The upgraded RAIDEN has increased its performance by a considerable margin, moving from an initial total theoretical computational performance of four petaflops (PFLOPS), to 54 PFLOPS, placing it in the top tier of Japan's systems. In having built this system, Fujitsu demonstrates its commitment to support cutting-edge AI research in Japan.
Since it began operations after system delivery in April 2017, the RIKEN AIP Center in Japan has put RAIDEN to use for R&D on next generation AI technology. Such cutting-edge AI research is conducted with enormous neural networks, a machine learning method particularly in deep learning. The increasing scale of neural networks promises to improve factors such as the accuracy with which the networks handle more complex characteristics, but has also led to a drastic increase in computational volume. Moreover, even in AI research beyond deep learning, computational time is increasing due to the increasing complexity of algorithms and the volumes of data involved.
The RIKEN AIP Center undertook this upgrade in light of its expanded usage needs, with a view toward increasing the efficiency of research and development and promoting further AI research using RAIDEN.
For the GPU servers specialized for deep learning, the system has been upgraded from NVIDIA DGX-1 servers featuring NVIDIA Tesla P100 GPUs to the latest NVIDIA Tesla V100 GPUs. In addition, by increasing the number of DGX-1 servers from 24 to 54, the upgraded system achieves a computational performance of 54 PFLOPS.
For the computational cluster servers capable of more general-purpose processing, in addition to the 32 existing Fujitsu Server PRIMERGY RX2530 M2 x86 servers, Fujitsu has newly deployed 64 additional Fujitsu Server PRIMERGY CX2550 M4 servers. In addition, it has deployed one Fujitsu Server PRIMERGY RX4770 M4 unit as a compute server that handles high-volume data.