Arch Linux

joanmanel

Member

Registered: 2012-11-06

Posts: 250

GPU (RTX 2070 Super) not working with Tensorflow

Hi All,

Around 1 year ago I was using this computer (with an RTX 2070 Super gpu) for Tensorflow. As far as I remember everything worked out of the box. Everything was installed using pacman (cuda, cudnn, tensorflow, ...). I don't remember having any problems, and everything worked fine.

I am trying it again now, to run Tensorflow in this computer (after 1 year hiatus). The computer is fully updated, and it has Tensorflow 2.11, Cuda 11.8 and cudnn 8.6.0. Checking the compatibility table for Tensorflow 2.11, it doesn't recommend these versions of cuda and cudnn. But I do remember one year ago, my versions of these libraries were always more updated than the recommended ones, and it worked fine.

Tensorflow works but it does not detect the GPU. Out of the blue as I do "import tensorflow as tf" I get:

2023-01-16 15:08:17.957602: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: SSE3 SSE4.1 SSE4.2 AVX AVX2 AVX512F AVX512_VNNI FMA To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.

When running things that use a GPU, I don't see anything going in nvidia-smi. And when doing

tf.test.is_built_with_cuda()

it returns False.

Installing Tensorflow through conda with "conda create -n tf -c conda-forge cudatoolkit tensorflow" seems to work fine. The previous command (tf.test.is_built_with_cuda()) returns a true. And when I create a model in Tensorflow, I can see it loaded in nvidia-smi.

I guess this works and I can continue with that. But why the stock installation of TF, Cuda and Cudnn does not work? Checking what Conda installed, it is TF 2.11, cuda 11.8 and cudnn 8.4.1. Is CuDNN breaking it all? I tried another simpler Conda environment, just witn tensorflow

Installing a simpler conda environment, with "conda create --name myenv" and "conda install tensorflow" installs tensorflow 2.10.0, and this doesn't work.

Doing instead (after creating the empty conda environment) "pip install tensorflow" installs 2.11 and again it doesnt work, but if I do "pip install tensorflow=2.10.0" install TF 2.10.0 and it works fine. And I guess it is using cuda 11.8 and cudnn 8.6.