Skip to main content
Last Updated: April 20, 2026

​
Overview

Nvidia’s RTX 5090 and RTX 5080 GPUs (Blackwell architecture) require CUDA 12.8. PyTorch 2.7.0 was the first stable release to add native sm_120 support - shipping pre-built CUDA 12.8 wheels with updated cuDNN, NCCL, and Triton 3.3. The current recommended version is 2.11.0. Any PyTorch image with the -cuda12.8- tag on version 2.7.0 or later will work with RTX 50-series GPUs. Use the latest images from the official PyTorch Docker Hub:
  • pytorch/pytorch:2.11.0-cuda12.8-cudnn9-runtime
  • pytorch/pytorch:2.11.0-cuda12.8-cudnn9-devel
Earlier versions with CUDA 12.8 support (2.7.0 minimum):
  • pytorch/pytorch:2.9.1-cuda12.8-cudnn9-runtime
  • pytorch/pytorch:2.9.1-cuda12.8-cudnn9-devel
  • pytorch/pytorch:2.8.0-cuda12.8-cudnn9-runtime
  • pytorch/pytorch:2.7.0-cuda12.8-cudnn9-runtime

​
Example Dockerfile

FROM pytorch/pytorch:2.11.0-cuda12.8-cudnn9-runtime

WORKDIR /app
COPY requirements.txt .
RUN pip install --no-cache-dir -r requirements.txt

COPY . .
CMD ["python", "train.py"]
If you intend to run a workload both on 50-series GPUs and older 40- or 30-series GPUs, you might need to maintain separate Docker images. Older GPUs might not support CUDA 12.8, so a 50-series image will not work on older hardware.