Compute Capability (CC) is Nvidia's version classification system that defines the features and capabilities supported by different GPU architectures. Each CC version corresponds to specific hardware features like supported CUDA operations, memory hierarchy, instruction sets, and architectural improvements. Higher CC versions generally indicate newer GPU generations with enhanced performance and additional functionality for parallel computing workloads.
The compute capability comprises a major revision number X and a minor revision number Y and is denoted by X.Y. Every major upgrade will show a difference in GPU architecture, while similar major versions usually share the same architecture. You can find out more at section 5.6 .
The table below illustrates the different architectures and the CC versions.
| Architecture | Year | CC |
|---|---|---|
| Fermi | 2010 | 2.0 |
| Kepler | 2012 | 3.0 |
| Maxwell | 2014 | 5.0 - 5.1 |
| Pascal | 2016 | 6.0 - 6.2 |
| Volta | 2017 | 7.0 |
| Turing | 2018 | 7.5 |
| Ampere | 2020 | 8.0 |
| Hopper | 2022 | 9.0 |
| Blackwell | 2024 - 2025 | 10.0, 11.0, 12.0 |
You can check the CC version supported on your GPU using:
You may encounter some issues if you install the incorrect version based on your CC.
1. Installing CUDA version too old for your CC
- Compilation errors: Your GPU's features won't be recognized
- No code generation: CUDA compiler won't generate code for your architecture
- Runtime failures: Applications may fail to detect or utilize your GPU
- Missing optimizations: You'll miss performance improvements designed for your architecture
2. Installing CUDA version too new for your CC
- Runtime errors: "CUDA driver version is insufficient for CUDA runtime version"
- Application crashes: Programs may fail to initialize CUDA context
- Driver incompatibility: Need to upgrade GPU driver to match CUDA version
Best Practice
Install the latest CUDA toolkit that your driver supports, as it will
be compatible with your GPU's CC version and provide the most features and
optimizations. The CUDA toolkit is backward compatible, meaning newer versions
support older GPU architectures, but older CUDA versions may not support newer
CC versions. Run nvidia-smi to see the
maximum CUDA version your current driver supports (shown in the top-right
corner). This is the highest CUDA toolkit version you can install without updating
your GPU driver.