Local machine with dedicated GPU for computing#

Do’s:

  • Move on to the way numpy handles things quickly. Most libraries going the extra mile depend on the well established numpy api.

  • Choose the hardware and OS wisely. Prefer UNIX like systems

  • Use libraries, such as numba first to move on in baby steps. Use CPU backend first, the proceed to the GPU backend.

  • Use specialized libraries to tweak performance to the limits. This will cost time at get you to the borde of what is considered to be pythonic. Only do it if the task and algorithm is worth it. Prior implementations should have made money before!

  • Learn how to think in kernels and stencils. E.g. write algorithms which can be applied to each element of an array.

Dont’s:

  • Start off with the newest kid in town first everyone is tweeting about. It’ll generate a mess in production environments.

  • Don’t leave out the initial steps recommended.