CUDA_VERSION_MATRIX.md



CUDA Version Matrix
Topmost entry = Tag latest


R
Python
SAGA¹

CUDA
cuBLAS
cuDNN
NCCL
TensorRT
Linux distro


4.4.1
3.12.4
7.3.0
12.5.1
12.5.3.2
8.9.7.29
2.22.3
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
8.5.3²

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
8.5.3²

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
8.5.3²

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
8.5.3²

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
8.5.3²

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
PyTorch³

TensorFlow⁴


3.12
12.5
2.2 ≤ version
2.16 ≤ version


3.11
11.8
2.0 ≤ version
2.12 ≤ version < 2.15


3.10
11.8
1.12 ≤ version
2.8  ≤ version < 2.15


Recommended NVIDIA driver (Regular)


CUDA
Linux driver version
Windows driver version⁵


12.5.1
≥ 555.42.06
≥ 555.85


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 535⁶

Driver version 470⁷


12.5.1
🟢
🟢


12.5.0
🟢
🟢


11.8.0
🟡
🟢


🟢: Works due to the CUDA forward compat package

🟡: Supported due to backward compatibility


qgisprocess image ↩


amd64 only ↩ ↩² ↩³ ↩⁴ ↩⁵


Installs its own CUDA binaries ↩


The expected TensorRT version is symlinked to the installed TensorRT
version.

❗ This relies on backwards compatibility of TensorRT, which may not always be
given. ↩


GPU support in Docker Desktop | Docker Docs

Nvidia GPU Support for Windows · Issue #19005 · containers/podman ↩


EOL: June 2026 ↩


EOL: July 2024 ↩