I am really sorry for coming here with common problem. I spend monthes to find correct configuration to connect my notebuuk to laptop NVIDIA. I tried to use suggestion that worked for other people, but it is still not working for me (I am really bad in python configurations). Please, help me =)
Following reccomendations and notebook requirenments I manage to recognize my GPU in notebook, but when I start model training using GPU in Jupyter notebook, kernel duy emidiatly. I belive I have poor compatability between CUDU, CuDNN and TF.
Here is what I am have installed:
cudatoolkit : 11.8.0
tensorflow==2.4.0 and tensorflow-gpu==2.4.0
NVIDIA-SMI 535.98 Driver Version: 535.98 CUDA Version: 12.2
(base) C:>nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Sun_Jul_28_19:12:52_Pacific_Daylight_Time_2019
Cuda compilation tools, release 10.1, V10.1.243
I also am currently having difficulty with Jupyter+GPU.
tf.config.list_physical_devices(‘GPU’) will list the GPU if it is running in a simple python terminal session or in a PyCharm IDE, but produces (i.e. empty list) when run inside a Jupyter notebook. Note, tf.test.is_built_with_cuda() produces “True” in all 3 scenarios.
I have a conda environment under Windows and I am using versions exactly as per (Install TensorFlow with pip) for Windows, i.e. cudatoolkit=11.2 and cudnn=8.1.0 installed using conda, and tensorflow=2.10.1 installed using pip.
Does Jupyter need special configuration to enable use of GPU?
Thanks
I believe I have located my error. I had not setup the ipykernel correctly in the conda environment. I could not figure out how to setup the kernel spec manually, so I have used Option 3. of https://stackoverflow.com/questions/58068818/how-to-use-jupyter-notebooks-in-a-conda-environment which suggests adding nb_conda_kernels to the base environment and ipykernel to each conda environment. This automatically create the correct jupyter kernelspec for each conda environment.
Having done this, my installation of cudatoolkit=11.2 cudnn=8.1.0 and tensorflow=2.10.1, as per the tensorflow installation page, works perfectly in my jupyter notebook and uses the GPU as expected.
I just managed to connect notebook to GPU. Following suggestion found on forum I created in my working directory new folder and subfolder .\nvvm\libdevice and copied inside files (libdevice.10.bc and nvvm64_40_0.dll) that I found in C:\Users\username\AppData\Local\anaconda3\pkgs\cudatoolkit-11.8.0-hd77b12b_0\DLLs