Workshops Archive

Mojo-Driven Performance Portability for LBM


Workshop: 2026 International Workshop on Performance, Portability, and Productivity in HPC

Authors: Giulio Malenza (University of Turin), Matthew A. Smith and Zadok Storkey (University of York), Massimiliano Meneghin (Autodesk Research), Marco Aldinucci and Robert Birke (University of Turin), and Steven A. Wright (University of York)

Abstract: Lattice-Boltzmann methods (LBM) have emerged as a prominent approach for simulating fluid dynamics, owing to their inherent parallelism and suitability for high-performance computing on both multicore CPUs and GPUs. In this study, we investigate Mojo, a low-level framework designed to bridge Pythonic programming with high-performance computing, and compare its performance portability against established parallel programming frameworks such as CUDA, HIP, and SYCL using lid-driven cavity simulations. Our results show that template metaprogramming in Mojo enables high-performance portability without requiring manual loop unrolling. Compared with FluidX3D, Mojo achieves application efficiencies above 80% across GPUs, although its performance is lower on CPUs, resulting in a 0.59 performance portability score for D3Q19-fp32 across all evaluated architectures, programming models and optimisations. These results highlight Mojo's potential as a productive and efficient programming model for heterogeneous computing, but also suggest room for improvement on CPU platforms.


Back to 2026 International Workshop on Performance, Portability, and Productivity in HPC Archive Listing

Back to Full Workshop Archive Listing