conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0ĬPU Note: Starting with TensorFlow 2.10, Windows CPU-builds for x86/圆4 You can get the latest update from here:įor CUDA in WSL. This corresponds to Windows 10 version 21H2, the November 2021 Windows WSL2 Note: TensorFlow with GPU access is supported for WSL2 on Windows 10 19044 or Python -c 'import tensorflow as tf print(tf.config.list_physical_devices('GPU'))' # Anything above 2.10 is not supported on the GPU on Windows Native TensorFlow-DirectML-Plugin conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0 Or install tensorflow-cpu and, optionally, try the Starting with TensorFlow 2.11, you will need to install Windows Native Caution: TensorFlow 2.10 was the last TensorFlow release that Python3 -c 'import tensorflow as tf print(tf.reduce_sum(tf.random.normal()))' MacOS # There is currently no official GPU support for MacOS. Python3 -c 'import tensorflow as tf print(tf.config.list_physical_devices('GPU'))' conda install -c conda-forge cudatoolkit=11.2.2 cudnn=8.1.0Įxport LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/ Seeįor more information about this collaboration. The third party fails to release the pip package. The availability and integrity of this pip package. Tensorflow will use reasonable efforts to maintain Processors are built, maintained, tested and released by a third party: Linux Note: Starting with TensorFlow 2.10, Linux CPU-builds for Aarch64/ARM64 LD_LIBRARY_PATH=/usr/lib64/nvidia-bumblebee primusrun.

So, that looks something like this for the nbody example: I’m not sure if this is required on more basic programs yet. Contrary to what the Bumblebee documentation suggests, you must use optirun or primusrun to execute the samples.It was necessarily to override the library path to point to Bumblebee’s version of the cuda shared library instead of a 32-bit version that somehow ended up in /usr/lib.Aside from that, compilation and execution were simple enough with two catches: You’ll notice that I told the gcc configuration during build time to add the 49 suffix with the –program-suffix=49 option. I compiled GCC 4.9 from scratch and then manually modified the makefiles of each sample I wanted to compile to point to “g++49” instead of “g++”. I had trouble getting the examples to compile because Fedora 23, which I’m on, defaults to gcc 5.3.1 and, unfortunately, CUDA 7.5 requires gcc 4.9 or less. So, instead you can just runĪdditional configuration works as specified in the User’s Manual and Getting Started Guides. It is important to only install the Toolkit from this repository, not the version of CUDA that is provided since Bumblebee already configured an appropriate Optimus-ready NVIDIA driver. I installed the CUDA Toolkit using the dnf repository provided by NVIDIA. Primus: warning: timeout waiting for display worker Primus: warning: dropping a frame to avoid deadlock Using both optirun and primusrun works, however on glxgears they report different numbersġ0162 frames in 5.0 seconds = 2032.270 FPS I used the managed proprietary option to get the latest NVIDIA driver for my system. Getting CUDA 7.5 to work was straightforward using Bumblebee, as described on the Fedora Bumblebee Page. I have a Dell XPS 15 that has both Intel and NVIDIA graphics using NVIDIA’s Optimus Technology:Ġ0:02.0 VGA compatible controller: Intel Corporation HD Graphics 530 (rev 06)Ġ1:00.0 3D controller: NVIDIA Corporation GM107M (rev a2)

Nvidia cuda toolkit 7.5