30系/40系GPU使用TensorFlow1.x的方法

1. Windows平台

在Windows上平台比较简单，可以在这个网站找到编译好的cuda11.1版本的TensorFlow1.15的whl文件，然后安装cuda11.8就可以支持40系显卡了

TensorFlow1.15.4+cuda11.1+avx2https://github.com/fo40225/tensorflow-windows-wheel/tree/master/1.15.4%2Bnv20.12/py38/CPU%2BGPU/cuda111cudnn8avx2

2. Linux平台

首先在这里下载对应cuda版本的TensorFlow，具体版本对应参考段末列表

NVIDIA-TensorFlowhttps://developer.download.nvidia.com/compute/redist/nvidia-tensorflow

查看系统版本 

# 查看系统版本
lsb_release -a

# 查看cuda版本
ls -l /usr/local/cuda

 对于Ubuntu 18.02，只能创建Python3.6环境，不能用python3.8，否则会出现glibc2.29找不到的问题

conda create -n tf-1.15 python=3.6
conda activate tf-1.15
pip install --user nvidia-pyindex
export PATH=$PATH:$HOME/.local/bin
# 提前下载好 https://developer.download.nvidia.com/compute/redist/nvidia-tensorflow/nvidia_tensorflow-1.15.4+nv20.10-cp36-cp36m-linux_x86_64.whl
pip install nvidia_tensorflow-1.15.4+nv20.10-cp36-cp36m-linux_x86_64.whl opencv-python==3.4.5.20 matplotlib==3.3.0 -i https://pypi.tuna.tsinghua.edu.cn/simple

  如果是Ubuntu20，则Python3.8/Python3.6均可

conda create -n tf-1.15 python=3.8
conda activate tf-1.15
pip install --user nvidia-pyindex
export PATH=$PATH:$HOME/.local/bin
# tf15_cuda-11.1: https://developer.download.nvidia.com/compute/redist/nvidia-tensorflow/nvidia_tensorflow-1.15.4+nv20.12-cp38-cp38-linux_x86_64.whl
# cudnn-8.0.5.43: https://developer.download.nvidia.cn/compute/redist/nvidia-cudnn/nvidia_cudnn-8.0.5.43-py3-none-manylinux1_x86_64.whl 
pip install nvidia_tensorflow-1.15.4+nv20.12-cp38-cp38-linux_x86_64.whl opencv-python matplotlib==3.3.0 -i https://pypi.tuna.tsinghua.edu.cn/simple

如果是cuda12.x 

conda create -n tf-1.15 python=3.8
conda activate tf-1.15
pip install --user nvidia-pyindex
export PATH=$PATH:$HOME/.local/bin
pip install nvidia_tensorflow-1.15.5+nv23.03-7472065-cp38-cp38-linux_x86_64.whl opencv-python==3.4.10.35 matplotlib==3.3.0 tensorrt==8.6.0 -i https://pypi.tuna.tsinghua.edu.cn/simple

测试

import tensorflow as tf
tf.test.is_gpu_available() 

cuda-11.0  nvidia_tensorflow-1.15.2+nv20.06-cp36-cp36m-linux_x86_64.whl cuda-11.0 nvidia_tensorflow-1.15.3+nv20.07-cp36-cp36m-linux_x86_64.whl cuda-11.0 nvidia_tensorflow-1.15.3+nv20.08-cp36-cp36m-linux_x86_64.whl cuda-11.0 nvidia_tensorflow-1.15.3+nv20.09-cp36-cp36m-linux_x86_64.whl cuda-11.1 nvidia_tensorflow-1.15.4+nv20.10-cp36-cp36m-linux_x86_64.whl cuda-11.1 nvidia_tensorflow-1.15.4+nv20.11-cp36-cp36m-linux_x86_64.whl cuda-11.1 nvidia_tensorflow-1.15.4+nv20.12-cp38-cp38-linux_x86_64.whl cuda-11.2 nvidia_tensorflow-1.15.5+nv21.02-cp38-cp38-linux_x86_64.whl cuda-11.2 nvidia_tensorflow-1.15.5+nv21.03-cp38-cp38-linux_x86_64.whl cuda-11.3 nvidia_tensorflow-1.15.5+nv21.04-cp38-cp38-linux_x86_64.whl cuda-11.3 nvidia_tensorflow-1.15.5+nv21.05-cp38-cp38-linux_x86_64.whl cuda-11.3 nvidia_tensorflow-1.15.5+nv21.06-cp38-cp38-linux_x86_64.whl cuda-11.4 nvidia_tensorflow-1.15.5+nv21.07-cp38-cp38-linux_x86_64.whl cuda-11.4 nvidia_tensorflow-1.15.5+nv21.08-cp38-cp38-linux_x86_64.whl cuda-11.4 nvidia_tensorflow-1.15.5+nv21.09-cp38-cp38-linux_x86_64.whl cuda-11.4 nvidia_tensorflow-1.15.5+nv21.10-cp38-cp38-linux_x86_64.whl cuda-11.5 nvidia_tensorflow-1.15.5+nv21.11-cp38-cp38-linux_x86_64.whl cuda-11.5 nvidia_tensorflow-1.15.5+nv21.12-cp38-cp38-linux_x86_64.whl cuda-11.6 nvidia_tensorflow-1.15.5+nv22.01-3720650-cp38-cp38-linux_x86_64.whl cuda-11.6 nvidia_tensorflow-1.15.5+nv22.02-3927706-cp38-cp38-linux_x86_64.whl cuda-11.6 nvidia_tensorflow-1.15.5+nv22.03-4138614-cp38-cp38-linux_x86_64.whl cuda-11.6 nvidia_tensorflow-1.15.5+nv22.04-4387458-cp38-cp38-linux_x86_64.whl cuda-11.7 nvidia_tensorflow-1.15.5+nv22.05-4761017-cp38-cp38-linux_x86_64.whl cuda-11.7 nvidia_tensorflow-1.15.5+nv22.06-5077300-cp38-cp38-linux_x86_64.whl cuda-11.7 nvidia_tensorflow-1.15.5+nv22.07-5236135-cp36-cp36m-linux_x86_64.whl cuda-11.7 nvidia_tensorflow-1.15.5+nv22.07-5236135-cp38-cp38-linux_x86_64.whl cuda-11.7 nvidia_tensorflow-1.15.5+nv22.08-5542765-cp36-cp36m-linux_x86_64.whl cuda-11.7 nvidia_tensorflow-1.15.5+nv22.08-5542765-cp38-cp38-linux_x86_64.whl cuda-11.8 nvidia_tensorflow-1.15.5+nv22.09-6040196-cp36-cp36m-linux_x86_64.whl cuda-11.8 nvidia_tensorflow-1.15.5+nv22.09-6040196-cp38-cp38-linux_x86_64.whl cuda-11.8 nvidia_tensorflow-1.15.5+nv22.10-6183310-cp36-cp36m-linux_x86_64.whl cuda-11.8 nvidia_tensorflow-1.15.5+nv22.10-6183310-cp38-cp38-linux_x86_64.whl cuda-11.8 nvidia_tensorflow-1.15.5+nv22.11-6482588-cp36-cp36m-linux_x86_64.whl cuda-11.8 nvidia_tensorflow-1.15.5+nv22.11-6482588-cp38-cp38-linux_x86_64.whl cuda-11.8 nvidia_tensorflow-1.15.5+nv22.12-6638418-cp36-cp36m-linux_x86_64.whl cuda-11.8 nvidia_tensorflow-1.15.5+nv22.12-6638418-cp38-cp38-linux_x86_64.whl cuda-12.0 nvidia_tensorflow-1.15.5+nv23.01-7142245-cp36-cp36m-linux_x86_64.whl cuda-12.0 nvidia_tensorflow-1.15.5+nv23.01-7142245-cp38-cp38-linux_x86_64.whl cuda-12.0 nvidia_tensorflow-1.15.5+nv23.02-7195399-cp36-cp36m-linux_x86_64.whl cuda-12.0 nvidia_tensorflow-1.15.5+nv23.02-7195399-cp38-cp38-linux_x86_64.whl cuda-12.1 nvidia_tensorflow-1.15.5+nv23.03-7472065-cp36-cp36m-linux_x86_64.whl cuda-12.1 nvidia_tensorflow-1.15.5+nv23.03-7472065-cp38-cp38-linux_x86_64.whl

参考文档：

• 30系显卡tensorflow 1.15安装教程（python3.6 or 3.8/CUDA 11/tf1.15）（离线适用）

• How To Install TensorFlow 1.15 for NVIDIA RTX30 GPUs (without docker or CUDA install)