Developing high-performance custom GPU kernels is critical for pushing the boundaries of AI inference, yet it presents a steep learning curve and significant optimization bottlenecks. In this session, we will walk through the process of working with Triton, an open-source, python-like language for writing highly efficient custom kernels.

Kyle Yu will cover the essentials of writing Triton kernels, highlighting common pitfalls he experienced while developing kernels for BLAS routines. The session will also explore the cutting-edge realm of AI-generated kernels, demonstrating how this innovative approach can accelerate development and potentially unlock new levels of performance. Join us to leverage Triton effectively and explore the future of kernel development.