Table of Contents

Installing the native libraries

NVMath.Net does not ship NVIDIA binaries. Install the ones you need and make sure they can be found when your program starts. A library is loaded lazily on first use, so you only need the libraries for the modules you actually call.

What each module needs

Module Native library Windows file Linux file Version
Core (devices, buffers, streams) CUDA runtime cudart64_13.dll libcudart.so.13 CUDA 13.x
Linear algebra cuBLAS, cuBLASLt cublas64_13.dll, cublasLt64_13.dll libcublas.so.13, libcublasLt.so.13 13.x
FFT cuFFT cufft64_12.dll libcufft.so.12 12.x
Sparse cuSPARSE cusparse64_12.dll libcusparse.so.12 12.x
Sparse direct solver cuDSS cudss64_0.dll libcudss.so.0 0.x
Dense solvers cuSOLVER (+ cuBLAS) cusolver64_12.dll libcusolver.so.12 12.x
Random numbers cuRAND curand64_10.dll libcurand.so.10 10.x
Tensors cuTENSOR cutensor.dll libcutensor.so.2 2.x

cuBLAS, cuBLASLt, cuFFT, cuSPARSE, cuSOLVER and cuRAND come with the CUDA Toolkit. cuTENSOR and cuDSS are separate downloads and are only needed for the tensor module and SparseDirectSolver.

Installing

CUDA Toolkit

Download the CUDA Toolkit 13.x installer from NVIDIA (or use your distribution's packages) and install it. The GPU driver must support the toolkit version.

cuTENSOR and cuDSS

Both are available from NVIDIA's developer downloads (pick the build for your CUDA major version), as pip wheels (cutensor-cu13, and the corresponding cuDSS wheel) or through conda-forge. On Windows the archives install into C:\Program Files\NVIDIA cuTENSOR\vX.Y\ and C:\Program Files\NVIDIA cuDSS\vX.Y\, with the DLLs under bin\12 and bin\13.

How the library finds them

Libraries are resolved by name in this order:

  1. The operating system's normal search path: PATH on Windows, LD_LIBRARY_PATH / ldconfig on Linux.
  2. Windows: the CUDA Toolkit folders under C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v*\bin (newest first), and the standard install folders of cuTENSOR and cuDSS (…\bin\<cuda-major>, newest first). The CUTENSOR_ROOT and CUDSS_ROOT environment variables can point at a custom location.
  3. Linux: /usr/local/cuda/lib64 and /usr/lib/x86_64-linux-gnu.

In practice a default CUDA Toolkit, cuTENSOR and cuDSS installation on Windows works without touching PATH. If you installed elsewhere, add the folder containing the DLLs (or .so files) to PATH / LD_LIBRARY_PATH.

Checking your setup

using NVMathNet;

Console.WriteLine($"Devices: {CudaDevice.GetDeviceCount()}");
Console.WriteLine($"Driver:  {CudaDevice.GetDriverVersion()}  Runtime: {CudaDevice.GetRuntimeVersion()}");

If a library cannot be found you get a DllNotFoundException naming the file. See Troubleshooting.

Edit this page