Training does not use gpu

[JRMeyer]

>>> amir.kovaci
[February 4, 2021, 6:16am]

Hi, slash
I am trying to train a model in a GPU powered machine but the GPUs do
not start and all the work is done on CPUs only. The version of
tensorflow of GPU is installed accordingly (1.15.4) but it does not
work. The version of CUDA is downgraded to version 10 but when I run the
command nvidia-msi it shows the CUDA version 11.2 but the command nvcc
shows the version 10.1, does this has any thing to do with it as the
wrong version might be pointed out?. Tried to uninstall the tensorflow
only version (CPU one) and leave only tensorflow GPU but in this case
training hangs on the beginning on initializing the variables.

Withe the same setup I am able to run the training in another less
powerful machine using GPU.

Thanks in advance for the help.

[This is an archived TTS discussion thread from discourse.mozilla.org/t/training-does-not-use-gpu]

[JRMeyer]

>>> othiele
[February 4, 2021, 7:43am]

CUDA 10.0 and check this
thread.

nvidia-smi will always show the latest installed version, nvcc and the
check the CuDNN.

[Archived Post]