Jetson Nano GPU Usage Checking in C++

The Jetson Nano is a popular embedded computer that packs a lot of power in a small package. It includes an NVIDIA GPU that can be used for a wide range of tasks, from machine learning to graphics rendering. In this tutorial, we'll explore how to check GPU usage on the Jetson Nano using C++ and the NVIDIA Jetson SDK.

Prerequisites

To follow along with this tutorial, you will need a Jetson Nano developer kit and a basic understanding of C++. You should also have the NVIDIA Jetson SDK installed on your system. If you haven't already done so, you can download the latest version from the NVIDIA website.

Step 1: Initializing CUDA

Before we can start checking GPU usage, we need to initialize the CUDA environment. CUDA is a parallel computing platform that enables developers to use NVIDIA GPUs for general-purpose computing. Here's how to initialize CUDA in C++:

#include <cuda_runtime_api.h>

int main() {
    // Initialize CUDA
    cudaError_t cudaStatus = cudaSetDevice(0);
    if (cudaStatus != cudaSuccess) {
        printf("cudaSetDevice failed!  Do you have a CUDA-capable GPU installed?");
        return 1;
    }

    // Your code here...

    return 0;
}

In this example, we use the cudaSetDevice function to initialize CUDA and select the first available GPU. If cudaSetDevice returns an error, we print an error message and exit the program.

Step 2: Checking GPU Usage

Once CUDA is initialized, we can check the current GPU usage using the nvidia-smi command. This command-line tool is included with the NVIDIA driver and provides a wealth of information about the current GPU state. To invoke nvidia-smi from C++, we can use the system function:

#include <stdlib.h>

int main() {
    // Your CUDA initialization code here...

    // Check GPU usage
    system("nvidia-smi");
    
    // Your code here...

    return 0;
}

This will execute the nvidia-smi command and print its output to the console. You can also redirect the output to a file or parse it programmatically using C++.

Step 3: Using the NVIDIA Management Library (NVML)

For more fine-grained control over GPU usage, we can use the NVIDIA Management Library (NVML). NVML is a C-based API that provides programmatic access to GPU information, management, and monitoring functions. Here's how to use NVML to check GPU usage in C++:

#include <nvml.h>

int main() {
    // Your CUDA initialization code here...

    // Initialize NVML
    nvmlReturn_t result = nvmlInit();
    if (result != NVML_SUCCESS) {
        printf("Failed to initialize NVML: %s\n", nvmlErrorString(result));
        return 1;
    }

    // Get GPU utilization
    nvmlDevice_t device;
    result = nvmlDeviceGetHandleByIndex(0, &device);
    if (result != NVML_SUCCESS) {
        printf("Failed to get device handle: %s\n", nvmlErrorString(result));
        return 1;
    }

    nvmlUtilization_t utilization;
    result = nvmlDeviceGetUtilizationRates(device, &utilization);
    if (result != NVML_SUCCESS) {
        printf("Failed to get utilization rates: %s\n", nvmlErrorString(result));
        return 1;
    }

    printf("GPU utilization: %d\n", utilization.gpu);

    // Clean up NVML
    nvmlShutdown();

    return 0;
}

In this example, we use the nvmlInit function to initialize NVML and select the first available GPU. If nvmlInit returns an error, we print an error message and exit the program.

We then use the nvmlDeviceGetHandleByIndex function to get a handle to the first GPU. We use this handle to get the current GPU utilization rates using the nvmlDeviceGetUtilizationRates function. We print the GPU utilization rate to the console and then clean up NVML using the nvmlShutdown function.

Conclusion

In this tutorial, we explored how to check GPU usage on the Jetson Nano using C++ and the NVIDIA Jetson SDK. We used both the nvidia-smi command-line tool and the NVIDIA Management Library (NVML) to obtain GPU usage information. By integrating these tools into your own C++ applications, you can gain better visibility into your Jetson Nano's GPU usage and optimize your code for maximum performance.