Ellis Group said on the 2nd that it will team up with Arista Networks Korea, a global artificial intelligence (AI) network corporations, to implement Korea's first "Ethernet-based large-scale GPU clustering."
On the 1st, Ellis Group signed a memorandum of understanding (MOU) with Arista Networks Korea for "next-generation AI network technology cooperation." The two sides said they plan to focus on shifting AI infrastructure, which had relied on costly dedicated networks, to Ethernet, a general-purpose network standard.
Ellis Group said that while the AI network market had mainly used Nvidia's data center–dedicated network "InfiniBand," corporations have recently been shifting to Ethernet-based networks that are more expense-efficient.
Under this agreement, the two sides will build an AI network based on "RoCE v2" technology. "RoCE v2" is a next-generation core technology that boosts data transfer speeds on existing general-purpose Ethernet networks and eliminates bottlenecks to support large-scale AI computation.
Ellis Group decided to introduce a RoCE v2–based high-performance network environment into Ellis Cloud, an AI-specialized cloud, and into its AI PMDC (portable modular data center). Arista Networks Korea will handle high-performance network design and technical support.
Park Jung-guk, Ellis Group chief technology officer (CTO), said, "The core of AI infrastructure competitiveness depends on how efficiently tens of thousands of graphics processing units (GPUs) exchange data," adding, "Together with Arista, we will present an economical and efficient AI infrastructure standard and take the lead so that corporations at home and abroad can transition to AI faster and more easily."
Kim Se-jin, head of Arista Networks Korea, said, "Collaboration with Ellis Group is an important opportunity to realize innovation in RoCE v2–based AI data center networks."