⚡ Key takeaways
- nvidia-smi shows the highest CUDA version your driver supports, not a version you installed.
nvcc --versionshows the CUDA toolkit, which you need to compile code yourself.- PyTorch brings its own CUDA runtime; torch.version.cuda tells you which one.
- On a GTX 10-series card, CUDA 13 and new PyTorch builds no longer support your GPU.
An app says it needs CUDA 12.8, you run a command, and Windows shows a different number, or none at all. That is normal. A Windows 11 PC has three CUDA numbers, and each one answers a different question. Here is how to check CUDA version for each of them, and which one the app actually means.
Table 1 sums up the three. The rule of thumb: for apps that bring their own CUDA, such as PyTorch builds, “needs CUDA 12.8” means your nvidia-smi number must be 12.8 or higher. Instructions that compile code mean the toolkit.
| Number | Command | What it tells you | When it matters |
|---|---|---|---|
| Driver | nvidia-smi |
The highest CUDA version your driver can run | When an app brings its own CUDA: that should not be newer than this |
| Toolkit | nvcc --version |
The CUDA compiler you installed | When you compile CUDA code yourself |
| PyTorch | torch.version.cuda |
The CUDA runtime inside your PyTorch install | When a PyTorch tool will not use your GPU |
You need PowerShell or Command Prompt and the NVIDIA driver; step 3 also needs Python with PyTorch.
Step 1: Check CUDA version your driver supports
What it does: the NVIDIA driver decides which CUDA versions can run on your PC.
nvidia-smiWhat you should see: a table with your graphics card. The header shows the driver version and, at the top right, “CUDA Version”. That number is the highest CUDA version your driver supports, not a version you installed.
If Windows cannot find nvidia-smi, the NVIDIA driver is missing or damaged, because the tool comes with the driver. Install the current driver from NVIDIA and try again.
Step 2: Check the installed CUDA toolkit
What it does: the toolkit holds the compiler, nvcc, and the libraries for building CUDA programs. You need it to compile code, such as building PyTorch from source or a custom CUDA extension.
nvcc --versionWhat you should see: the toolkit release. It often differs from the number in step 1. A toolkit at or below that number works with your driver.
If Windows says nvcc is not recognized, there is no toolkit on your PATH, which is usually fine for running local AI apps. If you did install one, look for it in the folder below and add its bin folder to your PATH:
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vX.YStep 3: Check which CUDA PyTorch uses
What it does: PyTorch from pip or conda ships with its own CUDA runtime, and your local toolkit is not used, as a PyTorch maintainer explains on the PyTorch forums. Run this in the same Python environment your tool uses:
python -c "import torch; print(torch.__version__, torch.version.cuda, torch.cuda.is_available())"What you should see: the PyTorch version, its CUDA version and True. If the CUDA version says None, you have a CPU-only build. If it says False at the end, see troubleshooting below.
Older cards: GTX 10 series and CUDA 13
Check your card’s compute capability, NVIDIA’s number for its generation:
nvidia-smi --query-gpu=name,compute_cap --format=csvIf it starts with 5, 6 or 7.0, new CUDA software is leaving your card behind. NVIDIA removed compilation and library support for Maxwell (GTX 900), Pascal (GTX 10 series) and Volta in CUDA 13.0, according to the CUDA features archive. CUDA 12.x still covers them. On Linux, NVIDIA confirmed the 580 driver branch as the last one for these cards.
PyTorch follows. Its September 4, 2026 release notice makes 2.14 the last version with CUDA 12.x wheels. From 2.15, prebuilt wheels drop these GPUs, so you stay on 2.14 or earlier, or build from source. From my own use: my GPU is a GTX 1060 with 6 GB, compute capability 6.1, too weak for local models, so I run AI on cloud models instead.
Troubleshooting
An app needs a newer CUDA than nvidia-smi shows
Update the NVIDIA driver, restart, and run nvidia-smi again. Newer drivers still run software built for older CUDA versions. On a GTX 10-series card, first check that a newer driver still exists for your GPU.
torch.cuda.is_available() returns False
If torch.version.cuda is None, reinstall PyTorch with a CUDA build, using the selector on pytorch.org. If it shows a version, your driver is probably older than that runtime: update the driver.
What’s next
If your card is recent enough, the local AI stack with LM Studio puts it to work. If it is not, local vs cloud AI explains what stays private when your models run in the cloud.
Correction, October 17, 2026: An earlier version presented the CUDA version in nvidia-smi as the installed version. It is the highest version your driver supports.
John Lock
Liked this? There's more where it came from.
Get our digest — articles worth your time, no spam, unsubscribe in one click.
Subscribe to Factnetize →
