SambaNova’s AI paired with Fugaku supercomputer to develop 'digital twins'

All part of the Japan's Society 5.0 project

AI systems specialist SambaNova has been chosen by Japan’s RIKEN scientific research institute to provide a DataScale system for the Fugaku supercomputer to assist with research into Japan’s “Society 5.0” vision.

SambaNova develops integrated systems based on specialized hardware designed for machine learning models, which the biz claims have a 6x performance advantage over GPU-based systems. The Fugaku supercomputer, meanwhile, was formerly the most powerful system in the world until surpassed in the Top500 rankings by the Frontier exascale supercomputer last year, and comprises 152,064 of Fujitsu’s 48-core Arm-based A64FX processor chips.

SambaNova said that its DataScale system will be delivered to the RIKEN Center for Computational Science (R-CSS) in order to accelerate the AI capabilities of Fugaku for the purposes of research on digital twins (digitized versions of physical objects) for the Japanese government's Society 5.0 plan for the nation's future.

The government plan views Society 1.0, as hunter-gatherers, the agricultural revolution led to Society 2.0, as did industrialization for Society 3.0. We're now in the information age (Society 4.0) and the goal is to create a new 'super-smart society' for the fifth version, with digital technologies powering a stable and eco-friendly vision for Japan.

In a statement, R-CSS director Professor Satoshi Matsuoka said the move will provide the Center with a new option for accelerating the integration of HPC simulations and AI with the Fugaku supecomputer.

"The new SambaNova system will boost research into the convergence of HPC and AI, including ultra-high-resolution computer vision for building a digital twin for the Society 5.0 era," he said.

Tech giant Fujitsu discussed digital twin technology at its ActivateNow: Technology Summit last year, saying that it could bring together the digital world and behavioral sciences to allow civic authorities to rehearse how policies designed to address various problems could change human behaviors.

SambaNova said that R-CCS researchers will use DataScale to develop surrogate models to improve the accuracy of ultra-high-resolution 3D computer vision. This would be used for more precise inspection of highways and other social infrastructure and to process ultra-high-resolution image datasets that it said are difficult to handle with conventional technologies.

“SambaNova's unique reconfigurable dataflow architecture and its large memory capacity provide a comprehensive AI solution capable of processing higher resolution 3D images beyond 512x512x512. Combined with Fugaku, it provides a new resource for digital twin simulations,” said the company’s Country Sales Director Toshinori Kujiraoka.

The DataScale systems are built around SambaNova's Reconfigurable Dataflow Unit (RDU) chips. Each of these is a grid of configurable compute and memory elements linked by an on-chip communication fabric. The latter is configurable by software so the flow of data through the chip elements mirrors the dataflow graph of the machine-learning algorithm it is running.

SambaNova's DataScale units have already been paired with other supercomputers, such as the Corona system at the Lawrence Livermore National Laboratory (LLNL), and also used at the Argonne National Laboratory in Illinois. ®