Why nvidia-smi and nvcc can show different CUDA versions

Two labels, two questions

nvidia-smi

What the driver can support

The displayed CUDA value is a driver capability indicator. It does not prove that the matching Toolkit is installed.

nvcc --version

Which compiler was invoked

This identifies the Toolkit compiler on the current path. Multiple toolkits or containers can complicate that lookup.

[32][41][42]

Compatibility is directional

A newer driver can often support applications built with older CUDA toolkits. Minor-version compatibility adds bounded behavior within a major toolkit family, subject to documented driver floors and feature limits. Forward-compatibility packages are a separate mechanism for qualified configurations.

“Newer” does not mean “every combination works.” Read the table that matches the driver, toolkit family, operating system, GPU, and application.

[32][33]

A no-install diagnostic checklist

Record the exact GPU model and operating system.
Record the driver version and the CUDA capability displayed by the driver tool.
Record which nvcc executable is on the path, if any.
Identify whether the application bundles its own CUDA runtime.
Check the framework, cuDNN, or TensorRT support matrix.
Change software only after the required combination is explicit.

This sequence gathers evidence. It is intentionally not an installation recipe.

[30][32][33][35][34]