Fujitsu Labs has achieved the highest performance on the MLPerf HPC benchmark, which measures large-scale machine learning processing on supercomputers.

RIKEN-Fujitsu-Supercomputer-Fugaku
Working with the Japanese National Institute of Advanced Industrial Science and Technology (AIST) and RIKEN, the record was taken using part of the ABCI (AI-Bridging Cloud Infrastructure) supercomputer system operated by AIST, and part of supercomputer Fugaku (pictured), currently under development by RIKEN and Fujitsu.
“Utilising about half the computing resources of its system, ABCI achieved processing speeds 20 times faster than other GPU-type systems,” said Fujitsu. “That is the highest performance among supercomputers based on GPUs, computing devices specialised in deep learning. Similarly, about one tenth of Fugaku was utilised 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.”
Fugaku is remarkable for other reasons. It is the successor to Japan’s record-breaking ‘K’ and is currently the world’s fastest super computer according to the Top500 ranking – at 442Pflop. According to the Top500 organisation: “This puts it three times ahead of the number two system”. It has 7,630,848 Arm A64FX cores. Fugaku was made by Fujitsu and is installed at the RIKEN Center for Computational Science in Kobe.

The MLPerf HPC v0.7 results were presented last week at the virtual International Conference for High Performance Computing, Networking, Storage, and Analysis (SC20).
MLPerf HPC is based on MLPerf, established in May 2018 to rank systems running machine learning applications.
The HPC version was created to compare supercomputers, and is mooted as a potential industry standard. It consists of two benchmark programs:
CosmoFLow – deep learning models were trained to predict cosmological parameters from the results of three-dimensional simulations of dark matter distributed in space
DeepCAM – a deep learning model trained to identify abnormal weather phenomena with global climate prediction simulation data
Half of ABCI ranked first in metrics of all registered systems in the CosmoFlow benchmark program and a tenth of Fugaku ranked second.
ABCI was first amongst all registered systems in the DeepCAM benchmark program.
“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,” said Fujitsu.
But why half of a computer?
“In accordance with the rules of this time’s MLPerf HPC v0.7, the measurement was executing using not the whole system, but just half of ABCI’s resources,” said Fujisu, which went on to say a similar thing about one tenth of Fugaku.
Fujitsu, AIST and RIKEN plan to release the software stacks, including the library and the AI framework. “This move will make it easier to use large-scale machine learning with supercomputers, while its use in analysing simulation results is anticipated to contribute to the detection of abnormal weather phenomena and to new discoveries in astrophysics,” said Fujitsu. “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.”