GPU Architecture · Reference Card

CUDA Dev Setup — Windows 11 (RTX 50 / Blackwell)

The four things to install, in the order that avoids pain, plus how to prove each one works.

Track A — WSL2, existing Ubuntu (no Visual Studio)

You already have WSL2 + Ubuntu, so we just install the toolkit into it. The one rule: install only the CUDA toolkit inside WSL, never a Linux display driver — the Windows driver provides the GPU. Follow NVIDIA's CUDA on WSL guide alongside these steps.

1. Confirm the pieces are in place

   Install/refresh the latest Windows NVIDIA driver via the NVIDIA App (the only driver you install — it feeds the GPU into WSL). Then check your distro in PowerShell:

   wsl -l -v

   Should list your Ubuntu as VERSION 2. Note the exact Ubuntu version (lsb_release -a inside it); 22.04 or 24.04 are both fine for CUDA-on-WSL.

2. CUDA Toolkit (inside Ubuntu)

   Enter Ubuntu (wsl), then use the WSL-Ubuntu repo and the toolkit-only package. Representative commands — copy the exact, current ones from the CUDA downloads page → Linux → x86_64 → WSL-Ubuntu → deb (network), since the keyring version changes:

   wget https://developer.download.nvidia.com/compute/cuda/repos/wsl-ubuntu/x86_64/cuda-keyring_1.1-1_all.deb sudo dpkg -i cuda-keyring_1.1-1_all.deb sudo apt-get update sudo apt-get -y install cuda-toolkit-13-3

   critical Install cuda-toolkit-13-3, not cuda-13-3 or cuda — those pull a Linux GPU driver that overwrites the Windows driver stub and breaks the passthrough. The WSL-Ubuntu repo's toolkit package omits the driver by design.

3. Host compiler + PATH

   sudo apt-get install -y build-essential echo 'export PATH=/usr/local/cuda/bin:$PATH' >> ~/.bashrc echo 'export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH' >> ~/.bashrc source ~/.bashrc

   build-essential gives gcc/g++ (nvcc's host compiler — the "no Visual Studio" win). The PATH lines put nvcc and the CUDA libs where the shell can find them.

Track B — Native Windows (Build Tools, no IDE)

1. NVIDIA driver

   Latest Windows driver via the NVIDIA App. Reboot.

2. Build Tools for Visual Studio (not the IDE)

   On the downloads page scroll to "Tools for Visual Studio" → "Build Tools for Visual Studio 2022." In the installer check "Desktop development with C++." No IDE is installed — just the command-line MSVC (cl.exe) + Windows SDK that nvcc needs.

3. CUDA Toolkit 13.3

   Download → Windows → x86_64 → 11 → exe (local). Choose Custom and you can uncheck "Visual Studio Integration" (only useful with the full IDE). nvcc + libraries + Nsight still install. Compile from the "x64 Native Tools Command Prompt" Build Tools adds to the Start menu.

Verify (either track)

Driver version, GPU name, and the max CUDA the driver supports. In WSL this runs inside Ubuntu and still sees your 5050. Proves driver + GPU are alive.

The toolkit compiler version (should say 13.3). Proves nvcc is on your PATH.

First program

Grab the official samples (they moved out of the toolkit to GitHub) and build the classic device query:

…then build deviceQuery (in Samples/1_Utilities/deviceQuery) — it prints every number from these lessons for your real GPU: SM count, registers/SM, shared memory/block, L2 size, memory bandwidth. Or ask your teacher for a minimal hand-written version to compile with nvcc hello.cu -o hello.

Gotchas & good-to-knows

• (Track B) "nvcc can't find cl.exe": compile from the "x64 Native Tools Command Prompt" — it puts MSVC on the PATH.
• (Track A) never install a Linux GPU driver in WSL — use the WSL-Ubuntu toolkit package, which omits it. The Windows driver is the only one.
• Targeting your GPU: Blackwell consumer is compute capability 12.0. CUDA 13.3 compiles for it by default; to be explicit, nvcc -arch=sm_120 ….
• Driver vs. toolkit versions: the driver's CUDA version (in nvidia-smi) only needs to be ≥ the toolkit's. A newer driver with an older toolkit is fine.
• Debugging bad output: run kernels under compute-sanitizer ./prog — it catches out-of-bounds shared/global accesses (the hardware won't) and reports the line.
• Profiling later: Nsight Systems (timeline) and Nsight Compute (per-kernel, shows Lesson 2's "warp stall reasons") ship with the toolkit.