2025-07-25 –, Main Room 5
This session provides a brief introduction to our new package, TrixiCUDA.jl, which offers CUDA acceleration support for solving hyperbolic PDEs on GPUs.
Trixi.jl has provided a numerical simulation framework for solving PDEs on CPUs for years, but it lacks GPU support for greater acceleration. TrixiCUDA.jl is designed to address this gap by serving as an acceleration package for Trixi.jl, offering CUDA-based support for solving PDEs on GPUs.
This new package is implemented based on CUDA.jl from JuliaGPU and Trixi.jl from the Trixi-Framework. Its current acceleration efforts focus on the computationally intensive core of the PDE solver: semidiscretization—a high-level description of spatial discretizations specialized for certain PDEs. GPU kernels are implemented to achieve high-performance semidiscretization on GPUs, and minor optimizations are applied to further enhance performance.
The package is currently under active development, testing, and optimization. Future steps will focus on the deep optimization of GPU kernels and initiating the migration of the mesh initialization process to GPUs.
TrixiCUDA.jl on GitHub: https://github.com/trixi-gpu/TrixiCUDA.jl
Comprehensive Documentation: https://trixi-gpu.github.io
Draft Talk Slides (Latest update on Jan 8, 2025): https://trixi-gpu.github.io/assets/files/juliacon25.pdf