Setting Up on Windows + RTX 4090

The fastest dev box for CUDA in 2026: Windows 11 + RTX 40-series

If you have a recent NVIDIA GPU and want zero friction, Windows is genuinely the path of least resistance. Drivers ship same-day with new GPUs, Visual Studio integration is first-party, and Nsight Compute / Nsight Systems run natively. The Linux setup is doable but the Windows setup is faster.

Step-by-step

Install Visual Studio 2022 (Community edition is free). Pick the 'Desktop development with C++' workload. CUDA's installer needs the C++ compiler and Windows SDK from VS to integrate properly. Without VS, NVCC has nothing to call as the host compiler on Windows.
Install the latest Game Ready or Studio driver from NVIDIA's website. Make sure the version supports the CUDA Toolkit version you want (check the Toolkit release notes). For CUDA 13.2 on Windows, you need driver 555.x or newer.
Install CUDA Toolkit 13.2. Download from NVIDIA, run the installer, pick 'Express (Visual Studio Integration).' This installs NVCC, libraries (cuBLAS, cuDNN if bundled, cuFFT), Nsight tools, and the VS integration that makes 'CUDA Runtime' a project template in VS.
Verify the toolchain from a Developer Command Prompt for VS 2022 (this is the prompt where cl.exe is on PATH).

Code

Verify NVCC + cl.exe in the same shell·bash

:: Open: Developer Command Prompt for VS 2022
nvcc --version
:: Should print: 'release 13.2, V13.2.xx'

cl.exe
:: Should print: 'Microsoft (R) C/C++ Optimizing Compiler Version 19.xx'
:: NVCC needs cl.exe as the host compiler — this proves both are visible.

Compile and run hello.cu directly·bash

:: Save the hello kernel from the next lesson as hello.cu, then:
nvcc -arch=sm_89 hello.cu -o hello.exe
hello.exe
::  Hello from block 0, thread 0
::  Hello from block 0, thread 1
::  Hello from block 0, thread 2
::  Hello from block 0, thread 3

Exercise

Even if your dev machine is a Mac, walk through the Windows path on a friend's PC or a cloud Windows instance once. Why? Because nearly every CUDA tutorial assumes Windows or Linux + NVIDIA, and seeing the smooth Windows install yourself makes you fluent in the docs. Stop after nvcc --version works — the rest of this track will work on either OS.