HPL-AI
Overview
HPL-AI stands for "the High Performance LINPACK for Accelerator Introspection" [1], which allows mixed-precision arithmetic to solve a linear equation system and seeks to highlight the convergence of HPC and artificial intelligence (AI) workloads. We have developed a highly tuned HPL-AI benchmark code on Fugaku, on which Fujitsu A64FX and Tofu-D technologies are harnessed to yield excellent performance. As far as the official page of the HPL-AI benchmark [1, 2], a number of HPC centers have benchmarked HPL-AI, starting with Oak Ridge National Laboratory’s Summit machine in 2019, and the latest list includes that our benchmark on Fugaku, which achieved two exaFLOP/s in mixed-precision performance (see also our early report in [3]). Our reference code will be released as Open Source Software, which does not include a copyright-concerned source code, while that is not our champion code.
[1] https://www.icl.utk.edu/hpl-ai/
[2] https://icl.bitbucket.io/hpl-ai/
[3] S.Kudo, K.Nitadori, T.Ina, and T.Imamura: "Implementation and Numerical techniques for One Eflop/s HPL-AI benchmark on Fugaku", ScalA20 conj SC20, 2020.
Downloads
- HPL-AI (tgr.gz, 88KB) (May 25, 2021)
- A distributed-memory implementation of HPL-AI benchmark for Fugaku and others. We have replaced a copyright-sensitive source code with generic or emulated source code. You need to contact us to know further information.
Publications
- S.Kudo, K.Nitadori, T.Ina, and T.Imamura: "Implementation and Numerical techniques for One Eflop/s HPL-AI benchmark on Fugaku", ScalA20 conj SC20, 2020.