Bit-reproducibilty on GPU-accelerated Fortran

117 · Fortran Lang Discourse · July 27, 2026, 3:19 p.m.
Summary
The blog post discusses achieving bit-reproducibility in GPU-accelerated Fortran through the implementation of bitwise-reproducible algorithms. By using OpenACC and a Fortran interface, the author maintains consistent results between CPU and GPU calculations, noting a performance penalty and sharing insights on the challenges and future testing with CUDA/HIP. The source code is available on GitHub. The author seeks feedback from others who may have explored similar techniques.