When working in fields related to data science, machine learning, or other computational areas that utilize GPU computing, you may often find yourself needing to determine the CUDA version installed on your system. CUDA (Compute Unified Device Architecture) is a parallel computing platform and application programming interface (API) model created by NVIDIA. It allows software developers to use a CUDA-enabled GPU for general purpose processing.

Identifying the CUDA Version

Knowing the version of CUDA installed on your system is crucial for compatibility with various software frameworks, especially those in machine learning like TensorFlow, PyTorch, and others. Here are several methods to find out the CUDA version:

Method 1: Using the nvcc Command

The simplest approach to discover the CUDA version is by using the NVIDIA CUDA Compiler (nvcc), which is included with the CUDA Toolkit. You can execute the following command in the terminal:

nvcc --version

This command directly outputs the CUDA version, among other details. For instance, the output might look like this:

1nvcc: NVIDIA (R) Cuda compiler driver
2Copyright (c) 2005-2020 NVIDIA Corporation
3Built on Mon_Nov_30_19:08:53_Pacific_Standard_Time_2020
4Cuda compilation tools, release 11.2, V11.2.152

Here Cuda compilation tools, release 11.2 indicates that the version is 11.2.

Method 2: Inspecting the CUDA Toolkit Directory

In most installations, the CUDA toolkit is installed under /usr/local/cuda, which is symlinked to the specific version like /usr/local/cuda-11.2. You can view the version by examining this symlink:

ls -l /usr/local/cuda

The output might look like:

lrwxrwxrwx 1 root root 9 Jan 24 2021 /usr/local/cuda -> cuda-11.2

Method 3: Through Environment Variables

If the CUDA_HOME environment variable is set, it typically points to the directory where CUDA is installed. You can echo this variable:

echo $CUDA_HOME

The output might provide a clue about the version if the installation path includes it.

Method 4: Using CUDA Runtime API

For developers working with CUDA in C or C++, the CUDA Runtime API provides functions to determine the version programmatically. Here’s a simple example:

1#include <cuda_runtime.h>
2#include <stdio.h>
3
4int main() {
5    int driver_version = 0, runtime_version = 0;  
6    cudaDriverGetVersion(&driver_version);
7    cudaRuntimeGetVersion(&runtime_version);
8    printf("Driver Version: %d.%d\n", driver_version / 1000, (driver_version % 100) / 10);
9    printf("Runtime Version: %d.%d\n", runtime_version / 1000, (runtime_version % 100) / 10);  
10    return 0;
11}

This code snippet when executed, outputs the CUDA Driver and Runtime versions, which can be crucial for debugging compatibility issues.

Summary Table

Each of these methods has its uses depending on your access and the type of information you need—either for a quick lookup or for integration into software for dynamic compatibility checking.