Isolating GPUs (without Docker) ΒΆ

You can isolate the GPU that your software will be able to see using a NVIDIA environment called CUDA_VISIBLE_DEVICES, where you can select which GPUs will be visible to your software. Let’s say you want to execute my_script.py and you want only the GPU 0 (zero) to be visible for your process:

  CUDA_VISIBLE_DEVICES=0 python my_script.py

If you execute your script with this environment set, your script will only be able to see the GPU 0.