GPU Programming: A day-long workshop by Verda

Low-level GPU programming is often framed around setting up matrix-multiplication pipelines. While those dominate modern AI workloads, what's typically under-represented is everything outside of them, namely, memory-bound kernels. Verda and Erik Schultheis (GPU MODE) are hosting a one-day intensive GPU programming course aimed at bridging the gap between writing performant memory-bound kernels on consumer GPUs versus on datacenter GPUs. The course is designed to be interactive: participants will be able to run their code live on Verda's infrastructure, including RTX 6000, B200, and H200.
The one day course is intended for anyone with a general interest in GPU programming and will be exclusively in CUDA, and knows the basics of C. The agenda is as follows:

9:30 - 10:00 - Introduction + setup.
10:00 - 12:00: Introduction to GPU programming: CPU v GPU, Memory Hierarchy, Execution Hierarchy.
12:00 - 13:00: Lunch.
13:00 - 17:30: Advanced topics: Warp-level programming, Vectorized Memory Access, Asynchronous Data Movement.
The idea of the day is to be interactive with quick fire concepts then explored through hands-on exercises!

P.S. Feel free to skip the intro morning session if you already have some experience with CUDA, just let us know in the registration.