Graham Markall - Accelerated Python with Numba

Numba is a Python compiler focused on accelerating workloads common in the Data Science, AI, and HPC ecosystems. It supports compiling many numerical functions as part of its basic functionality, but is also highly extensible and embeddable. It sits within an ecosystem of related tools and libraries; composability of Numba with these other frameworks is a key feature.

Many users use it directly in their Python code to improve performance whilst maintaining a single source language for an entire codebase. In other use cases, library and application developers enable their users to customise applications by writing User-Defined Functions (UDFs) in Python that are compiled by Numba and executed as part of an application workflow.

Numba also enables Python functions to run on NVIDIA GPUs, both standalone and in conjunction with accelerated CUDA C/C++ libraries. It is used for custom workflows in RAPIDS, NVIDIA, and third-party libraries including cuDF, DALI, and NeMo.

This talk will give an overview of the use of Numba and how it works, summarising the pipeline from an interpreted, duck-typed language to compiled, statically-typed implementations. Some tips on using Numba both as a straightforward Python compiler and as an embedded UDF compiler will be given, with the aim that attendees will feel well-placed to start using Numba to accelerate their applications.