CUDA on Nvidia MX130 GPU

Welcome to this guide on how to enable CUDA on an Nvidia MX130 GPU for machine vision inference on laptops. CUDA is a parallel computing platform and programming model developed by NVIDIA for general computing on graphical processing units (GPUs).

The Nvidia MX130 is a dedicated GPU that is commonly found in laptops. It is not as powerful as the RTX GPU families, but it can still run some machine vision tasks efficiently. In this guide, we will cover the installation of necessary drivers and software, as well as any configuration required to properly utilize the GPU for these types of computations using Pytorch. Keep in mind that the mileage may vary when using other frameworks.

At the end, I will demo the performance using SuperGlue's demo script.

Checking the GPU you have

Usually, a laptop with an NVIDIA GPU already has NVIDIA drivers preinstalled along with the NVIDIA Control Panel. So, open that to see your GPU version.

From the specs page, it says it can support CUDA. If you have other Nvidia GPUs, you can still follow along with this guide but your mileage may vary.

Optionally, you can verify the CUDA support using Python.

from numba import cuda # run pip install numba
cuda.detect()

Output:

Found 1 CUDA devices
id 0    b'NVIDIA GeForce MX130'                 [SUPPORTED (DEPRECATED)]
                      Compute Capability: 5.0
                           PCI Device ID: 0
                              PCI Bus ID: 1
                                    UUID: GPU-7b092133-34da-571d-9506-9de68403ed55
                                Watchdog: Enabled
                            Compute Mode: WDDM
             FP32/FP64 Performance Ratio: 32
Summary:
	1/1 devices are supported

Process finished with exit code 0

Checking CUDA availability using PyTorch

Even though the script above says that our GPU supports CUDA, PyTorch still cannot 'see' the GPU yet.

Create a new .py file and paste the following code:

import torch as torch # pip install torch

yes_cuda = torch.cuda.is_available()
print(yes_cuda)

Run the code, you'll probbably get the following output:

False

If your output is True, you can skip to Setup Pytorch section right away. If not, we'll need to install some tools to make our CUDA visible to PyTorch.

Install Visual Studio

You may need to install Visual Studio 2022 to correctly install CUDA.

But for my case, I don't install the full VS due to storage constraints, but I have the VS Build Tools already installed with C++ development. I just downloaded Microsoft Visual C++ Redistributable for Visual Studio 2022 just in case. Visit Visual Studio download page for more info.

Setup CUDA

Download supported CUDA version

The latest CUDA version is 12.0. However (at the time of writing), PyTorch seems doesn't have support for it yet. So, you'll need an older version of CUDA (11.6 or 11.7). Go to Cuda Toolkit Archives to download version 11.7.1.

You may select Installer Type to local, and proceed with downloading.

Install CUDA

Once the file is downloaded, double-click it to begin the installation.

Click the Express Installation. It will install the CUDA toolkit, some other things, and your display driver.

When the installation finishes. I got notified about trouble installing Nsight.

As the description said, it may not be related to CUDA so just hit Next and complete the installation.

Verify the installation

Open Command Prompt or Powershell, and run the following command:

nvidia-smi

Setup Pytorch

Go to PyTorch's Get Started page. Select the setting accordingly, copy and run the command generated.

Command:

pip3 install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu117

If you re-run the code previously. The output should be True indicating that PyTorch is now able to recognize the CUDA GPU.

Demo

I'm using the demo of the machine vision project SuperGlue Inference and Evaluation Demo Script without CUDA (using CPU) and with CUDA running on MX130.

Without CUDA

Running inference on device "cpu"

Average FPS = 0.4

With CUDA

Running inference on device "cuda"

Average FPS = 1.0

The performance with CUDA has improved, but it's important to note that it may not achieve significantly higher FPS due to the GPU's limitations in handling intensive calculations.

References

• https://pytorch.org/get-started/locally/
• https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/