显卡驱动版本 CUDA 版本官方详解 cudnn gcc工具包版本约束依赖关系 tensorflow gpu 版本

今天药忘吃喽~ 2022-01-22 23:15 468阅读 0赞

驱动是向下兼容的：所以如果想使用最新的tf代码库，那么需要最新的驱动，并且旧的虚拟环境还是能用的。显卡驱动，要在考虑自己的系统下，选择最新的版本。

<table>  Table 1. CUDA Toolkit and Compatible Driver Versions  
 <thead> 
  <tr> 
   <th colspan="1">CUDA Toolkit</th> 
   <th colspan="1">Linux x86_64 Driver Version</th> 
   <th colspan="1">Windows x86_64 Driver Version</th> 
  </tr> 
 </thead> 
 <tbody> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 10.1 Update 1</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 418.67</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 425.25</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 10.1.105</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 418.39</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 418.96</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 10.0.130</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 410.48</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 411.31</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 9.2 (9.2.148 Update 1)</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 396.37</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 398.26</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 9.2 (9.2.88)</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 396.26</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 397.44</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 9.1 (9.1.85)</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 390.46</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 391.29</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 9.0 (9.0.76)</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 384.81</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 385.54</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 8.0 (8.0.61 GA2)</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 375.26</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 376.51</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 8.0 (8.0.44)</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 367.48</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 369.30</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 7.5 (7.5.16)</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 352.31</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 353.66</td> 
  </tr> 
  <tr> 
   <td colspan="1" style="vertical-align:top;">CUDA 7.0 (7.0.28)</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 346.46</td> 
   <td colspan="1" style="vertical-align:top;">&gt;= 347.62</td> 
  </tr> 
 </tbody> 
</table>

cuda可以安装什么版本取决于驱动，所以如上图所述，你有最新的驱动，你可以安装任意的虚拟环境。anaconda，docker，都是没问题的。docker使用虚拟系统的时候注意下驱动版本是否支持。

<table> 
 <tbody> 
  <tr> 
   <th> <p>&nbsp;</p> <p>版本</p> </th> 
   <th>Python 版本</th> 
   <th>编译器</th> 
   <th>编译工具</th> 
   <th>cuDNN</th> 
   <th>CUDA</th> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.13.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.19.2</td> 
   <td>7.4</td> 
   <td>10.0</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.12.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.15.0</td> 
   <td>7</td> 
   <td>9</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.11.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.15.0</td> 
   <td>7</td> 
   <td>9</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.10.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.15.0</td> 
   <td>7</td> 
   <td>9</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.9.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.11.0</td> 
   <td>7</td> 
   <td>9</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.8.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.10.0</td> 
   <td>7</td> 
   <td>9</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.7.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.9.0</td> 
   <td>7</td> 
   <td>9</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.6.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.9.0</td> 
   <td>7</td> 
   <td>9</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.5.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.8.0</td> 
   <td>7</td> 
   <td>9</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.4.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.5.4</td> 
   <td>6</td> 
   <td>8</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.3.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.4.5</td> 
   <td>6</td> 
   <td>8</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.2.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.4.5</td> 
   <td>5.1</td> 
   <td>8</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.1.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.4.2</td> 
   <td>5.1</td> 
   <td>8</td> 
  </tr> 
  <tr> 
   <td>tensorflow_gpu-1.0.0</td> 
   <td>2.7、3.3-3.6</td> 
   <td>GCC 4.8</td> 
   <td>Bazel 0.4.2</td> 
   <td>5.1</td> 
   <td>8</td> 
  </tr> 
 </tbody> 
</table>

至于cuda，cudnn gcc以及多个工具包的版本依赖关系，这个完全没必要考虑，**虚拟环境安装软件（anaconda，docker） 会把python api版本依赖统统考虑到。**

综上所述，根据系统安装最新的显卡，利用虚拟环境 根据你想要的tf安装需要的cuda。这里提一句，TF实际上也是向下兼容的，但是不向上兼容，所以一切安装最新的即可。

**所以不管你用什么深度学习的工具，驱动版本一定要保持最新。**

参考：[https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html][https_docs.nvidia.com_cuda_cuda-toolkit-release-notes_index.html]

[https://www.tensorflow.org/install/source\#linux][https_www.tensorflow.org_install_source_linux]

[https_docs.nvidia.com_cuda_cuda-toolkit-release-notes_index.html]: https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html
[https_www.tensorflow.org_install_source_linux]: https://www.tensorflow.org/install/source#linux