Training with GPU memory but with 0% GPU utilization

Recently, I am trying to use GPU to train a model with CUDA and WSL2 to speed up the training process.

However, in my experiment, I found only my GPU’s memory was used, but the utilization of the GPU is almost 0% and computation seemed to be done by CPU. I would be very grateful if someone can help me out.

BTW, I am using TensorFlow -2.3.1, Cuda 10.1, and CuDNN 7.

Hi @fanfandouble ,

Could you please share the sample model and script so that we can help better?


Could be related Training WSL 2 CUDA hangs over several training steps - #6 by Cchivriga.