Tensor cores provide most of the computing power of modern NVIDIA GPUs, but programming them directly has long meant writing low-level, architecture-specific code. With the 1.0 release of cuTile.jl, that changes. Built on NVIDIA’s CUDA Tile programming model, cuTile.jl lets you write GPU kernels in ordinary Julia that operate on whole tiles of data, while the compiler takes care of tensor-core instructions, data movement and memory layout, on every GPU from Ampere to Blackwell.

In this webinar, we start from a matrix multiplication of about twenty lines that runs within a few percent of cuBLAS. We then move on to the kernels where writing your own pays off: fused GEMM epilogues, attention, mixture-of-experts layers and low-precision formats such as FP8 and FP4. Along the way, we show how cuTile.jl uses Julia’s strengths: generic code through multiple dispatch, Base functions and broadcasting inside kernels, seamless interoperability with CUDA.jl, and tools to inspect and debug every stage of compilation. We close with how cuTile.jl compares to NVIDIA’s cuTile Python and what’s coming next.

Whether you’ve seen our earlier introduction to cuTile.jl or are new to GPU programming in Julia, you’ll leave knowing how to write, verify and tune your own tensor-core kernels.