fp21axpy
software
posted on 2021-01-08, 22:03 authored by WACCPD Author 3WACCPD Author 3, Takuma Yamaguchi, Kohei Fujita, Tsuyoshi Ichimura, Akira Naruse, Muneo Hori, Maddegedara LalithFP21 data type is our proposed floating-point number format, which is composed of 1-bit sign part, 8-bit exponent part, and 12-bit fraction part. FP21 has the advantage of the same dynamic range as FP32 and a better accuracy than FP16.
3 × FP21 numbers are stored into 1 × 64-bit element, so it is suitable for 3-D simulations.
For memory bandwidth bound computations, 1.5-fold speedup can be expected by using FP21 data types compared to the standard FP32 data types.
Baseline experimental setup, and modifications made for the paper
Relevant hardware details: Each compute node of ABCI has four NVIDIA Tesla V100 GPUs and two Intel Xeon Gold 6148 CPUs (20 cores).
Operating systems and versions: Red Hat 4.8.5-28
Compilers and versions: PGI v19.3
Libraries and versions: OpenMPI v2.1.6, CUDA v10.0
Key algorithms: conjugate gradient
Input datasets and versions: Strong ground motion of The Southern Hyogo prefecture earthquake in 1995 observed at Kobe JMA observatory
Output from scripts that gathers execution environment information is described in environment_info.dat.
