Workshop: 2026 International Workshop on Performance, Portability, and Productivity in HPC
Authors: Andrew Morgan (University of Bristol, AWE Nuclear Security Technologies)
Abstract: Developing high-performance computing codes is challenging: exploiting modern hardware requires reasoning about architectural detail, and heterogeneity makes this harder still. Even portability frameworks leave much of a code specialised: across five C++ implementations of the same mini-application, only 16% of the source is shared by all backends. Languages such as Chapel/Julia promise a single-source approach, targeting CPUs and GPUs from multiple vendors while trading some control for productivity and portability. We implement BabelStream, CloverLeaf and TeaLeaf in Chapel and Julia, and compare them against CUDA, HIP and OpenMP C++ baselines across diverse hardware using the Pennycook performance-portability metric. Across three GPU platforms, CloverLeaf achieves P=89% (Julia) and 86% (Chapel); on TeaLeaf the ordering inverts, falling to 77% (Chapel) and 62% (Julia). Both languages need 45-64\% fewer lines of code, and where the C++ backends diverge by 0.63 on average, Chapel and Julia share almost all source between CPU and GPU.
Back to 2026 International Workshop on Performance, Portability, and Productivity in HPC Archive Listing
Back to Full Workshop Archive Listing