The compact definition
A compute-capability value is attached to a GPU architecture implementation. CUDA tools and applications use it to reason about hardware features and which compiled code can execute. The authoritative lookup is NVIDIA's current CUDA GPU table.
compute capability ≠ toolkit version ≠ performance
[30][5]Route each question to the right evidence
| Question | Look up | Do not infer |
|---|---|---|
| What hardware features can this GPU expose? | GPU compute capability | Application speed |
| What compiler is installed? | nvcc / Toolkit installation | From nvidia-smi alone |
| What CUDA level can the driver support? | Driver and compatibility documentation | That every toolkit is installed |
| Will this framework build work? | Framework plus library support matrices | From the GPU family name alone |
A five-step compatibility check
- Identify the exact GPU model.
- Find its compute capability in the maintained GPU table.
- Identify the installed driver branch and its supported CUDA level.
- Identify the locally installed toolkit and compiler separately.
- Check the application, framework, cuDNN, and TensorRT requirements that actually apply.
This reference explains the model; it does not inspect a reader's computer or certify an installation.
[30][31][33][34][36][37]Rubin example: 10.7 does not imply every software path is ready
TensorRT 11.3.0 lists Vera Rubin with compute capability 10.7 for its named Linux and Windows platforms. Tile IR 13.4 lists sm_107 as developer preview. The hardware designation does not override either component’s support qualification.