Implementation of D2C Liquid Cooling System; Immersion Cooling Model Also on Display
Featuring NVIDIA B200 GPUs and AI Server Racks
AI-Dedicated Network and Power Infrastructure Established
On December 11, KT Cloud announced the launch of its "AI Innovation Center," a technology demonstration hub that combines artificial intelligence (AI) data center infrastructure with a showroom. KT Cloud plans to demonstrate and apply AI infrastructure technologies at this center, including liquid cooling technology capable of rapidly cooling high-heat AI chips.
The center replicates the actual environment of an AI data center, featuring the latest NVIDIA B200 graphics processing units (GPUs), AI server racks, direct-to-chip (D2C) liquid cooling devices, and AI networks. In particular, the center showcases cooling technologies that KT Cloud is currently commercializing and validating.
KT Cloud has implemented the D2C liquid cooling system, which is the global standard for AI servers, as a priority at the center. This technology involves attaching plates to GPU chips to allow direct contact with chilled water. KT Cloud demonstrated the operation of a server-type D2C liquid cooling system based on load modules, consisting of eight 1kW-class modules, at the center. Last month, KT Cloud became the first in Korea to commercialize this technology at its Gasan AI Data Center.
To achieve this, KT Cloud independently developed both server-type and rack-type load modules, and has verified conditions such as coolant flow rate, pressure, and temperature to match the B200 GPU and NVL72 specifications. NVL72 refers to NVIDIA's high-speed GPU interconnect technology, NVLink, expanded to support up to 72 connections.
The company also unveiled a working model of immersion cooling technology, which cools AI servers by submerging them directly in a special liquid. KT Cloud explained that, through proof-of-concept (PoC) testing with actual loads, the company confirmed up to 60% energy savings.
In addition, KT Cloud partnered with Arista, a specialist in AI networking, to build an AI-dedicated network at the center based on RoCEv2 (RDMA over Converged Ethernet v2). This network supports ultra-high-speed data transmission between AI servers and features next-generation Ethernet RDMA technology, which improves cost efficiency and scalability compared to the existing high-speed InfiniBand network. RDMA is a technology that reduces latency and increases speed by enabling direct memory-to-memory data transfers between servers without involving the CPU.
KT Cloud also introduced its self-developed standard rack-based AI power infrastructure. This standard rack features a high-density power supply structure that supports AI servers of over 20kW, and applies a 48V DC power supply to enhance energy efficiency and stability. Its modular design allows for easy expansion and replacement, and it is compatible with global Open Compute Project (OCP) standards.
The center also demonstrated the operation of AI-based solutions. Pathfinder, developed and patented by KT Cloud, enables autonomous power control through digital twin-based simulation.
A demonstration was also held using autonomous robots to patrol the server room. The center provides a space where visitors can experience and understand AI training and machine learning operations (MLOps) based on the B200 GPU.
Choi Jiwoong, CEO of KT Cloud, stated, "The AI Innovation Center is not just a showroom but a core platform for demonstrating next-generation data center technologies and building AI infrastructure focused on customer value. KT Cloud will set a new standard for data center technology innovation and lead the growth of Korea's AI ecosystem."
© The Asia Business Daily(www.asiae.co.kr). All rights reserved.
