Supported by tensorflow-gpu builds located on PyPI This is an article that just makes a note of the CUDA and cuDNN versions.
I will add it to this article after confirming the operation with the new version.
This page is made by an individual as a personal memo Please see for reference only when there is no official description of tensorflow.
tensorflow official https://www.tensorflow.org/install/source#tested_build_configurations
Windwows10
| tensorflow | CUDA | cuDNN | Remarks |
|---|---|---|---|
| 2.3.0 | 10.1 | 7.6 | Check the operation in the hand environment |
| 2.2.0 | 10.1 | 7.6 | Check the operation in the hand environment |
| 2.1.0 | 10.1 | 7.6 | Described in Release Note |
| 2.0.0 | 10.0 | 7.4 | Described on the official website |
Ubuntu 18.04
| tensorflow | CUDA | cuDNN | Remarks |
|---|---|---|---|
| 2.3.0 | 10.1 | ||
| 2.2.0 | 10.1 | 7.6 | Check the operation in the hand environment |
| 2.1.0 | 10.1 | 7.6 | Described in Release Note |
| 2.0.0 | 10.0 | 7.4 | Described on the official website |
The tensorflow-gpu build installed by pip install now references a specific version of CUDA, so it won't work if you include a different version of CUDA.
Which version of tensorflow-gpu is built to use which version of CUDA here for older versions There is no problem because it is listed in>, but as of August 2020, tensorflow 2.3.0 has been released, but only 2.1.0 is listed in the list. https://www.tensorflow.org/install/source#tested_build_configurations
Then, if it is listed in Release Notes, although the corresponding CUDA version is written in tensorflow 2.1.0, it is not listed in Release Notes after tensorflow 2.2. https://github.com/tensorflow/tensorflow/releases
This is a problem, so if the new version works properly, I will write down the version at that time in this article. Instead of writing a new article about the new version, I will add it here.
If CUDA is not included or the version is not appropriate, the following error message will be displayed when importing. You can tell the version by the file name of the dll displayed at that time. Below, I'm trying to read cudart64_101.dll, so I feel like I probably need CUDA 10.1.
python
>>> import tensorflow as tf
2020-08-05 00:33:55.037723: W tensorflow/stream_executor/platform/default/dso_loader.cc:59] Could not load dynamic library 'cudart64_101.dll'; dlerror: cudart64_101.dll not found
2020-08-05 00:33:55.037919: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.
Windows10 Anaconda3 Ryzen7-1700X GTX-1080Ti
Ubuntu18.04 Anaconda3 core i9-9900K TitanV
Recommended Posts