Workshop: 2026 International Workshop on Performance, Portability, and Productivity in HPC
Authors: Jonas Schuhmacher and Hans-Joachim Bungartz (Technical University of Munich)
Abstract: We benchmark 14 GPU programming paradigms—more than in any prior comparison under a common performance-portability metric—on six GPUs from NVIDIA, AMD, and Intel, across four workloads: vector addition, matrix multiplication, and two scientific applications driving the GPU migration of the particle simulation AutoPas and ESA's polyhedral gravity model. We track performance portability across problem sizes, alongside SLOC and Halstead difficulty—a vocabulary-based, paradigm-aware code-complexity proxy—and explain key results with kernel profiles. Four conclusions arise: (i) abstraction need not cost performance—Kokkos, Boost.Compute, and RAJA provide the strongest portable results for complex scientific kernels; (ii) OpenCL is verbose but exceptionally dependable, running on all six platforms, often among the leaders in performance; (iii) implicit parallelism and directive-based offloading convince only for simple kernels; (iv) graphics APIs and shading languages are viable only on consumer GPUs—their kernels match CUDA, but they pay with per-dispatch latency and the highest code complexity.
Back to 2026 International Workshop on Performance, Portability, and Productivity in HPC Archive Listing
Back to Full Workshop Archive Listing