GPU-accelerated Python 3 kernel for NVIDIA Optimus

This small hack enables an easy way to switch to GPU accelerated hardware with NVIDIA’S Optimus hardware. This gear can be found in laptops where two different GPUs are provided: an integrated Intel GPU and a dedicated, more powerful NVIDIA GPU. These provides a less powerful, efficient GPU for normal use and a more capable and energy-demanding one.

Although Windows takes care of this by itself, GNU/Linux systems usually require greater configuration to work. A common alternative is to use bbswitch and bumblebee, which allow seamless switching between cards when properly configured. These utilities provide the optirun command, which turns on and off the dedicated GPU just for the command that is being invoked into. For that reason, using Jupyter notebooks with the NVIDIA GPU can be achieved by running optirun jupyter notebook. However, this requires the server to be completely restarted if GPU usage is required. On the other hand, running this command every time can be produce increased battery consumption even if it is not needed.

For that reason, this kernel can provide an easier way and better granularity for notebooks. To install this kernel, run the following commands:

$ git clone https://github.com/DiegoVicen/gpu-python3-kernel.git
$ cd gpu-python3-kernel
$ jupyter kernelspec install --user optirun_python3

This should be enough configuration to enable a Jupyter Kernel with GPU support that can be assigned to certain notebooks and provide seamless execution while minimizing the power consumption. You can check if it has been correctly installed by running this command:

$ jupyter kernelspec list
Available kernels:
  optirun_python3    <anaconda>/jupyter/kernels/optirun_python3
  python3            <anaconda>/jupyter/kernels/python3

And then selecting the kernel “Python 3 [GPU]” in the notebook once it is necessary.

This is a hacky solution and probably not a general enough one: some modification on your side may be required to adapt it to your requirements. This code is licensed under MIT so feel absolute freedom to modify or redistribute it. If you run into any problem or have any suggestion, do not hesitate to create a new issue.