Date: 14.08.2025
NVIDIA GPU architectures, drivers, and CUDA — quick cheat sheet
The table below shows, for each NVIDIA architecture generation: its Compute Capability (SM), the last Linux driver branch that still supports it, and the newest compatible CUDA Toolkit you can use.
For your Tesla M10 (Maxwell, SM 5.0): the last driver branch is R580, and the newest CUDA that still supports it is 12.x (e.g., 12.8).
Architecture (examples) | Compute Capability (SM) | Last Linux driver branch that supports the architecture | Newest compatible CUDA Toolkit |
---|---|---|---|
Kepler (GTX 780, Tesla K80) | 3.0 / 3.5 / 3.7 | R470 (Legacy) — support removed after this branch | 11.4 for SM 3.5/3.7; for SM 3.0 use 10.2 |
Maxwell (Tesla M10/M60, GTX 9xx) | 5.0 / 5.2 / 5.3 | R580 — last branch for Maxwell | 12.x |
Pascal (P100/P40, GTX 10xx) | 6.0 / 6.1 / 6.2 | R580 — last branch for Pascal | 12.x |
Volta (V100/Jetson Xavier) | 7.0 / 7.2 | R580 — last branch for Volta | 12.x |
Turing (RTX 20xx/Quadro RTX) | 7.5 | Supported by current branches (R580 and newer) | 13.x and newer |
Ampere (A100/RTX 30xx/Orin) | 8.0 / 8.6 / 8.7 | Supported by current branches | 13.x and newer |
Ada Lovelace (RTX 40xx/L40) | 8.9 | Supported by current branches | 13.x and newer |
Hopper (H100/GH200) | 9.0 | Supported by current branches | 13.x and newer |
Blackwell (B200/GB200) | 10.0 | Supported by current branches | 13.x and newer |
Tips
- For Maxwell / Pascal / Volta stay on R580 (or earlier) and use CUDA 12.x. CUDA 13.0+ no longer targets SM 5.x / 6.x / 7.0–7.2.
- On Ubuntu 24.04, prefer installing cuda-toolkit-12-x from NVIDIA’s APT repo (and avoid the
cuda
/cuda-drivers
metapackages if you already manage the driver separately).