Performance and Numerical Aspects of Decompositional Factorizations with FP64 Floating-Point Emulation in INT8
Mixing precisions for performance has been an ongoing trend as the modern hardware accelerators started including new, and mostly lower-precision, data formats. The advantage of using them is a great potential of performance gain and energy savings. The disadvantage are the numerical issues not present in the standard-mandated floating-point formats. Split integer emulation of FP64 takes this to an extreme with the computation performed only by fixed-point tensor core units. We present the new issues the emulation faces for practical cases involving dense linear solver. We show extensive numerical tests indicating the effect of extended numerical range of matrix entries. We also scaled the input sizes to study the performance and numerical profiles on the NVIDIA Hopper GPUs.
Authors
- Albert Reuther (ORCID: https://orcid.org/0000-0002-3168-3663)
- Julia Mullen (ORCID: https://orcid.org/0000-0002-0015-6182)
- Michael E. Houle (ORCID: https://orcid.org/0000-0001-8486-8015)
- William Arcand
- Michael T. Jones (ORCID: https://orcid.org/0000-0002-2365-2525)
- William Bergeron
- Jeremy Kepner (ORCID: https://orcid.org/0000-0001-9668-2613)
- David Bestor (ORCID: https://orcid.org/0009-0002-7684-1191)
- Piotr Łuszczek (ORCID: https://orcid.org/0000-0002-0089-6965)
- Daniel J. Burrill
- Matthew Hubbell
- Andrew Prout (ORCID: https://orcid.org/0000-0002-5137-0672)
- Chansup Byun (ORCID: https://orcid.org/0009-0003-0183-914X)
- Guillermo Morales
- Vijay Gadepally (ORCID: https://orcid.org/0009-0004-5782-7137)
- Antonio Rosa (ORCID: https://orcid.org/0000-0002-4595-1999)
- Larry W. Anderson
- Charles Yee
- Peter Michaleas
- Alex Bonn
Institutions
- MIT Lincoln Laboratory (US)
Publication Details
- Published
- 2025-09-15
- DOI
- https://doi.org/10.1109/hpec67600.2025.11196192
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00