Successful Demonstration of Distributed Computing Platform Using Optical-Electrical Fusion and Next-Gen Distributed Database

Due to the rapid expansion of generative AI and cloud services, the volume of data moving between and within data centers is exploding, leading to a concurrent increase in power consumption. Conversely, sudden server troubles causing service outages can significantly impact social activities. Therefore, to evolve critical social services such as autonomous driving, telemedicine, smart cities, and scientific computing, a mechanism to manage massive amounts of data swiftly, efficiently, with low power consumption, and high fault tolerance is indispensable.

Under the commissioned project by the New Energy and Industrial Technology Development Organization (NEDO) titled "Development of AI Chip and Next-Generation Computing Technology Enabling High-Efficiency and High-Speed Processing / R&D Item [2] Development of Next-Generation Computing Technology (JPNP16007) / Development of High-Efficiency and High-Speed Distributed Computing System Technology Using Heterogeneous Integrated Optoelectronics," this project successfully demonstrated a next-generation distributed computing platform. To simultaneously resolve the issues of "expanding data processing volume," "reducing power consumption," and "improving fault tolerance (high availability)," the platform transmits data via light with low latency and utilizes remote computing resources. Its defining feature is the integrated optimization of communication and computation by combining optical chips, optical transceivers, optical networks, and distributed databases, rather than optimizing them individually.

At the core, the optical chip realizes a high-speed, low-power consumption optical device using heterogeneous material integration technology that combines InP and silicon. The optical transceiver aims to reduce power consumption by shifting part of the electrical processing to optical processing and supports high capacity by utilizing multiple wavelengths. Furthermore, the multi-route elastic optical network allows flexible switching of communication routes and bandwidth as needed.

In the demonstration, real-field transmission using a 400 Gbps x 2 wavelength optical transceiver prototype, optical network control connecting multiple locations, and the operation of an application demonstrating low latency and high availability using the distributed database "Tsurugi" compatible with this optical network were confirmed. This charts a course toward realizing the 10 Tbps-class high-capacity, low-power communication required in the AI era and the effective utilization of widely distributed computing resources.

This outcome paves the way for a future vision where computing resources in various locations are dynamically linked and utilized as a single massive computing platform, rather than merely aggregating data in one place for processing. It is expected to be utilized as a foundational technology supporting social infrastructure in the AI era. Moving forward, it is anticipated to be deployed across a wide range of fields, including data centers, telecom infrastructure, and smart cities, as a technology that manages growing communication traffic while curbing power consumption.

## 1. Background

With the proliferation of generative AI, the volume of data processing required for AI training and inference is soaring. To support this processing volume into the future, it is necessary to effectively utilize regionally distributed computing resources. Additionally, to address the energy consumption issues associated with this increased processing, data centers and networks are required to achieve both higher capacity and lower power consumption simultaneously.

Conventionally, the increase in communication volume was handled by adding faster equipment. However, simply adding more equipment in the future will face severe constraints regarding power, installation space, and operational costs. Moreover, fixed communication bandwidth allocation makes it difficult to operate flexibly according to the congestion or availability of computing resources. From a carbon-neutral perspective, technology that operates communication and computation with less power is crucial. Especially as the power demand of data centers supporting AI processing rises, reducing the power consumption of the communication segment is a socially critical issue. Furthermore, realizing highly available databases has been sought from an information security standpoint.

Therefore, this project integrated optical chips, optical transceivers, optical networks, and distributed databases into a single distributed computing platform. By connecting remote locations with low latency and allocating necessary communication bandwidth to where it is needed, it aimed to realize a new ICT platform capable of handling the computing demands of the AI era while achieving high availability.

## 2. Current Achievements

### ① Heterogeneous Material Integrated Optical Chip

As a heterogeneous material integration platform technology, an InP chiplet/SOI*1 wafer bonding, which excels in positional accuracy between the InP-based active region and the Si waveguide and enables high-density, multi-functional integration, was introduced and applied to optical devices for digital coherent transmission*2. Regarding the wavelength tunable laser*3 acting as the light source, the project achieved the world's first device integrating two differently designed InP-based gain regions onto a single silicon (Si) photonics*4 circuit, covering not only the C-band (wavelength: 1530 nm - 1565 nm) but also the L-band (wavelength: 1565 nm - 1625 nm).