PyTorch Versions and Supported NVIDIA GPU Compute Capability Levels (Updated June 2026)
Companies with legacy PyTorch codebases often run into problems like "our old code doesn't run on the latest PyTorch" or "it doesn't run on the latest GPUs (such as the GeForce RTX 50 series, i.e., Blackwell)." In this article, we organize the relationships between PyTorch versions, supported GPU Compute Capabilities (SM levels), and supported CUDA versions, updated with information as of June 2026.
🔄 June 2026 Update
- The Blackwell generation (GeForce RTX 50 series = sm_120 / data center B100 and B200 = sm_100) is supported by CUDA 12.8 and later.
- PyTorch officially added Blackwell support in 2.7.0 (April 2025), providing CUDA 12.8 build (
cu128) wheels. The subsequent 2.8 / 2.9 / 2.10 lines also support it. - Starting with the cu128 builds of 2.8, Maxwell / Pascal (sm_50 and sm_60) are no longer supported (Volta sm_70 remains supported).
- The current stable line is 2.10 (January 2026). In addition to CUDA 12.6 / 12.8, a CUDA 13.0 build is also available, and Python is supported up to 3.14.
- CUDA 13.0 drops support for Maxwell / Pascal / Volta (sm_50 through sm_72); supported architectures now start at Turing (sm_75). If you rely on older GPU assets, you need to stay on the CUDA 12.x line.
GPU Compute Capabilities Supported by PyTorch
| PyTorch version | Supported Compute Capability (SM) levels |
|---|---|
| 1.0.0 - 1.3.1 | SM_35, SM_37, SM_50, SM_60, SM_61, SM_70 |
| 1.4.0 - 1.7.1 | SM_37, SM_50, SM_60, SM_61, SM_70, SM_75 |
| 1.8.0 - 1.8.1 | SM_37, SM_50, SM_60, SM_70, SM_75, SM_80 |
| 1.9.0 - 1.12.1 | SM_37, SM_50, SM_60, SM_70, SM_75, SM_80, SM_86 |
| 1.13.0 - 2.0.1 | SM_37, SM_50, SM_60, SM_70, SM_75, SM_80, SM_86, SM_89, SM_90 |
| 2.1.0 - 2.6.0 | SM_50, SM_60, SM_70, SM_75, SM_80, SM_86, SM_89, SM_90 |
| 2.7.0 | SM_50, SM_60, SM_70, SM_75, SM_80, SM_86, SM_89, SM_90, SM_100, SM_120 (Blackwell support, cu128 wheels) |
| 2.8.0 and later | SM_70, SM_75, SM_80, SM_86, SM_89, SM_90, SM_100, SM_120 (cu128 drops Maxwell sm_50 and Pascal sm_60) |
Reference page for installing older PyTorch versions
https://pytorch.org/get-started/previous-versions/
Python Versions Supported by PyTorch
| PyTorch version | Supported Python versions |
|---|---|
| 1.4.0 - 1.7.1 | 3.6, 3.7, 3.8 |
| 1.8.0 - 1.9.1 | 3.6, 3.7, 3.8, 3.9 |
| 1.10.0 - 1.12.1 | 3.7, 3.8, 3.9, 3.10 |
| 1.13.0 - 2.0.1 | 3.8, 3.9, 3.10, 3.11 |
| 2.1.0 - 2.5.1 | 3.8, 3.9, 3.10, 3.11, 3.12 |
| 2.6.0 - 2.8.0 | 3.9, 3.10, 3.11, 3.12, 3.13 |
| 2.9.0 and later | 3.10, 3.11, 3.12, 3.13, 3.14 |
Compute Capability and Representative GPU Lineups
| SM generation | Data center / professional GPUs | GeForce GPUs |
|---|---|---|
| SM_120 (Blackwell) | - | GeForce RTX 5090, RTX 5080, RTX 5070, RTX 5060 |
| SM_100 (Blackwell) | NVIDIA B100 (GB100), B200, GB202, GB203, GB205, GB206, GB207, NVIDIA B40, RTX PRO 6000 Blackwell | - |
| SM_90 (Hopper) | NVIDIA H100, NVIDIA H200 | - |
| SM_89 (Ada Lovelace) | NVIDIA L4, NVIDIA L40, RTX 6000 Ada Generation, L40s Ada | GeForce RTX 4090, GeForce RTX 4080, GeForce RTX 4070 Ti / 4070, GeForce RTX 4060 Ti / 4060, GeForce RTX 4050 |
| SM_86, SM_87 (Ampere) | NVIDIA A100, NVIDIA A30, NVIDIA A40, RTX A2000, A3000, RTX A4000, A5000, A6000, A10, A16, A2 Tensor Core GPU, A800 40GB | GeForce RTX 3090, GeForce RTX 3080, GeForce RTX 3070, GeForce RTX 3060, GeForce RTX 3050 |
| SM_75 (Turing) | Quadro RTX 8000, RTX 6000, RTX 5000, RTX 4000, T4 | GeForce RTX 2080 Ti, GeForce RTX 2070, GeForce GTX 1660 Ti |
| SM_70, SM_72 (Volta) | Tesla V100, Quadro GV100, Titan V, Xavier NX | - |
| SM_60, SM_61, SM_62 (Pascal) | Tesla P100, Quadro GP100 | GeForce GTX 1080, GTX 1070, GTX 1060, GTX 1050, GTX 1030, GT 1010, Titan Xp |
| SM_50, SM_52, SM_53 (Maxwell) | Tesla M-series, Quadro M6000 | GeForce GTX 980 Ti, GTX Titan X, GTX 980, GTX 970 |
| SM_35, SM_37 (Kepler) | Tesla K40, Tesla K80 | - |
| SM_30 (Kepler) | - | GeForce 700 series, GT-730 |
| SM_20 (Fermi) | - | GeForce 400, 500, 600, GT-630 |
Source: https://journal.qualiteg.com/nvidia-gpu-capability-level/
Supported CUDA Versions
| SM generation | Architecture | Minimum compatible CUDA version | Status in latest CUDA (12.8 / 13.0) |
|---|---|---|---|
| SM_120 | Blackwell | CUDA 12.8 or later | Supported |
| SM_100, SM_101 | Blackwell | CUDA 12.8 or later | Supported |
| SM_90 | Hopper | CUDA 12.0 or later | Supported |
| SM_89 | Ada Lovelace | CUDA 11.8 or later | Supported |
| SM_86, SM_87 | Ampere | CUDA 11.1 or later | Supported |
| SM_80 | Ampere | CUDA 11.0 or later | Supported |
| SM_75 | Turing | CUDA 10.0 or later | Supported |
| SM_70, SM_72 | Volta | CUDA 9.0 or later | Support ended in CUDA 13.0 |
| SM_60, SM_61, SM_62 | Pascal | CUDA 8.0 or later | Support ended in CUDA 13.0 |
| SM_50, SM_52, SM_53 | Maxwell | CUDA 6.0 or later | Support ended in CUDA 13.0 |
| SM_30, SM_35, SM_37 | Kepler | CUDA 5.0 or later | Support ended |
| SM_20 | Fermi | CUDA 3.2 or later | Support ended |
Get advice on GPU environment selection and buildout from a team with real operational experience.
Combinations of PyTorch, CUDA, and GPU generations raise no end of questions—from just getting things running to operating them afterward. We operate our own GPU clusters and develop LLM products in-house.
From GPU selection (H100 / A100 / RTX series) to inference optimization (quantization, vLLM) and multi-GPU, multi-node distributed configurations, our advice is grounded in operational experience—not theory.
Explore our LLM infrastructure services →Related Posts
・2025 NVIDIA GPU Quick Search Tool・NVIDIA GPUs and Capability Levels
