CUDA driver compatibility: match the GPU, driver, toolkit, and framework

Start with the layer, not the largest number

CUDA-related layers and the evidence each one owns
Layer	What it describes	Do not infer
Driver / KMD	The installed kernel-mode display driver	Which Toolkit compiler is on your path
CUDA UMD	The CUDA user-mode driver level reported by the driver tool	That the matching Toolkit is installed
Toolkit / compiler	The development toolkit selected by `nvcc`	Which runtime a container or framework bundled
Framework / runtime	The build-specific runtime and library envelope	Universal support from one version string
GPU	Hardware features expressed through compute capability	Toolkit version or workload performance

[35][33][30]

What are you trying to run?

An existing CUDA applicationStart with its framework or vendor requirements, then check the host driver and GPU.Code you compile locallyCheck the selected nvcc Toolkit, target compute capability, and driver floor.A framework wheel or containerUse that build's runtime and support matrix; do not substitute the system Toolkit version.A newer Toolkit on an older driverDetermine whether backward, minor-version, or qualified forward compatibility actually applies.

[79][33][30][37][36]

What nvidia-smi and nvcc answer

nvidia-smi

What the driver environment reports

The displayed CUDA field describes the driver-side CUDA level. 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.

[35][33][43][44]

Older labels and newer KMD / CUDA UMD labels

NVIDIA's current documentation marks Driver Version and CUDA Version as deprecated field names, with KMD Version and CUDA UMD Version documented as their newer counterparts. An installed driver branch can still present older labels. NVIDIA also warns that human-readable nvidia-smi output is not guaranteed to stay backward compatible, so durable tooling should use NVML rather than scrape this display.

Illustrative older field names—not captured hardware output

Driver Version: <installed driver>
CUDA Version: <driver-supported CUDA level>

Illustrative newer field names—not captured hardware output

KMD Version: <installed driver>
CUDA UMD Version: <driver-supported CUDA level>

[35]

Frameworks and containers can package another runtime

A framework wheel, container image, or application can carry CUDA runtime libraries that differ from a system Toolkit selected by nvcc. That packaged runtime still depends on an adequate host driver and on the framework's own support matrix. A missing nvcc command can simply mean no compiler Toolkit is on the current path; it does not establish that no CUDA-enabled application can run.

[33][36][37]

Compatibility is directional

Backward compatibility is the ordinary case in which a newer driver supports an application built with an older Toolkit. Minor-version compatibility can let CUDA 11 and later applications run within the same major Toolkit family on a sufficiently new driver, but newer features and PTX can require a newer driver. Forward compatibility uses an additional compatibility package and is limited to supported data-center GPUs, selected NGC Server Ready RTX configurations, and Jetson; it is not a general GeForce workaround.

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

[33][34][62]

Check the combination in this order

GPU: record the exact model and find its compute capability on NVIDIA's GPU list.
Driver: record the installed driver and the driver-supported CUDA level shown by nvidia-smi.
Toolkit: if you compile locally, use nvcc -V to identify the compiler Toolkit on the current path.
Runtime or framework: identify the CUDA runtime bundled by the wheel, container, or application and read that project's support matrix.
Compatibility mode: compare the exact Toolkit release notes and CUDA compatibility table; do not infer support from matching major numbers alone.
Test: run the application's documented check or NVIDIA's deviceQuery sample before changing a working driver.

You can stop when the application's documented GPU, driver, runtime, and framework conditions all match and its functional check passes. A matching nvidia-smi label by itself is not that result.

[30][35][79][33][34][37][36]

Then identify the exact Toolkit release

The version number answers a different question from the driver and GPU checks above. Use the CUDA release tracker to find the listed release month, channel, reviewed material changes, and direct NVIDIA sources for a specific Toolkit version.

The tracker is a documentary index, not a compatibility verdict or install recommendation.

[32][31]