Comer, Kirkham, & Ratzel - New CUDA Toolkit packages for Conda | SciPy 2023

Ai

Learn about new CUDA Toolkit packages for Conda, empowering scientific computing with NVIDIA's GPU-accelerated solutions, and discover how to easily install, manage, and specify CUDA versions and tools using Conda.

Key takeaways
  • CUDA is a suite of application tools that allows you to take advantage of GPUs, developed by NVIDIA.
  • The CUDA Toolkit is a large package that includes a variety of applications and tools.
  • Conda is a package manager that allows you to easily install and manage packages for scientific computing.
  • The first generation of CUDA Toolkit packages for Conda were released in 2018.
  • The default channel is curated by Anaconda and includes a variety of packages, but it does not include all packages.
  • The NVIDIA channel, starting with CUDA 11.3, has a new hierarchy that allows for more flexible and easier installation of CUDA Toolkit packages.
  • The CUDA Toolkit packages are now available via Conda.
  • The CUDA version meta package is used to align the CUDA version across all dependent packages.
  • The new hierarchical structure of the CUDA Toolkit packages reduces the install footprint and makes it easier to install specific components.
  • The run constraint block is used in the CUDA version meta package recipe to specify the required CUDA version.
  • The NVIDIA images for Conda include the CUDA environment and pull from CUDA.
  • The RAPIDS AI channel is useful for Python developers who want to use AI-accelerated libraries.
  • The dependencies of the CUDA Toolkit packages have been updated to align with the new hierarchical structure.
  • The user can specify the CUDA version using the CUDA version meta package.
  • The user can also specify the CUDA tools they need, such as the debugger and profiler.