Date: 14.08.2025
CUDA & PyTorch Compatibility by Compute Capability (CC)
This cheat sheet maps Compute Capability (CC) → newest usable CUDA Toolkit → a recent PyTorch version with official wheels → ready-to-copy pip command.
Notes
- Maxwell/Pascal/Volta (CC 5.x/6.x/7.0–7.2) should stay on CUDA 12.x. CUDA 13.x drops support for these architectures in the toolkit.
- PyTorch wheels (cuXXX) bundle the CUDA runtime. You only need the system CUDA Toolkit if you compile custom CUDA extensions.
- Choose the CUDA flavor (cu121 / cu124 / cu126 / cu128) that matches your environment and driver capabilities.
| Compute Capability (CC) | Typical generations | Newest usable CUDA Toolkit | Recommended PyTorch (latest with such wheels) | Example pip command |
|---|---|---|---|---|
| 3.5 / 3.7 | Kepler (e.g., Tesla K80) | CUDA 10.2 | PyTorch 1.12.1 (cu102) | pip install torch==1.12.1+cu102 torchvision==0.13.1+cu102 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu102 |
| 3.5 / 3.7 | Kepler (e.g., Tesla K80) | CUDA 11.4 | PyTorch 2.2.0 | manual build Telsa K80 + PyTorch 2.2.0 |
| 5.0 / 5.2 / 5.3 | Maxwell (Tesla M10/M60, GTX 9xx) | CUDA 12.4 (12.1 also fine) | PyTorch 2.5.0 (cu124 or cu121) | cu124: pip install torch==2.5.0 torchvision==0.20.0 torchaudio==2.5.0 --index-url https://download.pytorch.org/whl/cu124 • cu121: ... --index-url https://download.pytorch.org/whl/cu121 |
| 6.0 / 6.1 / 6.2 | Pascal (P100/P40, GTX 10xx) | CUDA 12.4 | PyTorch 2.5.0 (cu124 or cu121) | same as above (cu124/cu121) |
| 7.0 / 7.2 | Volta (V100/Jetson Xavier) | CUDA 12.4 | PyTorch 2.5.0 (cu124 or cu121) | same as above (cu124/cu121) |
| 7.5 | Turing (RTX 20xx/Quadro RTX) | CUDA 12.8 / 13.x | PyTorch 2.7.1 (cu128; cu126 also available) | cu128: pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128 • cu126: ... --index-url https://download.pytorch.org/whl/cu126 |
| 8.0 / 8.6 / 8.7 | Ampere (A100/RTX 30xx/Orin) | CUDA 12.8 / 13.x | PyTorch 2.7.1 (cu128; cu126 also available) | use the 2.7.1 cu128/cu126 commands above |
| 8.9 | Ada Lovelace (RTX 40xx/L40) | CUDA 12.8 / 13.x | PyTorch 2.7.1 (cu128; cu126 also available) | use the 2.7.1 cu128/cu126 commands above |
| 9.0 | Hopper (H100/GH200) | CUDA 12.8 / 13.x | PyTorch 2.7.1 (cu128; cu126 also available) | use the 2.7.1 cu128/cu126 commands above |
Quick guidance
- If you’re on Tesla M10 (Maxwell, CC 5.0): pick PyTorch 2.5.0 with cu124 (or cu121) wheels. Ensure your NVIDIA driver is from the R570/R580 line.
- For Kepler (CC 3.x): modern PyTorch 2.x wheels don’t target it; use PyTorch 1.12.1 + cu102 or build from source with
TORCH_CUDA_ARCH_LIST="3.5;3.7".