Speaker
Dr
Issaku Kanamori
(R-CCS, RIKEN)
Description
We present a stabilized FP16 mixed-precision solver. Because the dynamic range of FP16 is much smaller than that of FP32, straightforward use of FP16 arithmetic may suffer from underflow and overflow. To address this issue, we introduce a rescaling procedure for the working vectors. The performance of the solver is evaluated on the Supercomputer Fugaku, which provides native FP16 SIMD support, and compared with that of a conventional FP32-based mixed-precision solver.
Authors
Dr
Issaku Kanamori
(R-CCS, RIKEN)
Tatsumi Aoyama
(U. of Tokyo)
Kazuyuki Kanaya
(University of Tsukuba)
Hideo Matsufuru
(KEK)
Yusuke Namekawa
(Fukuyama University)
Hidekatsu Nemura
(Osaka U.)
Keigo Nitadori
(RIKEN)