HOME > Research Results > 2019 > “ABCI” Cloud Computing System for AI Achieves World’s Fastest Deep Learning Speed

Update(MM/DD/YYYY):12/03/2019

“ABCI” Cloud Computing System for AI Achieves World’s Fastest Deep Learning Speed

– Also ranked fourth in the world for energy-efficient supercomputers –

(Summary of AIST press release on November 13, 2018)

Researchers: Department of Information Technology and Human Factors

Key Points

The “ABCI Grand Challenge” program provides support for external users who are seeking to maximize the abilities of ABCI.
In the second ABCI Grand Challenge in 2018, a research group of Sony Corporation has updated the record for the world’s fastest deep learning speed.
ABCI is ranked fourth in the world for energy efficient supercomputers, and fifth in the world for processing performance using the conjugate gradient method.

Left: A section of the ABCI computing node rack; Center: Part of the cooling system; Right: The AI Data Center

Social Background of Development

AI Bridging Cloud Infrastructure (ABCI) is a large-scale, energy efficient cloud computing system capable of advanced artificial intelligence (AI) processing, which commenced operation on August 1, 2018. ABCI was developed as a part of the Ministry of Economy, Trade and Industry’s Global AI Research Hub Project (Second Supplementary Budget for fiscal 2016). As an open, advanced, high-performance computing infrastructure for the development of AI technologies in Japan, ABCI’s aims are to promote the use of the system through cooperation among industry, academia and government, and by a diverse range of businesses, to accelerate R&D and verification of AI technologies utilizing a high level of computing power, and to advance social utilization of AI technologies. In addition, it also aims to address the most important issues in the field of AI.

Summary

The “ABCI Grand Challenge” program is being conducted in order to test the limits of the ability of ABCI, which commenced operation on August 1, 2018.

In the second ABCI Grand Challenge, which was conducted in October, a research group of Sony Corporation completed ResNet-50 learning using the ImageNet image classification dataset in approximately 3.7 minutes with 2176 GPUs, significantly reducing the previous fastest time of 6.6 minutes recorded by China’s Tencent in July. In addition, the results of benchmark tests conducted in conjunction with the second ABCI Grand Challenge placed the system at number 4 on the Green 500 List, with 14.423 gigaFLOPS per watt. Also, following on from being placed at number 5 on the Top 500 List in June, the system was placed at number 5 on the HPCG Performance List with a figure of 508.85 teraFLOPS.

By means of the “ABCI Grand Challenge” program, AIST will continue to support efforts to address the most important issues in the field of AI, which the enormous computing power of ABCI will make possible.

▲ ページトップへ