA forum. So, I do apologize for the confusion but hopefully these steps help someone else who may be going down a similar rabbit hole of frustration)

1. Uninstall everything nvidia and cuda. Warning, it really uninstalls all your nvidia and cuda drivers. (taken from this youtube video)

sudo apt --purge remove "cublas*" "cuda*" 
sudo apt --purge remove "nvidia*"
rm -rf /usr/local/cuda* 
sudo apt-get autoremove 
sudo apt-get autoclean

2. I reinstalled nvidia drivers using:

sudo apt install nvidia-driver-525

*Note, I tried using the Ubuntu Software & Updates app, and none of those drivers worked. This NVIDIA forum post helped me figure out that the drivers from that app seem to be missing some packages.

3. rebooted, and was able to see my gpu detected again with the linux command

nvidia-smi 

4. I installed a cudnn archive from here

sudo tar -xvf cudnn-linux-x86_64-8.4.0.27_cuda11.6-archive.tar.xz 

and it when I ran matlab, it detected my gpu and my day (while wasted) was salvaged.

[Drew Grant]

Also, @Mike Croucher, sure I'll update to the R2023a. Hopefully that'll help with other issues I'm having with MATLAB annihilating my RAM, though this is probably just another Linux-centered thing.

On another note, as excited as I was to get this GPU, I have spent the last week just trying to get it adapted to all of my work. So woe is me.

[Mike Croucher]

It'll be worth it. I have the older 3070 and have been very happy with it! Do you plan to run GPU Bench on it? GPUBench - File Exchange - MATLAB Central (mathworks.com) It was the first thing I did...Walking Randomly » Running MATLAB’s GPU Bench on an NVIDIA GTX 3070

[Drew Grant]

Hey @Mike Croucher, I just ran GPUBench! Thanks for the recommendation. It's pretty cool/helpful to know that my GPU specializes in single precision computations. I will definitely take that into consideration moving forward.

[Mike Croucher]

Wow! Very impressive single precision scores there!

Its double precision speed is still useful! Faster than your CPU on all counts. Very nice piece of kit!