GPU programming is difficult. As GPU programmers, not only do we have to understand complex hardware with unique performance features, but we commonly have to learn and use novel programming models and languages like HIP and CUDA. On top of all that, we want to write high-quality code that can run on many types of hardware, that our colleagues can understand, and that produces results that we can trust. So how do we achieve these goals without losing our minds? In short, we choose to care about software engineering. We learn how to design and profile and test GPU codes well. This collection of modules aims to elevate your existing understanding of GPUs beyond the mechanical aspects of writing CUDA kernels to knowing how to write good GPU programs.

What is in the course?

We've chosen to focus on a few key aspects of advanced GPU programming:

  • Using GPU-accelerated libraries
  • Effective testing
  • Design and architecture
  • Profiling and optimisation

In general, each module presents its content through some written material you are expected to read, questions for you to check your understanding, illustrative examples and hands-on exercises to help you explore and gain experience in the concepts.

How do I go through the course?

You are welcome to access the modules in any order but we recommend you start with the refresher module and work through the modules in order.

Who is this course for?

This course is not intended for total beginners in GPU programming but if you are reading this as a total beginner, we recommend you first look at the refresher section and see how you get on. If you want a more guided introduction to GPUs and CUDA/HIP, we recommend a course like DiRAC's Many Ways to GPU, focusing particularly on the CUDA section.

This course is intended for C++ programmers. If you are a Fortran programmer looking to accelerate your code with GPUs, our recommendation (as of 2026) is to use OpenMP target offloading. For Python programmers, we recommend you look into CuPy, Numba and JAX. We provide some examples in the refreshed module.

What hardware will I learn about?

We have tried to make these modules vendor agnostic where possible so all terms should be introduced with both AMD and Nvidia equivalents. However, Nvidia and CUDA remain dominant in the market and in our experience, so we may use the Nvidia terms more often.

Nearly all exercises and examples are CUDA focused but CUDA and HIP and nearly interchangeable and the tools and concepts transfers easily.

Last modified: Monday, 16 November 2026, 4:20 PM