@inproceedings{hoerold_raptor_2025,
  title = {{{RAPTOR}}: {{Numerical Profiling}} of {{Scientific Applications}}},
  booktitle = {Proceedings of the {{International Conference}} for {{High Performance Computing}}, {{Networking}}, {{Storage}} and {{Analysis}}},
  author = {Hoerold, Faveo and Ivanov, Ivan Radanov and Dhruv, Akash and Moses, William S. and Dubey, Anshu and Wahib, Mohamed and Domke, Jens},
  date = {2025-11},
  series = {{{SC}} '25},
  publisher = {IEEE Press},
  abstract = {The proliferation of low-precision units in modern high-performance architecture increasingly burdens the domain scientists. Historically, the choice in HPC was easy: can we get away with 32bit floating-point operations and lower bandwidth requirements, or is FP64 necessary? Driven by Artificial Intelligence, vendors introduced novel low-precision units for vector and tensor operations. Technology roadmaps point to a future of stagnating, or even a reduction in, FP64 support. Hence, scientists have to re-evaluate their codes, but a simple search-and-replace approach to go from FP64 to FP16 will not work. Hence, we introduce RAPTOR: a numerical profiling tool to guide the scientists in their search for code regions where precision lowering is feasible. Using LLVM, we transparently replace high-precision computations using low-precision units, or emulate user-defined lengths of exponent and mantissa. Our RAPTOR is the first feature-rich approach---with focus on ease of use---to change, profile, and reason about numerical requirements and instabilities.},
  keywords = {deep neural networks,GPUs,out-of-core}
}