Minimal-Precision Computing for High-Performance, Energy-Efficient, and Reliable Computations

Daichi Mukunoki (1) , Imamura Toshiyuki (1) , Yiyu Tan (1) , Atsushi Koshiba (1) , Jens Huthmann (1) , Kentaro Sano (1) , Fabienne Jézéquel (2, 3) , Stef Graillat (3) , Roman Iakymchuk (3, 4) , Norihisa Fujita (5) , Taisuke Boku (6)

1 RIKEN CCS - RIKEN Center for Computational Science [Kobe]
2 UP2 - Université Panthéon-Assas
3 PEQUAN - Performance et Qualité des Algorithmes Numériques
4 Fraunhofer ITWM - Fraunhofer Institute of Industrial Mathematics
5 Université de Tsukuba = University of Tsukuba
6 Graduate School of Systems and Information Engineering [Tsukuba]

We propose a new systematic approach for minimal-precision computations. This approach is reliable, general, comprehensive, high-performance, and realistic. Although the proposed system is still in development, this presentation shows that the system could be constructed by combining already available (developed) in-house technologies as well as extending them.

Domaines: Calcul parallèle, distribué et partagé [cs.DC]; Arithmétique des ordinateurs; Analyse numérique [math.NA]; Optimisation et contrôle [math.OC]; Logiciel mathématique [cs.MS]

Daichi Mukunoki, Imamura Toshiyuki, Yiyu Tan, Atsushi Koshiba, Jens Huthmann, et al.. Minimal-Precision Computing for High-Performance, Energy-Efficient, and Reliable Computations. France-Japan-Germany trilateral workshop: Convergence of HPC and Data Science for Future Extreme Scale Intelligent Applications, Nov 2019, Tokyo, Japan. ⟨hal-02616486⟩