I just got a GEFORCE GTX 1660 TI GPU. I installed CUDA and cuDNN. I installed tensorflow-gpu. However tensorflow is still running on the CPU. Please please help.
-Laura
Hi Laura,
Can you share the output of nvidia-smi to double check that your GPUs/CUDA are setup properly? Also please share the following info about your environment:
- CUDA version
- CUDNN version
- tensorflow version (pip freeze | grep -i tensorflow)
- tensorflow-gpu version (pip freeze | grep -i tensorflow)
- Operating System version
- Python version
One common solution to these kinds of problems is to make sure your GPUs are exposed to the CUDA/TF libraries like so if running on a linux machine (bash). I believe this would be setup correctly by default if you rebooted your machine, but here’s how you’d do it manually:
# This is for a single-gpu machine. 0,1 for 2 gpus, 0,1,2,3 for 4 gpus, etc.
export CUDA_VISIBLE_DEVICES=0
Then try running your tensorflow code again to see if it uses your GPUs.
It’s also possible that you have an unsupported combination of CUDA + Tensorflow, which is defaulting to CPU. For example, there might not be any tensorflow pip packages built for CUDA 10.2 yet, so tensorflow would default to CPU. You can resolve a lot of these config/version headaches by using nvidia-docker containers, such as these which will have everything setup for you inside the container: TensorFlow | NVIDIA NGC
Related resources: