How to Install CUDA Toolkit 13.4 on Windows

CUDA Toolkit 13.4 is now the current general-availability release, and setting it up on Windows has one detail that can easily cause confusion: since CUDA 13.1, NVIDIA no longer bundles the Windows GPU driver with the toolkit. You therefore need a compatible NVIDIA driver and the CUDA development tools installed separately before you can compile and run GPU programs. This guide walks through the setup, verification, and first Visual Studio project without mixing the driver, toolkit, and application layers together.

Check your Windows and NVIDIA hardware first

Before installing CUDA 13.4, confirm that the machine has a CUDA-capable NVIDIA GPU and a supported Windows version. NVIDIA's current Windows documentation lists Windows 11 25H2, 24H2, 23H2 and 22H2-SV2, Windows 10 22H2, Windows Server 2022 and Windows Server 2025 among the supported systems. CUDA 13.4 also supports Visual Studio 2026 18.x, Visual Studio 2022 17.x and Visual Studio 2019 16.x for the supported native Windows development configurations.

You can check the graphics adapter through Windows Device Manager. Open the Run dialog, enter control /name Microsoft.DeviceManager, and expand Display adapters. The important part is not simply seeing an NVIDIA logo; the specific GPU must be one of the CUDA-capable products listed by NVIDIA.

Install the NVIDIA driver before the toolkit

CUDA programs need both the toolkit and an NVIDIA driver that can communicate with the GPU. CUDA 13.4 corresponds to NVIDIA's R615 driver branch, while new CUDA 13.4 features and newly enabled platforms require R615 or later. For the RTX Spark Windows-on-Arm platform, NVIDIA specifies driver version 616.41 or later.

Download and install the appropriate NVIDIA driver for the GPU before installing CUDA Toolkit 13.4. This separation matters because installing the toolkit alone does not give Windows the complete driver stack required to execute CUDA workloads. If Windows Update is actively installing updates, allow that process to finish before beginning the CUDA installation.

Download CUDA Toolkit 13.4.1

The current CUDA 13.4 general-availability release is version 13.4.1, which supersedes the earlier 13.4.0 developer preview. NVIDIA provides both a network installer and a full installer. The network installer downloads selected components during setup, while the full installer contains the toolkit components locally and is more useful when the machine has restricted or unreliable internet access.

Choose the Windows platform and the appropriate architecture in NVIDIA's CUDA download process, then download the installer. For environments where software integrity matters, NVIDIA also publishes an MD5 checksum for the installer so the downloaded file can be checked before execution.

Install CUDA Toolkit 13.4

1. Run the CUDA Toolkit installer with administrator privileges.

2. Allow the installer to perform its system checks and continue to the component selection screen.

3. Install the CUDA compiler, runtime libraries, headers, development tools and Visual Studio integration you need.

4. Keep the default installation location unless your development environment requires a different directory.

5. Complete the installation and restart Windows if the installer requests it.

With the default configuration, CUDA 13.4 places the toolkit under C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.4. The installer also creates the CUDA_PATH environment variable, which Visual Studio projects can use to locate the selected toolkit.

Verify that nvcc can see the CUDA toolkit

Open a new Command Prompt after installation and run nvcc -V. The nvcc command is NVIDIA's CUDA compiler driver, and its version output confirms that the compiler is available through the installed development environment.

nvcc -V

A successful result should identify CUDA 13.4. If Windows reports that nvcc is not recognized, first open a new terminal so it receives the updated environment variables. If the command still fails, inspect the CUDA_PATH environment variable and verify that the expected CUDA installation directory exists.

Run deviceQuery to test the complete CUDA stack

Seeing a version number proves that the compiler is installed, but it does not prove that the driver, toolkit and GPU can work together. NVIDIA recommends building and running the deviceQuery sample from its CUDA Samples project. The sample queries the CUDA runtime and reports whether a CUDA-capable device can actually be detected.

Build the deviceQuery sample with the provided Visual Studio solution and run it. A successful result should identify your NVIDIA GPU and report that the test passed. If the sample says that no CUDA-capable devices are present, the problem is usually below the application level: check the GPU installation, NVIDIA driver and compatibility before changing application code.

Use bandwidthTest for a second verification

The bandwidthTest sample checks communication between the system and the CUDA-capable device and reports transfer measurements. The actual bandwidth numbers vary with the hardware and configuration, so do not treat a particular number as a universal CUDA performance target. The useful result is that the expected GPU is detected and the required tests pass.

deviceQuery
bandwidthTest

Running both samples gives you a more useful baseline than checking nvcc -V alone. The compiler test answers whether the development tools are installed; the samples answer whether the installed software can communicate with the GPU.

Create your first CUDA project in Visual Studio

Once the installation passes verification, you can create a CUDA application directly in Visual Studio. NVIDIA's CUDA 13.4 integration adds CUDA project templates and build customizations to supported Visual Studio installations.

1. Open Visual Studio with a supported MSVC version installed.

2. Select File, then New > Project.

3. Choose the NVIDIA CUDA project template that targets CUDA 13.4.

4. Create the project and allow Visual Studio to generate the CUDA build configuration.

5. Build the project in Debug or Release configuration.

The resulting project is based on a normal Visual Studio C++ project, but NVIDIA's build customization tells Visual Studio how to compile CUDA source files. CUDA source files must be treated as CUDA C/C++ files so that the appropriate compiler and build rules are applied.

Keep the toolkit version explicit in production projects

CUDA installations can coexist, which is useful when different applications require different toolkit versions. NVIDIA's documentation allows a Visual Studio project to use the CUDA_PATH environment variable, but explicitly selecting the toolkit version is often safer for established projects. A future CUDA installation can otherwise change what an existing project resolves to.

For a project that must remain reproducible, record the expected CUDA toolkit and driver requirements alongside the project's build documentation. This becomes particularly important for machine-learning applications because a framework, CUDA runtime, GPU driver and compiled extensions may each impose their own compatibility requirements.

What changes if you are using Windows on Arm

CUDA 13.4 adds Windows on Arm support specifically for NVIDIA RTX Spark devices, extending CUDA's Arm support beyond the Linux environments developers have traditionally used. This is not the same as saying that every Windows-on-Arm computer can suddenly use CUDA: the supported NVIDIA hardware and corresponding driver stack still matter.

NVIDIA's current RTX Spark documentation describes three broad approaches for Windows applications: running existing x86 software through emulation, porting selected components with ARM64EC, or producing fully native ARM64 applications. Developers preparing an application for RTX Spark can therefore begin checking dependencies and building Arm64-compatible components before deploying the final workload to supported hardware.

Fix the most common CUDA installation problems

If nvcc -V works but deviceQuery cannot find a CUDA device, focus on the NVIDIA driver and hardware rather than reinstalling your application. If neither command works, check the toolkit installation and the CUDA_PATH environment variable. If Visual Studio does not show CUDA project templates, verify that a supported Visual Studio version and the CUDA Visual Studio integration component were installed.

Also avoid judging a CUDA installation by one benchmark number. CUDA 13.4 introduces many changes across the compiler, libraries, developer tools, Python components and GPU management features, while the actual performance of an application depends on its GPU, memory access pattern, kernel design and workload. A clean installation is the starting point; profiling and application-specific testing come afterward.