CODEWITHSUNDEEP
tensorflow
Sungmin Lee
Sungmin Lee

Reputation: 13

Training with TF 1.15 on RTX 3090

During I test several cases, I have some questions. One of that is "Training with tf 1.15 on RTX 3090".

[MY current environments]

  1. python : v3.7.9
  2. tensorflow : v1.15.5
  3. cuda : v11.2
  4. cudnn : v8.1.0
  5. os : window 10

Can I proceed with the training in this environment?

Upvotes: 0

Views: 5154

Answers (3)

bastien enjalbert
bastien enjalbert

Reputation: 341

If you can't upgrade your code base to 2.X TF. You may want to use the nvidia-tensorflow version which is basically a nvidia maintained version of tensorflow 1.15 which is compatible with CUDA >11 (and so on RTX 30 gpus).

You'll find more information here : https://developer.nvidia.com/blog/accelerating-tensorflow-on-a100-gpus/

TL;DR

  • Eventually source your python venv
  • Remove tensorflow==1.15 dependency
  • pip install nvidia-pyindex
  • pip install nvidia-tensorflow[horovod]

Then you're good to go.

Upvotes: 3

mathandy
mathandy

Reputation: 2010

The following worked for me (and my 3090 w/ CUDA 11.3).

# create new conda env
conda create --name nv-tf15

# install nvidia-version of TF 1.15
pip install nvidia-pyindex
pip install nvidia-tensorflow

# install pytorch (likely optional)
pip install torch==1.10.2+cu113 torchvision==0.11.3+cu113 torchaudio==0.10.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

# tensorboard wasn't importing properly, so I did this
conda install tensorboard

Upvotes: 1

Filippo Grazioli
Filippo Grazioli

Reputation: 385

I believe this cannot work. Tensorflow 1.15 requires cuda 10:

https://www.tensorflow.org/install/source#gpu

I believe the RTX 3090 does not support cuda 10.

You might need to use either TF 1.15 on CPU. Or a 2.x TF release with a more recent cuda version.

Upvotes: -1

Related Questions