This content was published in the January 2026 issue of “BUSINESS NETWOK” by RIC TELECOME.

“SAKURAONE” Ranks 49th Globally in Supercomputer Performance: Why an 800-GPU Cluster Was Built Using SONiC

SAKURA internet has constructed one of Japan’s largest AI training clusters, featuring 800 GPUs, as the foundation for its cloud-based HPC (High-Performance Computing) service, “SAKURAONE.”

This new cluster, recognized globally with a 49th-place ranking in supercomputer performance, was built in just four months. Furthermore, the GPU-to-GPU interconnects were configured using white-box switches and the open-source OS, SONiC. What factors led to the success of this unprecedented challenge?

“SAKURAONE” achieved 49th place in the ‘Top500’ global supercomputer performance rankings.* We take pride in the fact that it is the highest-ranking system built using open technologies like Ethernet and SONiC,” says Takashi Inoue, an infrastructure developer in the Cloud Services Department of the Cloud Business Division at SAKURA internet, regarding the “SAKURAONE” cloud HPC system.

“SAKURAONE” is a service designed for large-scale LLM (Large Language Model) development, incorporating the construction and operational expertise cultivated by the company, a leader in Japan’s GPU cloud business.
As a pioneer, the company adopted SONiC when launching the initial version of “SAKURAONE”. Based on the know-how internationally recognized by the Top500 ranking, they have reconstructed the system with new technology and are now offering it as a commercial service.

In the aforementioned Top500, custom-built supercomputers dominate the top rankings. In terms of networking, the vast majority of GPU clusters are built using vendor-specific protocols like InfiniBand, a standard dedicated to HPC. In contrast, “SAKURAONE” is characterized by its implementation using general-purpose GPU servers and highly transparent Ethernet technology.

*Source: Results as of June 2025.