Fujitsu, AIST, and RIKEN Achieve Unparalleled Speed on the MLPerf HPC Machine Learning Processing

Fujitsu, the National Institute of Advanced Industrial Science and Technology (AIST), and RIKEN today announced a performance milestone in supercomputing, achieving the highest performance and claiming the ranking positions on the MLPerf HPC benchmark(1). The MLPerf HPC benchmark measures large-scale machine learning processing on a level requiring supercomputers and the parties achieved these outcomes leveraging approximately half of the “AI-Bridging Cloud Infrastructure” (“ABCI”) supercomputer system, operated by AIST, and about 1/10 of the resources of the supercomputer Fugaku, which is currently under joint development by RIKEN and Fujitsu.

Utilizing about half the computing resources of its system, ABCI achieved processing speeds 20 times faster than other GPU-type systems. That is the highest performance among supercomputers based on GPUs, computing devices specialized in deep learning. Similarly, about 1/10 of Fugaku was utilized to set a record for CPU-type supercomputers consisting of general-purpose computing devices only, achieving a processing speed 14 times faster than that of other CPU-type systems.

The results were presented as MLPerf HPC v0.7 on November 18th (November 19th Japan Time) at the 2020 International Conference for High Performance Computing, Networking, Storage, and Analysis (SC20) event, which is currently being held online.

Background

MLPerf HPC is a performance competition in two benchmark programs: “CosmoFlow”(2), which predicts cosmological parameters, and “DeepCAM”(3), which identifies abnormal weather phenomena. The ABCI ranked first in metrics of all registered systems in the “CosmoFlow” benchmark program, with about half of the whole ABCI system(4), and Fugaku ranked second with measurement of about 1/10 of the whole system(5). The ABCI system delivered 20 times the performance of the other GPU types, while Fugaku delivered 14 times the performance of the other CPU types. ABCI achieved first place amongst all registered systems in the “DeepCAM” benchmark program as well, also with about half of the system. In this way, ABCI and Fugaku overwhelmingly dominated the top positions, demonstrating the superior technological capabilities of Japanese supercomputers in the field of machine learning.

Fujitsu, AIST, RIKEN and Fujitsu Laboratories Limited will release the software stacks including the library and the AI framework which accelerate the large-scale machine learning process developed for this measurement to the public. This move will make it easier to use large-scale machine learning with supercomputers, while its use in analyzing simulation results is anticipated to contribute to the detection of abnormal weather phenomena and to new discoveries in astrophysics. As a core platform for building Society 5.0, it will also contribute to solve social and scientific issues, as it is expected to expand to applications such as the creation of general-purpose language models that require enormous computational performance.

Fujitsu, Executive Director, Naoki Shinjo, said, “The successful construction and optimization of the software stack for large-scale deep learning processing, executed in close collaboration with AIST, RIKEN, and many other stakeholders made this achievement a reality, helping us to successfully claim the top position in the MLPerf HPC benchmark in an important milestone for the HPC community. I would like to express my heartfelt gratitude to all concerned for their great cooperation and support. We are confident that these results will pave the way for the use of supercomputers for increasingly large-scale machine learning processing tasks and contribute to many research and development projects in the future, and we are proud that Japan’s research and development capabilities will help lead global efforts in this field.”

Hirotaka Ogawa, Principal Research Manager, Artificial Intelligence Research Center, AIST, added, “ABCI” was launched on August 1, 2018 as an open, advanced, and high-performance computing infrastructure for the development of artificial intelligence technologies in Japan. Since then, it has been used in industry-academia-government collaboration and by a diverse range of businesses, to accelerate R&D and verification of AI technologies that utilize high computing power, and to advance social utilization of AI technologies. The overwhelming results of MLPerf HPC, the benchmark for large-scale machine learning processing, showed the world the high level of technological capabilities of Japan’s industry-academia-government collaboration. AIST’s Artificial Intelligence Research Center is promoting the construction of large-scale machine learning models with high versatility and the development of its application technologies, with the aim of realizing “easily-constructable AI”. We expect that the results of this time will be utilized in such technological development.

Satoshi Matsuoka, Director General, RIKEN Center for Computational Science, maintained, “In this memorable first MLPerf HPC, Fugaku, Japan’s top CPU supercomputer, along with AIST’s ABCI, Japan’s top GPU supercomputer, exhibited extraordinary performance and results, serving as a testament to Japan’s ability to compete at an exceptional level on the global stage in the area of AI research and development. I only regret that we couldn’t achieve the overwhelming performance as we did for HPL-AI to be compliant with inaugural regulations for MLPerf HPC benchmark. In the future, as we continue to further improve the performance on Fugaku, we will make ongoing efforts to take advantage of Fugaku’s super large-scale environment in the area of high-performance deep learning in cooperation with various stakeholders.”