To verify your CUDA installation is working correctly, start by confirming that the NVIDIA driver recognizes your GPU using the nvidia-smi command. This outputs GPU model, driver version, and utilization statistics. If the GPU does not appear, the driver may not be installed or may be incompatible. Next, confirm that CUDA’s compiler (nvcc) is available by running nvcc --version, which should display the installed toolkit version. If both checks succeed, your environment is likely configured correctly.
The next step is to compile and run a simple CUDA sample. The CUDA Toolkit includes an example called “vectorAdd,” which performs element-wise addition on the GPU. Compiling and running this sample confirms that the compiler works, the runtime libraries load correctly, and kernels can execute on the GPU. Many developers also try writing a minimal kernel—such as one that prints thread IDs—to validate that kernel launches behave as expected.
Finally, if CUDA will be used in larger pipelines—such as vector indexing or similarity search in Milvus or Zilliz Cloud—you can test GPU functionality by running sample vector workloads. For example, generate random embeddings, compute L2 distances using a CUDA kernel, and compare results to a CPU implementation. If the GPU results match and performance increases, the CUDA installation is functioning properly. These checks ensure that your environment is ready for both standalone CUDA development and integration with GPU-accelerated vector search workflows.