Skip to content

Latest commit

 

History

History
73 lines (58 loc) · 3.78 KB

CUDA_VERSION_MATRIX.md

File metadata and controls

73 lines (58 loc) · 3.78 KB

CUDA Version Matrix

Image tags = R versions

Topmost entry = Tag latest

R Python SAGA1 CUDA cuBLAS cuDNN NCCL TensorRT2 Linux distro
4.4.3 3.12.9 7.3.0 12.8.1 12.8.4.1 8.9.7.29 2.25.1 10.9.0.34/
10.3.0.26		 Ubuntu 22.04
4.4.2 3.12.9 7.3.0 12.8.0 12.8.3.14 8.9.7.29 2.25.1 10.8.0.43/
10.3.0.26		 Ubuntu 22.04
4.4.1 3.12.7 7.3.0 12.6.2 12.6.3.3 8.9.7.29 2.23.4 10.6.0.26/
10.3.0.26		 Ubuntu 22.04
4.4.0 3.12.4 7.3.0 12.5.0 12.5.2.13 8.9.7.29 2.21.5 10.0.1.6 Ubuntu 22.04
4.3.3 3.11.9 7.3.0 11.8.0 11.11.3.6 8.9.6.50 2.15.5 8.5.33 Ubuntu 22.04
4.3.2 3.11.8 7.3.0 11.8.0 11.11.3.6 8.9.6.50 2.15.5 8.5.33 Ubuntu 22.04
4.3.1 3.11.6 7.3.0 11.8.0 11.11.3.6 8.9.0.131 2.15.5 8.5.33 Ubuntu 22.04
4.3.0 3.11.4 7.3.0 11.8.0 11.11.3.6 8.9.0.131 2.15.5 8.5.33 Ubuntu 22.04
4.2.3 3.10.11 n/a 11.8.0 11.11.3.6 8.7.0.84 2.15.5 8.5.33 Ubuntu 22.04
4.2.2 3.10.10 n/a 11.8.0 11.11.3.6 8.7.0.84 2.16.2 8.5.3 Ubuntu 20.04

PyTorch/TensorFlow compatibility

Python CUDA PyTorch4 TensorFlow5
3.12 12.8 version ≥ 2.2 2.18 > version ≥ 2.16
3.12 12.6 version ≥ 2.2 2.18 > version ≥ 2.16
3.12 12.5 version ≥ 2.2 2.18 > version ≥ 2.16
3.11 11.8 version ≥ 2.0 2.16 > version ≥ 2.12
3.10 11.8 version ≥ 1.12 2.16 > version ≥ 2.9

Recommended NVIDIA driver (Regular)

CUDA Linux driver version Windows driver version6
12.8.1 ≥ 570.124.06 ≥ 572.61
12.8.0 ≥ 570.117 ≥ 572.30
12.6.2 ≥ 560.35.03 ≥ 560.94
12.5.0 ≥ 555.42.02 ≥ 555.85
11.8.0 ≥ 520.61.05 ≥ 520.06

Supported NVIDIA drivers (LTSB)

Only works with NVIDIA Data Center GPUs or select NGC-Ready NVIDIA RTX boards.

CUDA Driver version 5357 Driver version 4708
12.8.1 🟢 🔵
12.8.0 🟢 🔵
12.6.2 🟢 🔵
12.5.0 🟢 🔵
11.8.0 🟡 🟢

🔵: Supported with the CUDA forward compat package only
🟢: Supported due to minor-version compatibility9
🟡: Supported due to backward compatibility

Footnotes

  1. qgisprocess image ↩

  2. amd64/arm64 ↩

  3. amd64 only ↩ ↩2 ↩3 ↩4 ↩5

  4. Installs its own CUDA binaries ↩

  5. The expected TensorRT version is symlinked to the installed TensorRT version.
    ❗️ This relies on backwards compatibility of TensorRT, which may not always be given. ↩

  6. GPU support in Docker Desktop | Docker Docs
    Nvidia GPU Support for Windows · Issue #19005 · containers/podman ↩

  7. EOL: June 2026 ↩

  8. EOL: July 2024 ↩

  9. or the CUDA forward compat package ↩