2026-08-13 –, Room 3
CUDA is well known for its SIMT programming model, available in Julia through CUDA.jl. This year, NVIDIA introduces cuTile, a new tile-based programming model for writing high-performance GPU kernels, with automatic tensor core utilization. cuTile.jl brings this model to Julia, compiling Julia kernels through a custom pipeline to Tile IR bytecode. In this talk, we'll cover the programming model, the compiler design, and performance benchmarks on Blackwell GPUs.
CUDA Tile is a significant evolution in GPU programming, so it's important that Julia GPU developers have access to it. This talk will introduce the cuTile programming model, why it matters, and how cuTile.jl makes it possible to write tile-based GPU kernels directly in Julia.
Specifically, the talk will cover:
- The programming model: what tile-based programming is, how it abstracts over threads and warps, and how it enables automatic tensor core utilization.
- The compiler pipeline: what the Tile IR bytecode is, and how we target it from Julia using a custom compiler.
- Performance benchmarks: how cuTile.jl performs on Blackwell GPUs, and how it compares to Python's cuTile.
- The relationship between cuTile.jl and CUDA.jl, and how they can complement each other for different types of GPU programming.
Tim Besard is a software engineer at JuliaHub, where he leads GPU support and development for the Julia programming language. He holds a Ph.D. in computer science engineering from Ghent University, Belgium, and has been a key contributor to Julia's GPU ecosystem since 2014. Tim maintains several foundational GPU packages including CUDA.jl, GPUArrays.jl, GPUCompiler.jl, and LLVM.jl, which together form the backbone of GPU computing in Julia.