Acryl to Localize GPU Cluster Network Technology... Selected for Ministry of Science and ICT Project

Joint Research with Yonsei University, Sungkyunkwan University, and Ajou University

"Addressing Vendor Lock-In Issues of InfiniBand-Centric Systems"

Acryl, a company specializing in artificial intelligence transformation (AX), is taking steps to localize GPU cluster network technology, which is considered a key bottleneck in Korea's artificial intelligence (AI) infrastructure.

On June 10, Acryl announced that it has been selected as the lead organization for a national project initiated by the Institute for Information & Communications Technology Planning & Evaluation (IITP), under the Ministry of Science and ICT. The project is titled "Development of Ethernet-Based GPU Cluster Network Fabric System and Optimization Technology for Maximizing Network Efficiency in Large-Scale GPU Cluster Environments."

This project is a large-scale research and development initiative with a total budget of approximately 6.7 billion won (with 5.5 billion won funded by the government). The project will run until December 2028. Acryl, as the lead organization, will conduct joint research with Yonsei University, Sungkyunkwan University, and Ajou University Industry-Academic Cooperation Foundations.

Acryl plans to integrate the technology acquired through this project into its commercial GPU infrastructure integration platform, "JONATHAN GPUBASE (GPUBase)." The company aims to complete a domestically developed GPU network fabric solution that can be applied to cloud service providers (CSPs), managed service providers (MSPs), public AI data centers, and medical AI infrastructure.

Currently, large-scale GPU cluster networks are dominated by NVIDIA's InfiniBand. However, operating limitations due to vendor lock-in—such as high licensing costs (OPEX), delays in the supply of dedicated equipment, and capital expenditure (CAPEX) burdens—have led to the emergence of open fabric solutions based on "RoCEv2," which leverage existing Ethernet assets, as alternatives. However, even when adopting RoCE, the separation of network operating systems (NOS) and fabric management systems by different vendors results in new limitations, such as increased complexity in multi-vendor operations, as operators become dependent on individual tools again.

Acryl is already addressing these issues through its commercial GPU infrastructure integration platform, "GPUBase," which supports both InfiniBand and RoCEv2 and provides a unified control layer for monitoring, managing, and optimizing multi-vendor RoCE fabrics. Through this project, Acryl plans to integrate open Ethernet GPU cluster fabric technology into GPUBase and, by the third year, conduct large-scale proof-of-concept testing on a RoCEv2 cluster established at its own IDC. This will position Acryl as a leader in driving the transition of Korea's AI infrastructure to Ethernet/RoCE.

Yeom Ikjun, Chief Technology Officer (CTO) at Acryl, stated, "Through this project, we will present an open Ethernet GPU network alternative that addresses the limitations of cost, supply, and asset utilization inherent to InfiniBand-centered architectures, and we will lay the foundation for a self-sustaining domestic AI infrastructure ecosystem."

This content was produced with the assistance of AI translation services.